Wu2013 PDF

Analog Circuits and Signal Processing
Series Editors
Mohammed Ismail
Mohamad Sawan
For further volumes:

http://www.springer.com/series/7381
Rong Wu
Johan H. Huijsing
Kofi A. A. Makinwa
Precision Instrumentation
Amplifiers and Read-Out
Integrated Circuits
123
Rong Wu Kofi A. A. Makinwa
e-mail: rongwu.grace@gmail.com e-mail: k.a.a.makinwa@tudelft.nl
Johan H. Huijsing
e-mail: j.h.huijsing@tudelft.nl
ISBN 978-1-4614-3730-7 ISBN 978-1-4614-3731-4 (eBook)

DOI 10.1007/978-1-4614-3731-4
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2012940238
Ó Springer Science+Business Media New York 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the
work. Duplication of this publication or parts thereof is permitted only under the provisions of
the Copyright Law of the Publisher’s location, in its current version, and permission for use must always
be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright
Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

To my parents and Zhiyu
Preface
Sensors are ubiquitous in our lives and indispensable in many applications, e.g.,
process control, weighing scales, environmental monitoring, and temperature
measurement. They can be found in wafer steppers, weighing scales, mobile
phones and automobiles, etc. While these sensors convert the physical signals into
electrical domain, their output voltage are small, in the millivolt-level, such as
thermocouples and bridge transducers (thermistor bridges, Hall sensors and load
cells). Therefore, they need amplifiers to boost such signals to levels compatible
with the input ranges of typical Analog-to-Digital Converters (ADCs). To achieve
sufficient signal-to-noise ratio, the input referred error of the amplifier should be
reduced to a low enough level that means the amplifier must have low thermal and
1/f noise, high accuracy, and low drift. Achieving all these is quite challenging in
today’s mainstream CMOS technology whose inherent precision is limited by
1/f noise, component mismatch, gain error, and drift. A further challenge is to
achieve good power efficiency since many sensor systems are battery-powered.
This is also essential for precision temperature measurement to restrict local self-
heating errors.
This book describes the use of power-efficient techniques to mitigate low fre-
quency errors, resulting in interface electronics with high accuracy, low noise, and
low drift. Since this book is mainly about techniques for eliminating low frequency
errors, it describes the nature of these errors and the associated dynamic offset
cancelation techniques used to mitigate them. It then shows how these techniques
can be applied to operational amplifiers. Then these techniques are extended to
current-feedback instrumentation amplifiers (CFIAs) which are well suited for
bridge readout. Since the main disadvantage of CFIAs is their limited gain
accuracy, the available techniques to improve this are discussed, such as resistor-
degeneration, dynamic element matching, etc. The advantages and disadvantages
of each of these techniques are analyzed.
Later, it presents the architecture design and implementation of a CFIA, in
which a new technique (offset reduction loop) is proposed to suppress the chopper
vii
viii Preface
ripple without causing noise folding. An improved version CFIA of the first CFIA
is described, which maintains the noise performance of the first design and also
achieves high gain accuracy without trimming. This is obtained by dynamic
element matching and another proposed new technique (gain error reduction loop).
The basic architecture of the first CFIA is then combined with an ADC to build
a readout IC. The system-level design of the readout IC together with imple-
mentation details and measurement results are presented. The CFIA and the ADC
collaborate at system level to achieve an optimum performance. Measurement
results show that the realized readout IC achieves state-of-the-art offset and drift
performance.
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Overview of Read-Out Electronics for Sensors . . . . . . . . . . . . . 3
1.3 Instrumentation Amplifier Topologies . . . . . . . . . . . . . . . . . . . 5
1.3.1 Three-Opamp Topology . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Switched-Capacitor Topology. . . . . . . . . . . . . . . . . . . . 5
1.3.3 Capacitively-Coupled Topology . . . . . . . . . . . . . . . . . . 6
1.3.4 Current-Mode Topology. . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.5 Current-Feedback Topology . . . . . . . . . . . . . . . . . . . . . 8
1.4 Current-Feedback Instrumentation Amplifier. . . . . . . . . . . . . . . 9
1.5 Read-Out ICs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Targeted Sensor Applications and Challenges . . . . . . . . . . . . . . 14
1.7 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Dynamic Offset Cancellation Techniques

for Operational Amplifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Low Frequency Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 Offset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.2 1/f Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.3 Drift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Dynamic Offset Cancellation Techniques . . . . . . . . . . . . . . . . . 23
2.3.1 Auto-Zeroing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.2 Chopping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Charge Injection Compensation Techniques
in Auto-Zeroed and Chopper Amplifiers. . . . .............. 30
2.4.1 Compensation Techniques for Charge Injection . . . . . . . 30
ix
x Contents
2.4.2 Charge Injection and Clock Feed-Through

in Chopper Amplifiers . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.3 Chopper Charge Injection Suppression Techniques . . . . . 35
2.4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5 Dynamic Offset Compensated Operational Amplifiers . . . . . . . . 37
2.5.1 Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.5.2 Ping-Pong Operational Amplifier . . . . . . . . . . . . . . . . . 38
2.5.3 Chopper-CDS Operational Amplifier . . . . . . . . . . . . . . . 40
2.5.4 Offset-Stabilized Operational Amplifiers . . . . . . . . . . . . 41
2.5.5 Chopper Offset-Stabilized Operational Amplifiers. . . . . . 42
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Current-Feedback Instrumentation Amplifiers and Gain Accuracy

Improvement Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1 Current-Feedback Instrumentation Amplifier. . . . . . . . . . . . . . . 51
3.1.1 Indirect Current-Feedback Instrumentation Amplifier . . . 52
3.1.2 Direct Current-Feedback Instrumentation Amplifier . . . . 53
3.2 Precision Current-Feedback Instrumentation Amplifiers . . . . . . . 54
3.2.1 Chopper-Stabilized Current-Feedback
Instrumentation Amplifier . . . . . . . . . . . . . . . . . . . . . . 55
3.2.2 Ping-Pong Auto-Zeroed Current-Feedback
Instrumentation Amplifier . . . . . . . . . . . . . . . . . . . . . . 56
3.2.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3 Gain Accuracy Improvement Techniques . . . . . . . . . . . . . . . . . 58
3.3.1 Current-Feedback Instrumentation Amplifier with
Resistor-Degenerated Input Stages . . . . . . . . . . . . . . . . 59
3.3.2 Chopper-Stabilized Current-Feedback Instrumentation
Amplifier with Auto-Gain Calibration . . . . . . . . . . . . . . 61
3.3.3 Ping-Pong-Pang Current-Feedback Instrumentation
Amplifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 A Chopper Instrumentation Amplifier with Offset

Reduction Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1 Amplifier Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Amplifier Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Offset Reduction Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3.1 Basic Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3.2 Transfer Function Analysis. . . . . . . . . . . . . . . . . . . . . . 77
4.4 Other Sources of Chopper Ripple . . . . . . . . . . . . . . . . . . . . . . 81
4.4.1 Cascode Buffer Isolation . . . . . . . . . . . . . . . . . . . . . . . 81
4.4.2 Chopper Ripple from the Intermediate Stage . . . . . . . . . 83
Contents xi
4.5 Applying ORL to General Purpose Instrumentation

Amplifiers and Operational Amplifiers. . . . . . . . . . . . . . . . . . . 84
4.6 Circuit Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.6.1 The Input Stages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.6.2 The Intermediate and Output Stages . . . . . . . . . . . . . . . 90
4.6.3 The Cascode Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.6.4 Constant-Gm Bias Circuit . . . . . . . . . . . . . . . . . . . . . . . 93
4.6.5 Chopper Clock Design and Layout . . . . . . . . . . . . . . . . 94
4.7 Measurement Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.8 Benchmark and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 102
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5 A Chopper Instrumentation Amplifier with Gain Error

Reduction Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2 Dynamic Element Matching . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.3 Analog Gain Error Reduction Loop . . . . . . . . . . . . . . . . . . . . . 109
5.3.1 Basic Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3.2 Qualitative Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.3.3 Quantitative Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4 Digitally-Assisted Gain Error Reduction Loop . . . . . . . . . . . . . 114
5.5 Comparison Between ORL and GERL . . . . . . . . . . . . . . . . . . . 116
5.6 The Effects of Chopping, DEM and GERL
on CFIA Performance . . . . . . . . . . . . . . . . . . . . . . . . . . .... 117
5.7 Circuit Implementations . . . . . . . . . . . . . . . . . . . . . . . . . .... 118
5.7.1 Current-Feedback Instrumentation Amplifier
with Analog Gain Error Reduction Loop . . . . . . . . .... 118
with Digitally-Assisted Gain Error Reduction Loop. . . . . 124
5.8.1 Noise. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.8.2 Output Ripple Measurement . . . . . . . . . . . . . . . . . . . . . 128
5.8.3 INL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.8.4 Gain Accuracy and Gain Drift . . . . . . . . . . . . . . . . . . . 131
5.8.5 Settling Behavior of Analog GERL
and Digitally-Assisted GERL . . . . . . . . . . . . . . . . .... 133
5.9 Benchmark and Conclusions . . . . . . . . . . . . . . . . . . . . . . .... 135
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 136
6 Read-Out Integrated Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

6.1 ADC Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.2 Architecture Design of the ADC . . . . . . . . . . . . . . . . . . . . . . . 140
6.2.1 Modulator Topology . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.2.2 Non-Idealities in the DR Modulator . . . . . . . . . . . . . . . 145
xii Contents
6.3 Gain Accuracy Improvement Techniques in the Read-Out IC . .. 151

6.3.1 Dynamic Element Matching . . . . . . . . . . . . . . . . . . . .. 151
6.3.2 Digitally-Assisted Gain Error Correction Scheme . . . . .. 153
6.4 Offset and 1/f Noise Suppression Techniques
in the Read-Out IC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 155
6.4.1 Previous Approach
(Multi-Stage Chopping and System-Level Chopping) . .. 155
6.4.2 Proposed Approach (Input-Stage Chopping
Combined with System-Level Chopping) . . . . . . . . . . . . 156
6.5 Error Correction Techniques Summary. . . . . . . . . . . . . . . . . . . 159
6.6 Circuit Implementations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.6.1 CFIA Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.6.2 ADC Implementation. . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.1 Original Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.2 Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.3 Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.4 Chapter 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.5 Main Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.6 Other Applications of this Work . . . . . . . . . . . . . . . . . . . . . . . 181
7.7 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
About the Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Chapter 1
Introduction
This thesis describes the theory, design and realization of precision instrumenta-
tion amplifiers and read-out ICs for interfacing bridge transducers and thermo-
couples. The goal of the work is to investigate power-efficient techniques to
eliminate low frequency (LF) errors in the read-out electronics, so as to achieve
high accuracy, low noise and low drift while preserving low power consumption.
In this chapter, the motivation and objectives of this work are described, then an
overview of prior art read-out electronics is given. This is followed by a
description of a challenging application: the read-out of a precision thermistor
bridge intended for high resolution temperature measurements in wafer steppers.
Finally, the highlights and organization of the thesis are presented.
1.1 Motivation
A sensor can be defined as a device that forms the interface between non-electrical
physical domains and the electrical domain. Examples of such physical domains
are the thermal, magnetic, mechanical, radiant and chemical domains. Sensors are
ubiquitous in our lives and indispensable in many applications, e.g. process con-
trol, weighing scales, environmental monitoring, and temperature measurement.
They can be found in wafer steppers, weighing scales, mobile phones and auto-
mobiles, etc. For example, there are more than 300 sensors in a modern car and the
overall value of the market is expected to grow from $9.9 billion in 2009 to $16.1
billion in 2014 [1].
While these sensors convert the physical signals into electrical domain, their
output voltage are small, in the millivolt-level, such as thermocouples and bridge
transducers (thermistor bridges, Hall sensors and load cells). This thesis focuses on
the design of the interface electronics for such sensors.
R. Wu et al., Precision Instrumentation Amplifiers and Read-Out Integrated Circuits, 1

Analog Circuits and Signal Processing, DOI: 10.1007/978-1-4614-3731-4_1,
2 1 Introduction
Vref Vref
Measurement
Rgauge1 junction
Vout Vout
Vout
Rgauge2 Reference
junction
(a) (b) (c)

Fig. 1.1 Examples of bridge transducers and thermocouples a Load cell b Thermocouple
c Thermistor bridge
Load cell
A load cell is a sensor that is used to convert a force into an electrical signal. It
usually consists of a number of strain gauges configured in a Wheatstone bridge
(Fig. 1.1a). Through a mechanical arrangement, the sensed force deforms the
strain gauge, thus changing its electrical resistance. A recent report has shown that
the global load cell market is forecast to reach $1.5 billion dollars by 2015 [2].
Thermocouple
A thermocouple consists of two wires made of different metals that are joined at
one end, called the measurement junction. At the other end of the conductors, a
reference junction is formed (Fig. 1.1b). If the measurement and the reference
junctions are at different temperatures, a voltage appears across the two terminals
that is a function of this temperature difference. Thermocouples are widely used in
industrial manufacturing environments.
Thermistor Bridge
A thermistor is a resistor whose value varies significantly with temperature. To
measure temperature, they are usually configured in a Wheatstone bridge structure
(Fig. 1.1c). Compared to thermocouples, thermistor bridges exhibit higher sensi-
tivity and lower noise, thus they are widely used in precision temperature mea-
surement applications, such as temperature control and compensation systems.
Hall Sensors
The principle behind the Hall Effect is that a magnetic field induces a voltage
between two points on the sides of a current-carrying conductor. A Hall sensor is
thus a four terminal device which can be modeled as a Wheatstone bridge. Hall
sensors can be used for contactless current sensing, since they are sensitive to the
induced magnetic field instead of the target (current) itself. Besides current-
sensing, they are also widely used for position measuring, speed detection and
proximity switching applications.
1.2 Overview of Read-Out Electronics for Sensors 3
Sensor Microprocessor
Vref
Rgauge1 Analog Digital
output Read-Out output
Electronics
Rgauge2
Fig. 1.2 Read-out electronics bridging the analog and digital worlds
1.2 Overview of Read-Out Electronics for Sensors
The electrical information produced by the sensor is usually an analog signal and
thus needs to be converted to a robust digital signal for further signal processing
[3, 4]. The system that converts the analog signal from the sensor to the digital
domain is called a sensor read-out system. Figure 1.2 shows a typical strain gauge
readout system. The analog output signal of the strain gauge is processed by the
read-out electronics and converted into a digital signal.
Given the broad applications of thermocouples and bridge transducers it is
important to investigate and improve the quality of their read-out electronics and
this will be the main goal of this thesis. Load cells, thermocouples, thermistor
bridges and Hall sensors typically output LF small signals in the millivolt-range.
Therefore, they need amplifiers to boost such signals to levels compatible with the
input ranges of typical Analog-to-Digital Converters (ADCs) (Fig. 1.3). A single
integrated chip on which both the preamp and the succeeding ADC is implemented
is called a read-out IC.
Although the differential output voltage of the sensor Vid can be as small as a
few millivolts (Fig. 1.3), the common-mode (CM) voltage VCM, depending on the
application, may be much larger and may even vary by a few volts during the
period of operation. Furthermore, the CM voltage of thermocouples may equal one
of the supply rails, usually ground. To accommodate this variable CM voltage, an
Instrumentation Amplifier (IA) is generally used for sensor read-out applications.
To accurately process the millivolt-level signal from the sensor, the input referred
error of the IA should be at the microvolt- or nanovolt-level. To cope with CM
variations of a few volts, the IA should have a common-mode rejection ratio
(CMRR) greater than 120 dB. Furthermore, it should have high input impedance
so as not to attenuate the sensor signal or load the sensor. This amplifier is very
critical since it determines the overall performance of the read-out IC. To sum up,
the main functions of this amplifier are to
1. Amplify the weak differential voltage (Vid)
2. Reject the sensor common-mode voltage (VCM) (CMRR [ 120 dB)
3. Be capable of handling CM voltages near the rails
4. Provide high input impedance for bridge read-out
4 1 Introduction
Fig. 1.3 Read-out IC for Sensor

bridge transducers that + Digital
output
consists of an instrumentation Vid A A Vid ADC
amplifier and an ADC -
Instrumentation
VCM Amplifier
As the amplifier is used to detect a small differential signal, its input-referred

error (due to noise and offset) must be reduced well below the minimum signal
amplitude. Since most sensor applications operate near DC with a bandwidth of a
few Hz, 1/f noise is the dominant noise source. Although bipolar technology is
well known for its low offset and low 1/f noise [5], nowadays, CMOS is the
preferred technology because of its low cost and the powerful capability of digital
signal processing. However, amplifiers realized in CMOS technology have no-
nidealities such as offset and 1/f noise. The worst-case offset can be as large as
10 mV, while the 1/f noise corner frequency can be a few tens of kHz. This
problem can be solved with dynamic offset cancellation techniques. The first MOS
chopper operational amplifier was reported in [6]. Later, CMOS amplifiers with
microvolt offset levels have been achieved [7–10] and amplifiers with 1/f noise
corner frequency of a few Hz [12, 13] and even a few tens of mHz [14] have been
reported.
Besides offsetgain error is another dominant source of error. It is usually
determined by component mismatch and has typical values of ±1 % in a standard
CMOS process. Furthermore, the amplifier’s gain should have very low temper-
ature drift, so that it can be used for temperature measurement applications (e.g.
for the read-out of thermocouple and thermistor bridge). Instrumentation ampli-
fiers meeting such specifications (low noise, low offset and low drift) are further
classified as precision instrumentation amplifiers. The succeeding ADC converts
the amplified analog output to the digital domain, while maintaining the signal-to-
noise-ratio from the precision instrumentation amplifier.
To conclude, achieving low thermal and 1/f noise, high accuracy and low drift is
quite challenging in today’s mainstream CMOS technology whose inherent pre-
cision is limited by 1/f noise, component mismatch, gain error and drift. A further
challenge is to achieve good power efficiency since many sensor systems are
battery-powered. This is also essential for precision temperature measurement to
restrict local self-heating errors.
This thesis addresses these challenges in two ways. First, it presents the design
of two stand-alone precision instrumentation amplifiers which can be indepen-
dently used in many practical systems to drive an external ADC. Second, it
describes the design of a read-out IC that combines the instrumentation amplifier
and the ADC into one chip, so as to provide a digital output.
1.3 Instrumentation Amplifier Topologies 5
Fig. 1.4 Bridge

Vin+ + R3 R4
instrumentation amplifier
realized by three operational OA1
amplifiers -
R21
+ -
R1 Vout
- +
R22
- Vref
OA2
Vin- + R5 R6
1.3 Instrumentation Amplifier Topologies
Instrumentation amplifiers can be built in several ways. The commonly used

topologies are the classic three-opamp, switched-capacitor, capacitively-coupled,
current-mode and current-feedback instrumentation amplifiers (IAs).
1.3.1 Three-Opamp Topology
The three-opamp IA uses voltage feedback to obtain a gain determined by resistors

[15]. Figure 1.4 shows a fully-differential three-opamp IA in which the first stage
is a fully-differential amplifier with a gain of (R21 ? R1 ? R22)/R1 and the second
stage is a differential amplifier.
The CMRR of the three-opamp IA is determined by the product of the finite
gain of its first stage, and the finite CMRR of its second stage. The latter is
determined by the matching of the feedback resistors R3-R6, which usually leads
to a CMRR of about 80 dB [16]. Furthermore, it can not sense the supply rails
because the input CM level of the amplifier must be set within its output voltage
range. This topology is also not very power-efficient, as it requires the use of two
high-gain opamps. However, it exhibits high input impedance and good linearity
over a wide input and output range. If a differential output is required, the first
stage can be used alone, which is known as the ‘‘two-opamp’’ topology.
1.3.2 Switched-Capacitor Topology
The switched-capacitor (SC) IA uses capacitors as the feedback elements [17, 18].
A fully-differential SC IA is shown in Fig. 1.5. When clock U1 is high, the input
signal is sampled on capacitor C1, while the integration capacitor C2 is reset. When
6 1 Introduction
Fig. 1.5 Switched-capacitor Φ1

instrumentation amplifier
Φ2 Φ1
Φ1 Φ2 C2
C1
+ + - +
V in OA Vout
- - + -
C1
Φ2 Φ1
C2
Φ1
Fig. 1.6 Capacitively- Cm

coupled instrumentation Rb
Iin Cin
amplifier
+ - + -
Vin Gm1 Gm2 Vout
- + - +
Iin CHin Cin CHout
Rb Cm
Cfb
Cfb CHfb
clock U2 is high, the charge stored on C1 is transferred to C2. The closed-loop gain
of the IA is determined by C1/C2. With careful layout, this gain can achieve 0.1 %
gain error and low drift over temperature. Furthermore, this topology accommo-
dates a large CM input range since the input sampling capacitors block DC.
However, it can not provide a continuous signal and the sampling procedure of the
SC amplifier increases noise level due to noise folding. The noise associated with
sampling is the well-known kT/C noise [19]. To reduce noise, input capacitor C1
needs to be increased. However, this results in a decrease in the input impedance
due to the switching impedance of input capacitors [20].
1.3.3 Capacitively-Coupled Topology
A recent development is the continuous-time capacitively-coupled (and chopped)

IA [12, 21]. As shown in Fig. 1.6, it employs an input chopper (polarity reversing
switch) to convert the input DC signal into an AC signal, which can be transported
via the capacitors. An output chopper then converts the amplified AC signal back
to DC. Since capacitors block DC signals, this topology exhibits a rail-to-rail input
CM range. Furthermore, it is very power-efficient since it has a continuous-time
Reset ( rst)
Slow chop ( sc)
Input
1
Mod k
s
ADC 1+z-1 2
On-chip Digital
sensors Output
f
Fast chop
( fc, fc_d)
Fig. 1.7 Block diagram of an improved capacitively-coupled analog-to-digital interface [22]
fc
Cfb
fc
rst s
Cm
h
In Cs int s Out
Gm1 Gm2 Gm3 Gm4
Ch
grd
V-to-I Opamp -C S&H Amplifier
Fig. 1.8 A closed-loop capacitive-feedback sampler [22]
signal path and its power consumption is mainly dominated by Gm1. Its input
impedance is typically in the order of several MX, determined by the impedance of
the input capacitors at the chopping frequency. This can be increased to a few tens
of MX by using an impedance boosting technique [21]. The disadvantage of this
topology is that the switched-chopper capacitor causes spikes (or glitches) and
more noise at the amplifier output.
An improved capacitively-coupled analog-to-digital interface was reported in
[22]. As shown in Fig. 1.7, it consists of a sampler and a sigma-delta (DR) ADC.
The sampler employs a closed-loop capacitively-coupled topology consisting of a
V-to-I converter, a Gm-C integrator followed by a sample-and-hold amplifier
(Fig. 1.8). Unlike [12, 21] in which the input chopper precedes the input capaci-
tors, here, the input chopper is shifted in the sensor. The sampler (Fig. 1.8) directly
processes the modulated sensor signal, thus its input capacitor Cs provide high
input impedance and also store the offset for coarse offset cancellation. The
residual offset, 1/f noise of the interface electronics (mainly from the ADC) and the
mismatch of the input capacitors are then eliminated by nested chopping that
chops the complete read-out chain.
8 1 Introduction
Fig. 1.9 Current-mode Current Mirror 1

instrumentation amplifier VDD
Vin+
+
A1 +
- Vout
A3
-
R2
I VDD
R1
Current Mirror 2 GND
+
Vin- A2
-
The presence of the sampler (Fig. 1.8) means that this circuit can not be, used
as a stand-alone IA with a continuous-time output signal. This implies that
capacitively-coupled IAs is more compatible with a SC sampled ADC. With
proper timing, sampling of the spikes at the IA output [21] can be avoided and the
non-continuous signal path [22] is not an issue for a SC sampled ADC.
1.3.4 Current-Mode Topology
Figure 1.9 shows a current-mode instrumentation amplifier [23–26]. The feedback

around the input amplifiers A1 and A2 forces the input voltage across the resistor R1.
The current through this resistor is mirrored by precision current mirrors and con-
verted into a voltage by R2, and then buffered by the output opamp A3. The CMRR of
this topology depends on the matching of the current-mirrors and the DC precision
of the current mirrors is essential for the overall offset, gain accuracy, drift and
linearity. Since the matching of an impedance-boosted current mirror is still
insufficient for the required DC precision, thin-film resistor-degenerated current
mirrors [26] is used and a CMRR larger than 120 dB is achieved. However, precise
thin-film resistors are not always available in CMOS technology.
1.3.5 Current-Feedback Topology
Figure 1.10 shows a current-feedback instrumentation amplifier (CFIA). The input

transconductor Gm2 and feedback transconductor Gm3 convert the input and
feedback voltages into corresponding currents. Their difference is then nulled by
the gain of Gm1. The overall feedback ensures that the output currents of Gm2 and
Gm3 cancel and thus the amplifier’s gain is given by
Fig. 1.10 Current-feedback C1

instrumentation amplifier R21 Vout+
Vin+ + - + -
Gm2 Gm1 R1
Vin- - + - +
R22
Vout+
Vfbk+ C2
+ +
Gm3
Vfbk- - -
Gm2 R1 þ R21 þ R22

Gain ¼ ð1:1Þ
Gm3 R1
The CFIA is well suited for bridge read-out. First, compared to the three-opamp
topology, it achieves higher CMRR, because the input transconductor Gm2 isolates the
input CM level by converting the input differential voltage to a differential current [27].
Its CMRR is mainly determined by the CMRR of Gm2 and can be made greater than
120 dB. Second, it is capable of handling input CM voltages that include either of the
supply rails [16]. Third, it is more power efficient because it can be seen as a merged
version of the three-opamp topology [28], in which the output stages are shared.
Compared to the switched-capacitor IA, the CFIA avoids noise folding.
Furthermore, the CFIA has higher input impedance than the capacitively-coupled
topology, and it does not produce output glitches. The CFIA is thus more suitable
for use as a stand-alone IA.
The main disadvantage of the CFIA is its limited gain accuracy. From (1.1),
assuming the open-loop gain of the CFIA is high enough and that precision
external feedback resistors (R1, R21 and R22) are used, the CFIA’s closed-loop gain
accuracy is mainly determined by the matching between the input and feedback
transconductors (Gm2 and Gm3). Furthermore, the linear range of a CFIA is often
limited by the input and feedback transconductors to several tens of mV. Although
this limited input range is not a problem for many bridge applications, extending it
will make the CFIA useful for other applications.
The first part of the thesis will focus on the design of improved CFIAs, while
their major disadvantages–limited gain accuracy, limited input range and nonlin-
earity will be addressed.
1.4 Current-Feedback Instrumentation Amplifier
The first CFIA was introduced by Analog Devices [29] in 1971, and was imple-
mented in bipolar technology. Later, the current-feedback concept was again
described by Huijsing in 1981 [30] and by Säckinger in 1987 [31]. In 1984, a
10 1 Introduction
bipolar CFIA used in low-power biomedical applications was reported by Hamstra

[32]. In 1993, Dool described another bipolar CFIA with a CM input range that
included the negative rail [16].
The first CFIA implemented in CMOS technology was presented in 1987 by
Steyaert for medical applications [33]. Although amplifiers in CMOS technology
typically has larger offset and 1/f noise than these in bipolar technology, the
existence of good MOS switches means that both these non-idealities can be
mitigated by dynamic offset cancellation techniques. Later, in 2004, Chan intro-
duced a CFIA that employs dynamic offset cancellation techniques to achieve low
offset (16 lV) and low 1/f noise [34].
Recently, several CFIAs have been reported [8, 11, 35] which achieve offsets
less than 5 lV or even less than 2 lV. Table 1.1 summaries the performance of
these CFIAs. Since power efficiency is an important design criterion in this work,
the noise efficiency factor (NEF) [33] is used to evaluate the power efficiency of
these CFIAs. It relates the amplifier’s noise PSD and supply current, as given by:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2Itot
NEF ¼ Vn ; ð1:2Þ
p Vt 4kT BW
where Vn is the input-referred noise voltage, Itot is the supply current of the
amplifier, Vt is the thermal voltage which equals kT/q, and BW is the amplifier
bandwidth.
As seen from Table 1.1, although these CFIAs all achieve offsets of a few lV,
they consume significant amount of power (NEF [ 24). Unlike noise, the offset
does not have a direct trade-off with power consumption. This means that it should
be possible to achieve low offset with low power.
Through a new technique (an offset reduction loop, ORL) [28], a precision
CFIA is presented in this thesis that achieves microvolt-offset with 3 9 improved
power efficiency (NEF = 8.8) compared to [8]. Moreover, its 1/f noise corner is in
the mHz range, which has not been achieved by previous stand-alone CMOS
amplifiers.
As mentioned before, the main disadvantage of the CFIA is its limited gain
accuracy, which is determined by the mismatch of the input and feedback trans-
conductances. If precision feedback resistors are used, in the worst-case, this
mismatch can be as large as 2 %. To reduce this mismatch, resistor-degeneration
was used [11, 35], since with careful layout, or by trimming, resistors can be made
to match better than transistors. As a result, a gain error of 0.1 % has been
achieved [11, 35]. However, resistor degeneration leads to a significant loss in
power efficiency (NEF = 43 [35] and 153 [11]). Moreover, trimming the degen-
eration resistors [11] increases production cost. This thesis presents a power-
efficient CFIA that achieves high gain accuracy without trimming. By using
dynamic element matching and a gain error reduction loop (GERL) [36], it
achieves an untrimmed gain error of 0.06 % in a power efficient manner
(NEF = 11.2).
1.5 Read-Out ICs 11
Table 1.1 Performance of low-offset CFIAs

Witte [8] Pertijs [35] Witte [11]
Year 2009 2010 2008
Supply voltage 5V 3.0 to 5.5 V 2.8 to 5.5 V
Supply current 325 lA 1.7 mA 850 lA
Input noise PSD 42 nV/HHz 27 nV/HHz 136 nV/HHz
CMRR 130 dB 142 dB 140 dB
Gain error 0.05 % (Untrimmed) ±0.1 % (Untrimmed) ±0.1 % (trimmed)
(Relative) (Absolute) (Absolute)
Offset \2.5 lV \2 lV \5 lV
GBW 640 kHz 800 kHz 1 MHz
NEF [33] 29.2 43 153
1.5 Read-Out ICs
The second part of the thesis is devoted to the design of a read-out IC. It consists of
a precision instrumentation amplifier followed by an ADC. The IA provides high
input impedance and relaxes the offset and noise requirements of the ADC.
Since the read-out IC acts as an ADC, a figure of merit (FOM) [37] is used to
evaluate its power efficiency. This FOM relates the read-out IC’s resolution and
bandwidth with its the power consumption, as given by
Power
FOM ¼ ð1:3Þ
2 BW 2ENOB
where BW is the bandwidth of the ADC, ENOB is the effective number of bits,
defining as
SNDR 1:76
ENOB ¼ ð1:4Þ
6:02
where SNDR is the Signal-to-Noise-Distortion-Ratio. Note that including ENOB
in the formula takes into account the distortion introduced by the ADC.
Many precision read-out ICs have been reported, which achieve more than 20-
bit resolution, low offset and gain drift (\15 ppm/°C) [38–42]. They are intended
for precision instrumentation and measurement applications. To achieve such high
resolution within a small bandwidth (\250 Hz), a delta-sigma (DR) ADC is a good
choice. This is due to the fact that the resolution of Nyquist-rate ADCs is limited
by component matching, while DR ADCs apply over-sampling technique which
trades speed for resolution. Furthermore, with noise shaping, DR ADCs can easily
achieve a resolution higher than 18-bit [43].
Figure 1.11 shows the block diagram of a DR ADC. It consists of a single-loop
DR modulator and a decimation filter. The DR modulator consists of a loop filter,
performing the noise-shaping, a low resolution quantizer, which is over sampled
12 1 Introduction
Decimation filter
Loop-filter Quantizer
In Digital Down Out
A/D filter sampler
D/A
Fig. 1.11 Block diagram of a DR ADC
and a digital-to-analog converter (DAC). The oversampling ratio (OSR) is defined

as OSR = fs/2fb, where fs is the sampling rate and fb is the signal bandwidth. The
DR modulator shapes the frequency response of the quantization errors in such a
way that the quantization error is reduced in the frequency band of interest, while
it is increased outside that band. Therefore, high resolution can be obtained in a
relatively small bandwidth. Since most of the quantization noise is shifted to
higher frequencies, it is necessary to eliminate the high frequency noise by using a
decimation filter.
For instrumentation applications, the input signal is a low-frequency or DC
signal, so the DR ADC typically operates in ‘single-shot’ mode, which means that
it powers up, produces a single conversion result and finally powers down again to
save power. This type of DR ADCs are called incremental DR ADCs. Their main
characteristics are:
1. The loop filter and the decimation filter are reset at the beginning of a
conversion.
2. The modulator does not operate continuously, but runs for a limited number of
N clock cycles, producing a bitstream of N bits.
Incremental DR ADCs provide very precise conversion with accurate gain, high
linearity and low offset [44]. A first-order DR modulator for instrumentation
applications was already introduced by Van der Plassche in 1978 [45], which
achieve 6-bit resolution reference to 1 V. In 2006, Quiquempoix reported a 22-bit
third-order incremental ADC reference to 5 V [44]. It achieves a 2 lV offset,
2 ppm gain error and 4 ppm INL with a FOM of 4.8 pJ/conv.
However, since bridge sensors typically produce millivolt-level signals, they
need precision IAs to boost the sensor signals to the typical ADC input range of a
few volts (Fig. 1.3). The IA is the most challenging and power consuming part,
since it determines the noise and accuracy performance of the read-out IC [39].
The IAs in previous precision read-out ICs generally employ switched-capac-
itor (SC) or two-opamp IA topologies. For instance, Analog Devices reported a
read-out IC in 1997 [38] that used a SC IA, as shown in Fig. 1.12. Due to the
sampling in the SC amplifier, its noise level increases due to noise folding.
Furthermore, an additional input buffer was necessary to provide high input
impedance for bridge read-out. To satisfy the noise specification, this read-out IC
1.5 Read-Out ICs 13
fLF
fout
Digital
ΣΔM FIR
Output
•••••
SC 0 fLF 2fLF 3fLF 4fLF Freq

Ain + Vos
Buffer Ain - Vos
Preamp Sinc3 filter
Sensor
Bridge
Transducer ADC
Fig. 1.12 Block diagram of the read-out IC in [39]
R Digital
Output
C M Decimation
filter
R
Sensor Two-opamp
Bridge IA
Fig. 1.13 Block diagram of read-out IC in [39]
Digital
V-I Output
Decimation
Converter Modulator filter
fslow ffast ffast fslow
Fig. 1.14 Block diagram of hall sensor interface in [46]
results in a high power consumption of 80 mW to achieve an 18-bit resolution with

10 mV full scale. The ADC is realized with a second-order incremental DR ADC.
System-level chopping is employed to chop the entire analog signal path at a slow
frequency of fLF [38]. The subsequent digital decimation filter creates notches to
suppress the modulated offset due to chopping. Overall, it achieves an offset drift
in the level of 10 nV/°C. The FOM of the total read-out IC is 43,000 pJ/Conv.
In 2000, Cirrus Logic described a read-out IC [39], shown in Fig. 1.13. It
consists of a two-opamp IA and a fourth-order incremental DR ADC. This
architecture is also not particularly power efficient, since the two-opamp IA
requires two high-gain amplifiers, thus resulting in a high power consumption of
40 mW to achieve a 19-bit with 28 mV full scale in a BW of 128 Hz
(FOM = 164pJ/Conv).
A Hall sensor interface for compass application was reported by van de Meer
et al. in 2005 [46] and its block diagram is shown in Fig. 1.14. It consists of a
voltage-to-current converter (VIC), whose differential current output is digitized
by a first-order sigma-delta modulator. The output of the modulator is averaged
over an entire spinning-current cycle by a decimation filter. To achieve high
14 1 Introduction
linearity over a wide dynamic range, the VIC consists of two opamps, each with a
DC gain of over 120 dB. This Hall sensor interface achieves 120 dB resolution
and less than 50nV offset with a consuming 21 mW.
Recently released read-out ICs from Analog Devices [40], Cirrus Logic [41]
and Texas Instruments [42], still use the SC or two-opamp IAs. That is part of the
reason why these read-out ICs are not very power efficient. The FOM of [40–42]
are 135, 9,000 and 172 pJ/Conv, respectively.
CFIAs, in contrast, provide superior power efficiency, since they avoid noise
folding and share output stages. Furthermore, the CFIAs exhibit high CMRR [8,
16] and rail-to-rail sensing capability [16, 35]. In order to show the potential of
CFIAs in read-out IC applications, this thesis presents the design of a read-out IC
[47] that combines a CFIA and an ADC. In this work, various dynamic cancel-
lation techniques are used to eliminate the 1/f noise, offset, gain error and drift.
Moreover, digital signal processing on the output of the ADC is explored to
improve the CFIA’s gain accuracy and gain drift. The CFIA and the ADC can then
collaborate together to achieve optimum performance. Compared to the state-of-
the-art [40], the proposed read-out IC achieves 10 9 better offset (50 nV), com-
parable gain drift (1.2 ppm/°C) and better power efficient (FOM = 111 pJ/Conv).
1.6 Targeted Sensor Applications and Challenges
As a test-case, the challenging task of developing interface electronics for a pre-

cision thermistor bridge was attempted. It is intended for use in wafer steppers
where lK-level temperature resolution is required [28].
In wafer steppers, thermal expansion is an important source of error and so
control loops are used to stabilize the temperature, and consequently the dimen-
sions of critical mechanical components. The mechanical stability requirements on
these components translate into allowable temperature drifts in the order of
100 lK per minute. To measure such slow drifts with sufficient resolution, the
total noise of the sensor read-out system should be less than 1lK (3r) from
21.1 °C to 22.9 °C, i.e. a 1.8°C range (equivalent to a 21-bit resolution) when
measured in a bandwidth ranging from 3 to 50 mHz.
Since the goal is to prevent temperature drift, the absolute accuracy of the
temperature measurement system may be much worse than its 1 lK resolution.
Thus absolute accuracy is established by periodic system-level calibrations. To
maintain accuracy during the (minute-long) intervals between calibrations, the
thermistors, as well as the interface electronics used to read them out should be
characterized by low LF noise, with 1/f noise corners in the order of only a few
mHz. In addition, the read-out electronics must have low offset and gain drift (a
few ppm/°C) to maintain system accuracy over temperature.
Compared to other temperature sensors such as thermocouples, negative tem-
perature coefficient (NTC) thermistors are well suited for high-resolution tem-
perature measurements because they can achieve high sensitivity, low thermal
1.6 Targeted Sensor Applications and Challenges 15
Fig. 1.15 Dual thermistor

bridge Vref =1. 22V
Rx R1
14nV/ √Hz
Vout
Metal Foil R1 Rx Thermistor

11.4k 11.4k
@22°C
noise, low 1/f noise corners (in the mHz range for high quality parts), and good
long-term stability (about 1 mK/year) [48].
In order to double the sensitivity, a double thermistor bridge consisting of two
thermistors and two metal foil resistors is shown in Fig. 1.15. In our case, the
resistance of the thermistor (Rx in Fig. 1.15) is 11.4 kX at 22°C, the same as the
resistance of the metal foil resistors (R1 in Fig. 1.15). Therefore, the bridge output
is zero at a temperature of 22°C. Due to the tolerance of its components, the bridge
has a gain error of ± 0.5 %. When biased by a band-gap reference at 1.22 V, the
common-mode voltage of the bridge is 0.61 V and its sensitivity is 27 mV/°C.
Thus, over the required 1.8°C range, the output range of the bridge is ± 24.3 mV.
Being only at the millivolt level, the output of the thermistor bridge should be
amplified before it is digitized or processed further. This requires the use of a low-
noise instrumentation amplifier followed by an ADC. The challenges associated
with the design of the read-out electronics are discussed below.
The first challenge is the required resolution: 0.33 lK (1r) in a 1.8°C range in a
bandwidth ranging from 3 mHz to 50 mHz. Together with the sensitivity of the
bridge (27 mV/°C), this translates into an input-referred noise density requirement
of 31 nV/HHz for the whole system. The thermal noise level of the thermistor
bridge is 14 nV/HHz, and so the amplifier’s white noise density was chosen to be
at roughly the same level, i.e. 16 nV/HHz. To achieve high power efficiency, the
amplifier’s noise should be white in the bandwidth of interest, which means that
the amplifier’s 1/f noise corner frequency must be below 3 mHz. To justify such
low noise specifications, the amplifier must also have high CMRR ([120 dB) and
PSRR ([120 dB).
The second challenge is the need for the amplifier to accommodate different
input and output CM voltages. Since the bridge is biased at 1.22 V, the input CM
is at 0.61 V. While the output CM is at 2.5 V, since the amplifier’s output is to be
digitized by an ADC with a 0–5 V input range.
Thirdly, since the sensor and the read-out electronics are calibrated as a single
system, the read-out electronics should exhibit very low offset and gain drift (a few
ppm/°C) to maintain system accuracy over temperature. Thus, the read-out
electronics aims to achieve gain and offset drift less than 1 ppm/°C and 10 nV/°C,
respectively.
16 1 Introduction
Table 1.2 Typical sensor specifications and targets of the read-out IC

Strain Gauge Thermo- Thermistor Hall sensor Target of
couple bridge read-out
electronics
Resolution 20-bit 18-bit 21-bit 20-bit 20-bit
(±10 mV) (±2 mV) (±48 mV) (±50 mV) (±40 mV)
Noise PSD H H 14 nV/HHz H 16 nV/HHz
1/f noise H H \3 mHz H 1 mHz
corner
BW 5 Hz H H H 5 Hz
Gain drift H H 1 ppm/°C H 1 ppm/°C
Gain error 0.02 % H H H 0.02 %
Offset H H H 50 nV 50 nV
Offset drift H H 10 nV/°C H 10 nV/°C
Supply H H 300 lA H \300 lA
current
Note for sensor specifications, only the most challenging one is shown in each row, while the
other two are represented with ‘H’
The final challenge involves self-heating. For the thermistor read-out applica-
tion in wafer steppers, the read-out electronics and the thermistor bridge are
located in the vacuum environment of a wafer stepper where heat sinking is a
significant problem. The power consumption of the interface electronics should not
be larger than that of the bridge (a few hundreds of lA) to restrict local self-
heating errors.
To make the interface electronics not only useful for thermistor bridge, but also
applicable for other voltage-out sensors, e.g. strain gauge and thermocouple and
Hall sensors, as shown in Table 1.2, the read-out IC was also designed to achieve
the same gain accuracy as stain gauges: less than 0.02 % [49]. Since Hall sensor
output with zero field conditions is less than 50 nV, in order to accurately process
sensor output, the interface electronics must have an offset less than 50 nV [46].
Furthermore, the read-out IC is designed to have a bandwidth of 5 Hz, to make it
useful for some strain gauges applications. Thus, the targeted noise specification is
a noise PSD of 16nV/HHz from 3 mHz to 5 Hz, which corresponds to a 20-bit
resolution reference to ±40 mV.
In summary, the low noise, low drift and low power qualities of the read-out
electronics presented in this thesis are beyond the capability of current available
interface electronics. These qualities make the demanding thermistor read-out
application in wafer steppers possible. In addition, the read-out electronics also can
be used for interfacing strain gauge, thermocouples and Hall sensors.
Although this research work is targeted for sensor applications, the new tech-
niques developed in this work also can be applied to other applications, such as
general purpose operational amplifiers, general purpose CFIAs and general pur-
pose read-out ICs.
1.7 Organization of the Thesis 17
Vout
Sensor CFIA
Stand-alone
(1)
Dout
Sensor CFIA ADC
Read-Out IC
(2)
Motivation &
challenges Chapter 1
(1)
Precision & low noise Theory
techniques Chapter 2
Precision CFIA & Gain Accuracy (1)

Theory
Improvement techniques Chapter 3
CFIA with (1)

Implementation
offset reduction loop Chapter 4
CFIA with (1)

Gain error reduction loop Implementation
Chapter 5
(2)
Read-out IC Implementation
Chapter 6
Fig. 1.16 Organization of the thesis
1.7 Organization of the Thesis
This thesis has been divided into seven chapters. The organization of the thesis is
illustrated in Fig. 1.16. It is divided into two parts. The first part is indicated as
(1) in Fig. 1.16: the design of precision stand-alone IAs for bridge interfacing. The
second part is indicated as (2) in Fig. 1.16: the design of a read-out IC that
combines the IA and an ADC. The outlines for each chapter are discussed as
follows.
Since this thesis is mainly about techniques for eliminating low frequency
errors, Chap. 2 describes the nature of these errors and the associated dynamic
offset cancellation techniques used to mitigate them. It then shows how these
techniques can be applied to operational amplifiers.
In Chap. 3, these techniques will be extended to CFIAs. Since the main
disadvantage of CFIAs is their limited gain accuracy, this chapter discusses the
available techniques to improve this, such as resistor-degeneration, dynamic
18 1 Introduction
element matching, etc. The advantages and disadvantages of each of these tech-
niques are analyzed.
Chapter 4 presents the architecture design and implementation of a CFIA. A
new technique (offset reduction loop) is proposed to suppress the chopper ripple
without causing noise folding. This CFIA achieves low offset, low thermal and
1/f noise and simultaneously, low power consumption. A 1/f noise corner of
1 mHz is achieved at a noise PSD of 15nV/HHz with a NEF of 8.8.
Chapter 5 discusses an improved version CFIA of the first CFIA described in
Chap. 4. It maintains the noise performance of the first design and also achieves
high gain accuracy without trimming. This is obtained by dynamic element
matching and another proposed new technique (gain error reduction loop). It
achieves less than 3 lV offset, and 0.06 % untrimmed gain error in a power
efficient manner (NEF = 11.2). These results show that the CFIA achieves state-
of-the-art performance in terms of offset, 1/f noise, gain accuracy and power
efficiency.
The basic architecture of the CFIA discussed in Chap. 4 is then combined with
an ADC to build a read-out IC. Chapter 6 discusses the system-level design of the
read-out IC together with implementation details and measurement results. The
CFIA and the ADC collaborate at system-level to achieve an optimum perfor-
mance. Measurement results show that the realized read-out IC achieves state-of-
the-art offset and drift performance.
In Chap. 7, the main conclusions of the thesis are presented. Special sections
have been included to highlight the original contributions of this thesis and some
recommendations for future work are made.
References
1. Fitzgerald V (2010) Automotive sensor demand forecast 2008 to 2017: global economic
rebound sparks growth. Available at https://www.strategyanalytics.com/default.aspx?
mod=ReportAbstractViewer&a0=5758
2. Global load cells market to reach US$1.5 billion by 2015, according to a new report by global
industry analysts, Inc (2011). Available at http://www.prweb.com/releases/load_cells/
single_point_shear_beam/prweb8121165.html
3. Bakker V, Huijsing JH (2000) High-accuracy CMOS smart temperature sensors. Kluwer
academic publishers, Boston
4. Huijsing JH, Riedijk FR, van der Horm G (1994) Developments in integrated smart sensors.
Sensors Actuators 43(1–3):276–288
5. Erdi G (1981) Amplifier techniques for combining low noise, precision, and high-speed
performance. IEEE J Solid-State Circuits SC-16(6):653–661
6. Poujois R, Borel J (1978) A low drift fully integrated MOSFET operational amplifier. IEEE J
Solid-State Circuits 13:499–503
7. Enz CC, Vittoz EA, Krummenacher F (1987) A CMOS chopper amplifier. IEEE J Solid-State
Circuits SC-22(3):335–342
8. Witte JF, Huijsing JH, Makinwa KAA (2009) A chopper and auto-zero offset-stabilized
CMOS instrumentation amplifier. Paper presented at the IEEE symposium on VLSI circuits,
pp 210-211
References 19
9. Pertijs MAP, Kindt WJ (2009) A 140 dB-CMRR current-feedback instrumentation amplifier

employing ping-pong auto-zeroing and chopping. In Proceedings of the IEEE ISSCC, digital
technical papers, pp 324–325
10. Sakunia S, Witte F, Pertijs M, Makinwa KAA (2011) A ping-pong-pang current-feedback
instrumentation amplifier with 0.04 % gain error. Paper presented at the IEEE Symposium on
VLSI Circuits, pp 60–61
11. Witte JF, Huijsing JH, Makinwa KAA (2008) A current-feedback instrumentation amplifier
with 5 lV offset for bidirectional high-side current-sensing In Proceedings of the IEEE
ISSCC, digital technical papers, pp 74–75
12. Denison T et al (2007) A 2.2 lW 94nV/HHz chopper-stabilized instrumentation amplifier for
EEG detection in chronic implants. In Proceedings of the IEEE ISSCC, digital technical
papers, pp 162–163
13. Yazicioglu RF et al (2008) A 200 lW eight-channel acquisition ASIC for ambulatory EEG
systems. In Proceedings of the IEEE ISSCC, digital technical papers, pp 164–165
14. Kejariwal M, Ammisetti P, Thomsen A (2002) A 250 ? dB open loop gain feedforward
compensated high precision operational amplifier. In Proceedings of the ESSCIRC, digital
15. AD8250 data sheet (2007). Analog Devices Inc., Norwood, MA
16. van den Dool BJ, Huijsing JH (1993) Indirect current feedback instrumentation amplifier
with a common-mode input range that includes the negative rail. IEEE J Solid-State Circuits
28(7):743–749
17. Van Peteghem PM, Verbauwhede I, Sansen WMC (1985) Micropower high-performance
SC building block for integrated low-level signal processing. IEEE J Solid-State Circuits
SC-20(4):837–844
18. Martin K, Ozcolak L, Lee YS, Temes GC (1987) A differential switched-capacitor amplifier.
IEEE J Solid-State Circuits SC-22(1):104–106
19. Enz CC, Temes GC (1996) Circuit techniques for reducing the effects of op-amp
imperfections: autozeroing, correlated double sampling, and chopper stabilization. In
Proceedings of the institute of electrical and electronics engineers(IEEE) vol 84(11).
pp 1584–1614
20. Verma N, Shoeb A, Bohorquez J et al (2010) A micro-power EEG acquisition SoC with
integrated feature extraction processor for a chronic seizure detection system. IEEE J Solid-
State Circuits 45(4):804–816
21. Fan Q, Huijsing JH, Makinwa KAA (2010) A 1.8 lW 1 lV-offset capacitively-coupled
chopper instrumentation amplifier in 65 nm CMOS. In Proceedings of the ESSCIRC, digital
22. Ezekwe C et al (2011) A 6.7nV/HHz sub-mHz-1/f-corner 14b analog-to-digital interface for
rail-to-rail precision voltage sensing. In Proceedings of the IEEE ISSCC, digital technical
papers, pp 246–247
23. Toumazou C, Ligey FJ, Anding ME (1990) Extending voltage-mode op amps to current-
mode performance. In Proceedings of the IEE.-Circuits, Devices and Systems, vol 137(2).
pp 116–130
24. Azhari SJ, Fazlalipoor H (2009) CMRR in voltage-op-amp-based current-mode
instrumentation amplifiers (CMIA). IEEE Trans Instrum Meas 58:563–569
25. Koli K, Halonen KAI (2000) CMRR enhancement techniques for current-mode
instrumentation amplifiers. IEEE Trans Circuits Syst I: Fundam. Theory Applicat
47(5):622–632
26. Schaffer V, Snoeij MF, Ivanov MV, Trifonov DT (2009) A 36 V programmable
instrumentation amplifier with sub-20 lV offset and a CMRR in excess of 120 dB at all
gain settings. IEEE J Solid-State Circuits 44(7):2036–2046
27. Huijsing JH (2011) Operational amplifiers: theory and design, 2nd edn. Springer, Netherlands
28. Wu R, Makinwa KAA, Huisjing JH (2009) A chopper current-feedback instrumentation
amplifier with a 1 mHz 1/f noise corner and an AC-coupled ripple reduction loop. IEEE J
Solid-State Circuit 44(12):3232–3243
20 1 Introduction
29. Krabbe H (1971) A high-performance monolithic instrumentation amplifier. In Proceedings

of the IEEE ISSCC, digital technical papers, pp 186–187
30. Huijsing JH (1981) Comparative study of some types of differential–differential amplifiers.
Paper Presented at the European Conference on Electrotechnics, Eurocon, B 6–8(1)(2),
pp 22–26
31. Säckinger E, Guggenbühl W (1987) A versatile building block: the CMOS differential
difference amplifier. IEEE J Solid-State Circuits SC-22(2):287–294
32. Hamstra GH, Peper A, Grimbergen CA (1984) Low-power low-noise instrumentation
amplifier for physiological signals. Med Biol Eng Comput 22(3):272–274
33. Steyaert MSJ, Sansen WMC, Chang Z (1987) A micropower low-noise monolithic
instrumentation amplifier for medical purpose. IEEE J Solid-State Circuits SC-22(6):
1163–1168
34. Chan PK, Ng KA, Zhang XL (2004) A CMOS chopper-stabilized differential difference
amplifier for biomedical integrated circuits. In Proceedings of the the 47th IEEE international
midwest symposium on circuits and systems (MWSCAS), III-33-6, vol 3
employing ping-pong auto-zeroing and chopping. IEEE J Solid-State Circuits 45(10):
2044–2056
36. Wu R, Huijsing JH, Makinwa KAA (2011) A current-feedback instrumentation amplifier
with a gain error reduction loop and 0.06 % untrimmed gain error. In Proceedings of the
IEEE ISSCC, digital technical papers, pp 244–245
37. Murmann B, Boser B (2004) Digitally assisted pipeline ADCs: theory and implementation.
Kluwer Academic Publishers, Boston
38. McCartney D, Sherry, Sherry A et al (1997) A low-noise low-drift transducer ADC IEEE J
Solid-State Circuits 32(7):959–967
39. Thomsen A et al (2000) A DC measurement IC with 130nVpp noise in 10 Hz. In Proceedings
of the IEEE ISSCC, digital technical papers, pp 334–335
40. AD7193 datasheet: http://www.analog.com/en/analog-to-digital converters/adconverters/
ad7193/products/product.html.
41. CS5530 datasheet: http://www.cirrus.com/en/products/pro/detail/P1108.html.
42. ADS 1282 datasheet: http://focus.ti.com/docs/prod/folders/print/ads1282.html.
43. Norsworthy SR, Schreier R, Temes GC (eds) (1997) Delta-sigma data converters: theory,
design and simulation. Piscataway, IEEE Press, New York
44. Quiquempoix V et al (2006) A Low-power 22-bit incremental ADC. IEEE J Solid-State
Circuits 41(7):1562–1571
45. van der Plassche RJ (1978) A sigma-delta modulator as an A/D converter. IEEE Trans
Circuits Syst 25(7):510–514
46. van de Meer JC, Riedijk FR, van Kampen E, Makinwa KAA, Huijsing JH (2005) A fully
integrated CMOS hall sensor with a 3.65lT 3r offset for compass applications. In
Proceedings of the IEEE ISSCC, digital technical papers, pp 246–247
47. Wu R, Huijsing JH, Makinwa KAA (2011) A 21-bit ± 40 mV range read-out IC for bridge
transducers. In Proceedings of the IEEE ISSCC, digital technical papers, pp 110–111
48. Meijer GC (1994) Thermal sensors. Institute of physics publishing, Bristol, Philadelphia
49. Slattery C, Nie M (2005) A reference design for high-performance, low-cost weigh scales.
Available at: http://www.analog.com/library/analogDialogue/archives/39-12/weigh_scale.html
Chapter 2
Dynamic Offset Cancellation Techniques
for Operational Amplifiers
At low frequencies, offset, 1/f noise and drift are the dominant error sources of
operational amplifiers. This is especially true in CMOS technology. This chapter
reviews precision techniques that can be used to achieve low 1/f noise and low
offset in operational amplifiers.
There are three types of CMOS offset cancellation techniques: trimming,
chopping, and auto-zeroing. Trimming is usually performed during production to
eliminate offset. Auto-zeroing is a sampling technique in which the offset is
measured and then subtracted in subsequent clock phases. Chopping, on the other
hand, is a continuous-time modulation technique in which the signal and offset are
modulated to different frequencies. Due to the modulated offset and 1/f noise, a
chopper ripple appears at the amplifier output. Since chopping and auto-zeroing
are dynamic techniques that continuously reduce offset, they also remove low
frequency 1/f noise as well as offset drift over temperature or time.
In auto-zeroing amplifiers, the residual offset is mainly caused by charge
injection and clock feed-through. While in chopper amplifiers, the residual offset is
mainly caused by demodulated clock feed-through spikes. Several techniques can
be used to counteract these non-idealities.
Later in this chapter, several dynamic-offset-compensation techniques used in
operational amplifiers will be discussed, e.g. ping-pong auto-zeroing, offset
stabilization, and specifically, chopper offset stabilization of a low-frequency path
in a multi-path amplifier. To suppress chopper ripple, numerous ripple reduction
techniques can be used. It will be shown that these all have significant drawbacks,
and thus new techniques are required.

Springer Science+Business Media New York 2013
22 2 Dynamic Offset Cancellation Techniques for Operational Amplifiers
2.1 Introduction
For sensor applications, the bandwidth of interest is generally a few Hz. In this
bandwidth, offset, 1/f noise and drift are the dominant error sources. Thus,
dynamic offset cancellation techniques are required to mitigate these errors. Before
those dynamic offset cancellation techniques are discussed, it is necessary to first
understand the nature and origins of these error sources.
2.2 Low Frequency Errors
2.2.1 Offset
In CMOS technology, the worst-case offset of a differential input pair can be as

large as 10 mV [1]. This offset is caused by manufacturing variation or uncer-
tainty. For example, MOS devices exhibit threshold voltage (Vth) mismatch
because Vth is a function of the doping levels in MOS channels and the gates, and
these parameters vary randomly from one device to another.
On the other hand, the dimensions of MOS devices suffer from random,
microscopic, variations during fabrication and hence there is mismatch between
the equivalent lengths and widths of nominally identical transistors. This mismatch
can be reduced by using large devices. However, this increases chip area and
therefore production cost.
2.2.2 1/f Noise
1/f noise is mainly caused by the defects in the interface between the gate oxide
and the silicon substrate, so it depends on the ‘‘cleanness’’ of the oxide-silicon
interface and may be considerably different from one CMOS technology to another
[2, 3]. The typical 1/f noise corner frequency of CMOStechnology is in the order of
several kHz to tens of kHz, making the 1/f noise a dominant error source at low
frequencies. Related to the lifetime of the carriers, the 1/f noise can be modeled as
a function of frequency [2], given by:
K
Vn2 ¼ ð2:1Þ
WLCox f
where K is a process-dependent constant in the order of 10-25V2F, W and L are the
width and length of the MOS transistor, Cox is the gate capacitance per unit area,
and f is the operation frequency. Generally, 1/f noise in PMOS is much lower than
NMOS in most technologies.
2.2 Low Frequency Errors 23
Fig. 2.1 Low frequency Noise

noise spectrum for CMOS offset, drift
PSD
amplifier
dB 1/f noise
1/f corner
frequency
thermal noise
10Hz 10kHz log(freq)
In (2.1), the noise spectral density of the 1/f noise is inversely proportional to
the frequency. The inverse dependence of (2.1) on the area of the transistor WL
suggests that to decrease 1/f noise, the device area must be increased. However,
this again increases chip area.
2.2.3 Drift
Drift is caused by the cross-sensitivity of some error sources to temperature or

time. Low drift is a critical requirement for precision temperature measurement
(e.g. thermistor bridges and thermocouples), since the drift of the interface elec-
tronics can then not be distinguished from the sensor signal itself.
Drift mainly manifests itself as offset drift and gain drift. Thus it can be reduced
by suppressing the offset and gain error to a low enough level, and furthermore by
dynamically compensating for their temperature drift.
To conclude, Fig. 2.1 depict the low frequency errors in CMOS amplifier. As
can be seen in the bandwidth of a few Hz for bridge transducer applications, the
main errors are caused by 1/f noise, offset, and drift. To mitigate these errors,
dynamic offset cancellation techniques can be employed, which will be described
in the next section.
2.3 Dynamic Offset Cancellation Techniques
To reduce offset, three types of offset cancellation techniques can be applied:

trimming, auto-zeroing, and chopping. Trimming involves measuring and then
reducing the offset during production. While this approach can be used to obtain an
order-of-magnitude reduction of the offset, it is unable to reduce the initial mV-level
offset below a few tens of lV, because offset drift is not compensated for. Moreover,
trimming does not eliminate low-frequency noise, such as 1/f noise. Dynamic offset
cancellation techniques, such as auto-zeroing or chopping are therefore needed to
counteract this problem.
Fig. 2.2 Auto-zeroing with CK

output offset storage Vos C1
X
- +
Vin A Vout
Y
C2
CK
S1 S2 S3 S4
VCM VCM
2.3.1 Auto-Zeroing
Auto-zeroing is a discrete-time sampling technique. It involves sampling the offset

of the amplifier in one clock phase, and then subtracting it from the input signal in
the other clock phase. There are three basic topologies for auto-zeroing [4]: output
offset storage (also called open-loop offset cancellation), input offset storage (also
called closed-loop offset cancellation) and closed-loop offset cancellation using an
auxiliary amplifier.
2.3.1.1 Output Offset Storage
Figure 2.2 depicts an auto-zeroed amplifier with output offset storage. When CK is
high, the amplifier is in the auto-zeroing phase in which its inputs are shorted
together, driving its output to Vout = AVos. During this period, nodes X and Y are
shorted together as well. When all the node voltages are settled, AVos is stored
across C1 and C2. When CK turns low, the amplifier enters the amplification phase.
The differential input voltage together with Vos is amplified, and stored on C1 and
C2. Since Vos is already stored on C1 and C2, VX and VY does not see Vos, which is
fully cancelled.
When a switch opens, it injects some charge into the surrounding circuitry. This
charge consists of gate-source/drain channel charge and charge injected through
the overlap capacitances (also known as clock feed-through). In reality, the charge
injection in the switches S3 and S4 will not completely cancel. The mismatch
charge injection results in a residual offset, given by
qinj3 qinj4
Vos;res ¼ ð Þ=A; ð2:2Þ
C1 C2
where qinj3 and qinj4 are the charge injection caused by switches S3 and S4, A is the
DC gain of the amplifier. Note that if A is large, AVOS may saturate the amplifier’s
output. For this reason, A is typically chosen to be between 10 and 100 [2]. An
integrated amplifier with three cascaded auto-zeroed amplifiers with output voltage
storage has been described in [4]. In [5], these stages were chopped, resulting in a
low drift MOSFET operational amplifier.
2.3 Dynamic Offset Cancellation Techniques 25
Fig. 2.3 Auto-zeroing with S3

input offset storage
CK
C1 Vos
X
+ -
Vin A Vout
Y
C2
CK
S1 S2
S4
CK
VCM
2.3.1.2 Input Offset Storage
The output offset storage technique limits the maximum gain of the amplifier. If a
high gain is needed, storing the offset at the input storage capacitance would be a
better solution. Figure 2.3 shows the basic principle of input offset storage tech-
nique [2]. In the auto-zeroing phase when CK is high, the output and input of the
amplifier are shorted together by switches S1 and S2, placing the amplifier in a
unity-gain configuration.
When the node voltages are settled, the output voltage Vout is given by
A
Vout ¼ VOS : ð2:3Þ
1þA
The circuit reproduces the amplifier’s offset at nodes X and Y, storing the result
on C1 and C2. Note that for a zero differential input, the differential output is equal
to VOS. Thus, the input-referred offset voltage of the overall circuit equals VOS/A if
S3 and S4 match perfectly.
If S3 and S4 have any mismatch, this will cause mismatch charge injection and,
in turn, lead to a residual offset, which is given by
VOS qinj3 qinj4
Vres þð Þ; ð2:4Þ
Aþ1 C1 C2
where qinj3 and qinj4 are the charge injection caused by switches S3 and S4, and A is
the DC gain of the amplifier.
From (2.4), the offset Vos is suppressed by the gain of the amplifier. The charge
injection and the leakage of the capacitors can be reduced by increasing the size of
the capacitors, but cannot be suppressed by the gain because the capacitors are
directly at the amplifier input.
The drawback of input offset storage and output offset storage is that they
introduce capacitors in the signal path. The bottom-plate parasitic of the capacitors
decreases the amplifier bandwidth, thus degrading its phase margin and stability.
CK
Vos1
- +
Vin Gm1 R Vout
CK Vos2 S3
S1 S2
- +
Gm2 S4
VCM
CK
C2 C1
Fig. 2.4 Auxiliary amplifier placed in a feedback loop during offset cancellation
2.3.1.3 Closed-Loop Offset Cancellation with Auxiliary Amplifier
To mitigate the stability issue, closed-loop offset cancellation with an auxiliary

amplifier can be used to isolate the offset storage capacitors from the signal path,
as shown in Fig. 2.4.
In the auto-zeroing phase, the inputs of Gm1 are shorted. Thus, the output
voltage Vout can be calculated as
½Gm1 VOS1 Gm2 ðVout VOS2 ÞR ¼ Vout;AZ ð2:5Þ
Thus,
Gm1 RVOS1 þ Gm2 RVOS2
Vout;AZ ¼ : ð2:6Þ
1 þ Gm2 R
This voltage is stored on C1 and C2 after S3 and S4 turn off. The offset voltage
referred to the main input is given by
Vout;AZ VOS1 VOS2
VOS;res ¼ þ : ð2:7Þ
Gm1 R Gm2 R Gm1 R
The charge injection due to the mismatch of S3 and S4 contributes to the offset
of Gm2. In order to attenuate this charge injection, as seen from (2.7), Gm2 is
usually chosen to be at least 50 times smaller than Gm1.
Note that in an auto-zeroed amplifier, half of the clock period is used for auto-
zeroing, so the amplified output is only available during part of the clock period.
Such amplifiers cannot provide a continuous-time output, unless a ping-pong
topology is employed [4, 6].
As seen from the discussion above, these three offset cancellation techniques
cancel offset by periodically subtracting the offset obtained during the previous
sampling moment. This assumes that the offset does not change too much during
the amplification time. Since low-frequency noise and DC offset can not be
Fig. 2.5 Noise spectrum of a Noise

sampled system Power
Under-sampled
Pno = (2fn,BW/fs)Pni noise
Input noise
-3 -2 -1 0 1 2 3 f/fs
fs
2fn,BW
distinguished from each other, these techniques also eliminate 1/f noise and drift.
However, the sampling action of the auto-zeroing techniques affects the amplifier’s
noise performance at frequencies below the sampling frequency [4].
2.3.1.4 Noise in Auto-Zeroing
As discussed above, auto-zeroing is a sampling technique. To complete settle

within a half clock cycle, the noise bandwidth fn, BW (determined by the time
constant of the system) is usually chosen to be larger than the auto-zeroing fre-
quency fAZ, so that the under-sampled noise folds back to DC, increasing the noise
PSD at the baseband.
The amount of folded noise to DC depends on the noise bandwidth fn, BW and
the auto-zeroing frequency fAZ. An exact quantitative calculation for the folded
noise of auto-zeroing can be found in [4, 7]. Since, this is rather complex, a more
intuitive explanation of noise folding in auto-zeroing is described here.
The noise folding factor n is defined as the ratio between the noise bandwidth
and the auto-zeroing frequency, as given by
2fn;BW
n¼ : ð2:8Þ
fAZ
Due to under-sampling, the noise power after sampling increases by this factor
as compared to that before sampling (Fig. 2.5), thus incurring a noise penalty. This
implies that by choosing a small fn, BW, the folded noise can be restricted.
Applying this concept, a slow-settling nulling loop is used in [8, 9] to reduce the
noise bandwidth to a fraction of the auto-zeroing frequency. This approach will be
discussed in Sect. 2.3.2.
Fig. 2.6 Chopping principle fch fch fch

in the time domain
fch fch
Vos DC
DC AC
+ ~0
+
Vin A1 V1 V2 Vout
+ + LPF +
DC AC AC DC
Vin V1 /A1 V2 /A1 Vout /A

Vin Vin
Vos
V0os 0 0 0
2.3.2 Chopping
Unlike auto-zeroing, chopping is a continuous-time modulation technique that

does not cause noise folding. Figure 2.6 shows a chopped amplifier together with
its ideal waveforms. The input voltage Vin first passes through a chopper driven by
a clock at frequency fch, thus it is converted to a square wave voltage at fch. Next,
the modulated signal is amplified together with its own input offset. The second
chopper then demodulates the amplified input signal back to DC, and at the same
time modulates the offset to the odd harmonics of fch, where they are filtered out by
a low-pass filter (LPF). This results in an amplified input signal without offset.
Low-frequency errors, such as 1/f noise and drift will be modulated and filtered
out along with offset. This can be seen in Fig. 2.7, which depicts chopping in the
frequency domain. To completely remove the 1/f noise, the chopping frequency
should be higher than the 1/f noise corner frequency. At the beginning, the signal is
modulated, and the noise and offset are superposed onto this modulated signal
(Fig. 2.7b). After amplification and the second chopper, the modulated signal is
demodulated back to DC, while the low-frequency noise and offset are modulated
to the harmonics of the chopper frequency, appearing as a chopper ripple at the
amplifier output (Fig. 2.7c). A LPF is then used to remove the modulated offset
and 1/f noise, resulting in a clean low-frequency signal without offset or 1/f noise
(Fig. 2.7d).
From the above discussion, the offset is amplified by the DC gain of A1, while
the signal is amplified by the effective gain of A1 at the chopping frequency fch, To
maximize the effective gain of the stage consisting of A1 and two choppers, the
optimum chopping frequency should be around 3 dB bandwidth of A1 [4].
The amplitude of the chopper ripple can be calculated with the help of the
simplified block diagram shown in Fig. 2.8. The chopped offset of the input stage
is filtered by the main Miller compensation capacitor CM1 and appears as a
offset & noise

modulated signal Modulated
offset & noise
Vin V1 V2 Vout
signal
1 3 5 f/fch 1 3 5 f/fch 1 3 5 f/fch 1 3 5 f/fch
(a) (b) (c) (d)
Fig. 2.7 Chopping principle in the frequency domain
R1
f ch f ch
Vos CM11
R2
+ + + - +
Vin Va Gm2 Gm1 +
- -
- Vout
CH1 CH1
-
CM12
Fig. 2.8 Simplified block diagram of a chopper amplifier
triangular waveform at the output. The peak-to-peak amplitude of the ripple can
then be approximated as:
Vos Gm2
Vout;ripple ¼ : ð2:9Þ
2CM1 fch
From (2.9), ripple amplitude can be reduced by reducing input-stage offset Vos with
careful layout, by increasing the chopping frequency fch1 or by increasing the size of the
Miller compensation capacitor. For a worst-case 20 mV offset, with Gm2 = 250 lA/
V, CM1 = 80 pF, VDD = 5 V, and fch1 = 40 kHz, Vout, ripple & 0.8 V. This is quite
large compared to the amplifier’s maximum 5 V output range and so must be
suppressed.
2.3.3 Conclusions
Both auto-zeroing and chopping techniques have been introduced. Table 2.1
compares and summarizes these two techniques. Chopping is superior to auto-
zeroing because it is a continuous-time modulation technique that does not cause
noise folding. However, chopping gives rise to a chopper ripple at the amplifier
output. Auto-zeroing does not introduce ripple and its discrete-time nature is well
compatible with switched-capacitor circuitry. Since power efficiency is an
Table 2.1 Comparison of auto-zeroing and chopping techniques

Auto-zeroing Chopping
Low frequency noise ± +
Power-noise efficiency - +
Ripple + -
Residual offset + ++
Note auto-zeroing removes 1/f noise, but causes the noise to fold back to DC, thus auto-zeroing in
terms of low-frequency noise is denoted as ‘‘±’’
Fig. 2.9 Charge injection

compensation using dummy
switches
W/2L
Vin Vout
W/L qinj1 qinj2
CH
important concern in our work, chopping is applied here. In Sect. 2.5.5, various
techniques will be discussed to eliminate the chopper ripple.
2.4 Charge Injection Compensation Techniques

in Auto-Zeroed and Chopper Amplifiers
2.4.1 Compensation Techniques for Charge Injection
The residual offset of auto-zeroed amplifiers is mostly determined by charge

injection and clock feed-through. To mitigate these two effects in a sampling
circuit, several techniques can be applied.
Dummy Switches
Charge injection can be partially compensated for by adding dummy switches
that are driven by a complementary clock signal and that inject an amount of
charge which compensates for the charge injected by the main switch [10]. The
effectiveness of the compensation depends on the matching of the injected charges.
A clock signal with a high slew rate can be used to obtain an equal distribution of
the channel charge in the main switch’s drain and source terminals. A half-size
dummy switch can then be used for compensation (Fig. 2.9). However, the
assumption of equal splitting of the charge between the source and drain is gen-
erally invalid, making this approach less attractive.
2.4 Charge Injection Compensation Techniques in Auto-Zeroed and Chopper Amplifiers 31
CK CH
Vin1 qinj1
Electrons Vout1
qinj1
ΔV=Δqinj/CH
Vin Vout
Vout2
qinj2 CH Vin2
qinj2
Holes CH
CK
(a) (b)
Fig. 2.10 a Using complementary switches to reduce charge injection b Using differential circuit
to suppress charge injection
Using Complementary Switches

Another approach to reduce the charge injection is to use both PMOS and
NMOS devices such that the opposite charge packets injected by the two cancel
each other (Fig. 2.10a). However, this cancellation is only effective for a limited
range of the input signal around half of the supply voltage.
Using a Fully Differential Circuit
The best way to compensate for charge injection is to use fully differential
circuitry (Fig. 2.10b). If the charge injection in the two half circuits matches, the
charge injection only results in a change in the common-mode voltage. A differ-
ential voltage change only results from charge-injection mismatch. With this
compensation, a reduction in offset of at least 10 9 can be expected.
2.4.2 Charge Injection and Clock Feed-Through in Chopper

Amplifiers
For chopper amplifiers, residual offset is mainly caused by the following three
issues:
• Non-idealities in clock timing
• Demodulated clock feed-through current spikes
• Impedance mismatch between two input nodes
Firstly, the non-idealities in the clock timing, such as clock skew, non-overlap
and overlap in chopper clocks introduce residual offset. Clock skew is a phe-
nomenon in which the two complementary chopper clocks switch at the different
transition moments, as shown in Fig. 2.11a. Assuming the offset is 10 mV and the
clock skew is 0.01 %, the resulting offset is 1 lV. To ensure a perfect offset
clk1 clk2
clk1 50% 50% clk1 clk1
clk2 49.99% 50.01% clk2 clk2
(a) (b) (c)
Fig. 2.11 Non-idealities in clock timing a Clock skew b Non-overlap clock c Overlap clock
fCH C2
C4
R1 C1 C3
+ +
+ + - + +
Vin Ioffset V1 G1 V2 Vout LPF VLPF
- + - -
- - -
R2
CH1 CH2
Fig. 2.12 Charge injection model in a chopper amplifier
cancellation, the two complementary chopping clocks must both exhibit 50 % duty
cycle and have transitions at the same moments.
Having transitions at the same moment, means that both non-overlap and
overlap of the complementary chopper clocks must be avoided. An NMOS
chopper is driven by the non-overlap chopper clocks, as depicted in Fig. 2.11b. A
small time gap exists when clk1 and clk2 are both low, hence the differential signal
paths are being interrupted. This leaves the signal path not attenuated and may
cause glitches at the output of the amplifier in between the choppers.
For the overlap chopper clocks (Fig. 2.11c), there is a small time interval when
both clocks are high, thus causing a ‘‘short circuit’’ between the differential signal
paths. This causes low input impedance and also shorten the amplification time
between the choppers. Thus, the effective gain of the amplifier reduces, resulting in
increased noise and offset.
Complementary chopper clocks with a 50 % duty cycle and the same transitions can
be produced by a non 50 % duty cycle clock and a divider-by-two D-flipflop, as will be
described in Sect. 4.6.5. Extra buffers can be added to reduce the rise and fall time.
Secondly, due to clock feed-through, the imbalance of parasitic capacitors in the
choppers also causes a residual offset [11]. Figure 2.12 illustrates the charge
injection due to imbalanced parasitic capacitances of the input and output choppers
in a fully-differential chopper amplifier. Figure 2.13 depicts a zoom-in picture of
the input and output choppers in which the current spikes caused by imbalanced
parasitic capacitances are illustrated.
IDC,offset fCH fCH fCH fCH

IAC,spike IAC,spike
C11 C31
0
0 0 M31
+ M11 C22 + + C32 M32 +

M22
Vin V1 V2 C42 Vout
M12 C41 M42
C12
C21 - M41
- - -
M21
Input chopper CH1 Output chopper CH2

(a) (b)
Fig. 2.13 Current spikes caused by imbalanced parasitic capacitances in the input and output
choppers CH1 and CH2 that give rise to a residue offset
As shown in the input chopper CH1 (Fig. 2.13a), at the transition moments of
the chopper clocks, due to clock feed-through the mismatch between the capaci-
tances C11 and C12 causes an AC current spike at the node of V1+. For the same
reason, the mismatch between C21 and C22 also leads to another AC current spike
at V1-. The difference between these two current spikes results in an AC current
spike, as shown in Fig. 2.13a at V1. This AC current spike is rectified by CH1,
which appears as a DC spike current at the input of CH1, with an average value
given by:
Ioffset;DC ¼ 2ðDC1 DC2 Þ Vclk fCH ð2:10Þ
where DC1 = C11-C12, and DC2 = C21-C22, Vclk is the amplitude of the clock
signal, and fCH is the chopper frequency. The DC current spike contributes to the
input offset current Ioffset of the amplifier. This current goes through the series
impedance of the chopper and the input signal source, leading to a rectified input
voltage spike. The average DC value of the spike results in a residual offset
VOS, res1, as given by:
VOS;res1 ¼ 2ðR1 þ R2 Þ ðDC1 DC2 Þ Vclk fCH ð2:11Þ
where (R1 ? R2) is the input impedance including on-resistance of the chopper
switches and the impedance of the signal source. DC1 and DC2 are the mismatch of
clock feed-through capacitances, which is mainly due to the overlap capacitances
of the clock line and the source terminals of switches in the input chopper CH1. If
DC1 = DC2, then no residual offset occurs since it will only result in a common-
mode spike.
DC3 and DC4 are the mismatch of clock feed-through capacitances due to the
overlap capacitances of the clock lines and the amplifier G1 output. They will also
cause an AC current spike at V2 (Fig. 2.13b). To provide this AC current spike, the
input of G1 needs to generate an AC voltage spike. This voltage spike is
fCH C
C C
R1 Ibias C
+ +
+ + - + +
Vin V1 G1 V2 Vout LPF VLPF
- + - -
- - -
R2 Ibias
CH1 CH2
Fig. 2.14 Residual offset due to the bias current and impedance mismatch in a chopper amplifier
demodulated by the input chopper towards the input, resulting in a residual offset
VOS, res2 as given by:
2ðDC3 DC4 ÞVclk fCH
VOS;res2 ¼ ð2:12Þ
G1
where DC3 = C31-C32, DC4 = C41-C42, and G1 is the transconductance of the
chopped amplifier. It can be seen that an amplifier with a higher transconductance
is less vulnerable to the mismatch of DC3 and DC4. For example, a 10 kHz
chopper frequency with 1/G1 = 5 kX, no source impedance, and a 5 V driving
clock voltage would result in a residual offset per unit of capacitance mismatch
between DC3 and DC4 of 0.5 l V/fF. The total residual offset due to clock feed-
through is the sum of the offsets given in (2.11) and (2.12).
Thirdly, the source impedance mismatch (DR = R1-R2) causes another residual
offset. The charge injection and clock feed-through due to chopping action cause two
s (denoted as Ibias in Fig. 2.14) which both flow out of the amplifier’s input in the
same direction [12]. For bias current calculation, if DC1 = DC2 = DC as shown in
Fig. 2.14, the average value of the bias current Ibias (Fig. 2.14) can be calculated as:
Ibias;DC ¼ 2DC Vclk fCH ð2:13Þ
The mismatch between input impedances R1 and R2 is DR, so these two bias
currents also generate residual offset, given by:
VOS;res3 ¼ 2ðR1 R2 Þ DC Vclk fCH ð2:14Þ
If DC is 1fF, the chopper frequency fCH is 10 kHz with a 5 V chopper clock,
then according to (2.13), the resulting bias current is 0.1 nA. From (2.14), if the
mismatch between R1 and R2 is 100 kX, these two bias currents flow through these
two resistors (Fig. 2.14), resulting in an extra residual offset of 10 lV.
It can be seen from (2.11), (2.12) and (2.14) that, the charge injection and clock
feed-through of the choppers, cause input bias current, offset current and hence the
residual offset. These three errors can be minimized by:
• Decreasing the chopping frequency
• Decreasing the chopper clock amplitude
• Balancing or minimizing the overlap capacitors between the clock lines to the
input and output terminals of G1
(a) L H H L
+ + - - - - -
Vin A V1 V2 V3 Vout
- + + + + LPF +
-
(b) V V1 /A
spike
0 t
-Vspike
Vspike V2 /A Vos,res,H
0 t
-Vspike
V3 /A
Vspike fH=3•fL
0 t
-Vspike
Fig. 2.15 a Nested chopper amplifier b Charge-injection spikes from the high-frequency
choppers are chopped by the low-frequency chopper pair
• Ensuring matched and balanced input impedance for differential paths reduces
residual offset caused by the input bias current
In Chap. 4, a chopper layout that minimizes the charge injection and clock feed-
through will be presented.
2.4.3 Chopper Charge Injection Suppression Techniques
Besides minimizing the charge injection and clock feed-through with the afore-
mentioned methods, there are several techniques that can be used to suppress the
demodulated clock feed-through spikes.
Nested-Chopper Technique
Since the residual offset of a chopper amplifier is proportional to the chopper
frequency fch, as expressed by Eqs. (2.11), (2.12) and (2.14), it can be decreased by
reducing fch. However, fch cannot be lower than the 1/f noise corner, otherwise
1/f noise can not be completely removed. The nested chopper technique solves this
problem by using an extra pair of choppers that run at a much lower frequency.
The residual offset of the amplifier chopped by a high frequency chopper clock UH
is chopped out by a low-frequency chopper clock UL [13] (Fig. 2.15).
(a) V1
0 fch 3fch 5fch 7fch
(b) fch
R1
+ + - + - -
+
or
Vin V1 A1 A2 Vout
- - - + - + LPF +
R2
Fig. 2.16 a Modulated signal and spike harmonics of V1 b Chopper amplifier with an LP or BP
amplifier to remove clock feed-through spikes
The overall residual offset is only limited by the charge injection in the low-
frequency chopper, and therefore it is reduced by a factor fH/fL, where fH and fL are
high and low chopping frequencies, respectively.
The implementation of the nested chopper is very simple: only one extra pair of
choppers and a low-frequency clock signal are needed. A disadvantage of this
approach is that the useable signal bandwidth is reduced, since it is limited by fL
rather than fH. However, this is not a problem for bridge read-out applications.
Nested chopping can be used to chop the complete read-out chain which consists
of a preamp and a DR ADC. The low-frequency chopper spikes at fL can be filtered
out in the decimation filter following the DR ADC [14]. The nested chopping will
be applied to the read-out IC design, as will be discussed in Chap. 5.
Filtering of Spike Harmonics
The chopper clock is a square-wave signal that contains odd harmonics at fch,
3fch, 5fch, etc. Most of the energy of the chopper ripple is located at the first
harmonics [4]. Therefore, a low-pass (LP) or band-pass (BP) filter can be incor-
porated between the chopper switches to filter out the chopper harmonics at the
high frequencies, at the cost of a small reduction in the signal bandwidth [4]
(Fig. 2.16a). A LP filter was implemented in [15] to filter the spikes, achieving a
5 lV offset (Fig. 2.16b). A BP filter implementation is presented in [16] to sup-
press the DC offset. Here, the chopping frequency is designed to track the center
frequency of the BP filter. It achieves an offset of less than 600 nV. However, a
disadvantage of this technique is the significant amount of extra circuitry required.
Chopper with Guard Band
Another approach to filter out clock feed-through spikes is to introduce a small
guard time in the output chopper switch that prevents the spikes caused by the
input chopper from being demodulated, as shown in Fig. 2.17b. This technique has
Fig. 2.17 a Modulated fD

(a) fM
amplifier b Guard band clock
diagram + + + - - -
Vin M Va A1 Vb D Vout
- - - + + +
(b) fM
Va
fD
Vout
t
Average is 0
been used in [17–19] for custom sensor interfaces. An average offset of 200 nV
has been achieved in [18]. However, the residual offset with the guard band
technique is limited by the matching between the shape of the spike and the guard
time delay. Moreover, the output signal is no longer continuous-time due to the
gap in the guard time since Vout just holds the value before the guard time starts,
thus incurring a slight loss of gain and noise aliasing.
2.4.4 Conclusions
For the auto-zeroed amplifier, charge injection determines the residual offset. In
the chopper amplifier, the residual offset is caused by the non-ideality in clock
timing, the demodulated clock feed-through current spikes and the impedance
mismatch between two input nodes. To conclude, symmetry, matching and bal-
ancing the parasitics are essential to achieve low residual offset in chopper
amplifier.
2.5 Dynamic Offset Compensated Operational Amplifiers
This section discusses the basic principle of feedback and then reviews several
precision operational amplifiers that employ the dynamic offset cancellation
techniques discussed above, i.e. auto-zeroing and chopping. These amplifier
topologies include ping-pong, offset stabilization and chopper offset
stabilization.
Fig. 2.18 Block diagram of

E
a general feedback system +
X(s) H(s) Y(s)
–
2.5.1 Feedback
Feedback is a powerful technique that has found wide application in high-precision

signal processing. Figure 2.18 shows a negative feedback system, where H(s) and
b denote the high gain amplifier and the passive feedback network, respectively.
Part of the output signal is redirected back to the input and compared to the
incoming signal. The loop accurately controls the output to produce an amplified
or processed replica of the input signal.
The error, which is indicated as E in Fig. 2.18, is given by:
1
EðsÞ ¼ XðsÞ bYðsÞ ¼ XðsÞ ð2:15Þ
1 þ bHðsÞ
From (2.15), the higher the loop gain bH(s), the more effectively the error E is
minimized. This means that a feedback loop with high loop gain can be applied to
reduce an error.
This feedback concept will be used in the work presented in Chaps. 4 and 5 to
reduce the offset and gain error of an amplifier. Generally, the feedback concept is
used in operational amplifiers to obtain an accurate gain determined by 1/b. The
next few sections will review several precision operational amplifier topologies.
2.5.2 Ping-Pong Operational Amplifier
As discussed before, the auto-zeroing technique is not directly suitable for use in a
continuous-time general purpose amplifier, since the amplified output signal is
only available during one half of the clock period. To obtain a continuous-time
output signal, the ping-pong technique can be used. This involves the use of two
auto-zeroed amplifiers [6], as shown in Fig. 2.19. When one amplifier is being
auto-zeroed, the other is being used to amplify the signal. The same output stage is
shared by the two auto-zeroed amplifiers. Furthermore, the combination of auto-
zeroing and chopping is employed to achieve a noise PSD of 20 nV/HHz from DC
to 1.5 kHz, which rises to 48 nV/HHz at 20 kHz. It consumes a supply current of
800 lA.
Figure 2.20 shows the noise spectrum of chopping, auto-zeroing and the
combination of these two. As seen from Fig. 2.20b, chopping modulates low-
frequency 1/f noise to the chopping frequency, thus achieving a clean and flat noise
2.5 Dynamic Offset Compensated Operational Amplifiers 39
ping stage
ch ch
A A
+ + -
+ + +
Vin Z G1 Vb1 Va Gout Vout
- +
- - - -
Z
+
G3 -
Vc1 C1 C2
A A
+ -
+
Z G2 Vb2
- +
-
pong stage Z
+
G4 -
Vc2 C3 C4
Fig. 2.19 Basic concept of the ping-pong operational amplifier
Noise Noise
PSD PSD
f (Hz) fch f (Hz)

(a) (b)
No chopping or Chopping
auto-zeroing
Noise Noise
PSD PSD
fAZ f (Hz) fAZ fch f (Hz

2fAZ
(c) (d)
Auto-zeroing Chopped
Auto-zeroing
Fig. 2.20 Input-referred noise PSD of chopping, auto-zeroing and the combination of the two
(fch = 2fAZ)
Fig. 2.21 Simplified fch

chopper-CDS block diagram Vo C cds
Vin Vos Vout
+ A1 CDS A2
2 1
fch
1 CS
1 2
spectrum at low frequencies, but with ripple at the chopping frequency. Auto-
zeroing involves sampling, thus causing increased noise at DC due to aliasing
(Fig. 2.20c). In their combination, since the input stage is chopped at twice the
auto-zeroing frequency, this noise is then modulated away from DC to fch (or 2fAZ)
[6] (Fig. 2.20d).
A disadvantage of the ping-pong technique is that spikes are introduced because
the voltages Vb1 and Vb2 at the output of the first stage amplifiers have to switch
between the offset compensating voltages Vc1, Vc2 and the voltage required at the
input of the output amplifier Va. This results in spikes at the output. This effect can
be reduced by replacing C1-C4 with active integrators with the same input CM
voltage as the output stage Gout [1]. However, spikes still remain because
switching occurs within the signal path.
2.5.3 Chopper-CDS Operational Amplifier
As a hybrid of chopping and auto-zeroing, a ripple-free operational amplifier is

proposed in [20] that uses input chopping and correlated double sampling (CDS)
for demodulation. As shown in Fig. 2.21, the AC-coupling capacitor Ccds is
inserted between the first and the second stage, removing the offset from the first
stage without causing ripple. The CDS then demodulates the signal back to DC.
Since CDS operates on the modulated input signal, the folded noise spectrum
due to sampling is also around the fCDS, which is equal to the chopping frequency.
This implies that the gain of A1 around the chopping frequency fch should be large
since it suppresses the folded noise of CDS. The DC gain of A1 should be low
enough so that the amplified offset and 1/f noises do not saturate A1. A1 is
implemented with a band-pass response that exhibits low gain around DC but high
gain at fch. In the worst case, the amplified offset can still be too large. A band-
reject passive feedback network is used that decreases the gain at low and high
frequencies, but not at the chopping frequency. In this way, the DC gain of A1 is
reduced further. However, this scheme increases the circuit complexity. It achieves
a 2 lV offset and a noise PSD of 37nV/HHz with an NEF of 5.5.
R1
R2
Vos CM1
+ + + - +
Vin Va Gm2 Gm1 +
- -
- Vout
-
- + +
Gm4 Gm3 +
+ - -
Fig. 2.22 The concept of the offset-stabilized operational amplifier
2.5.4 Offset-Stabilized Operational Amplifiers
Offset stabilization is another technique that can be used to design a precision

wideband amplifier. The basic concept of offset-stabilization is shown in Fig. 2.22.
A main operational amplifier Gm2 with an offset Vos, is being offset-stabilized by a
stabilizing amplifier Gm4 with a hypothetical zero offset. Gm3 acts as an auxiliary
input of the main amplifier. The stabilizing amplifier Gm4 applies a voltage to the
inputs of Gm3, which drives a current to the output of Gm2 to compensate for its input
offset voltage [4]. The feedback resistors R1 and R2 determine the gain of the
amplifier.
With the negative feedback configuration, the differential input voltage Va of
Gm2 is approximately equal to the offset Vos. The residual offset due to the finite
gain of the combined amplifier can then be expressed as:
Am2
Vos;res Vos ð2:16Þ
Am4 Am3
where Am4, Am3 and Am2 are the DC voltage gains of the stabilizing amplifier Gm4,
the main amplifier Gm3, and the auxiliary input of the main amplifier Gm2,
respectively. Equation (2.16) indicates that the combined voltage gain of Gm4 and
Gm3 has to be much larger than the voltage gain of the main amplifier Gm2. If Gm3
is 50 times lower than Gm2, then to reduce a 10 mV worse-case offset to 1 lV, Gm4
must have a minimum voltage gain of 114 dB. To maintain stability, frequency
compensation must also be implemented.
This topology can also be seen as a multi-path amplifier in which the cascode of
Gm4, Gm3 and Gm1 form the high-gain low-frequency path, and the main amplifier
Gm2 and Gm1 form the low-gain high-frequency path. Low-frequency character-
istics will, therefore, be determined by the low frequency path. In other words, the
low-frequency noise, residual offset and gain accuracy are determined by Gm4,
while the unity gain frequency is determined by the main amplifier Gm2. To
achieve lV-level offset, the offset-stabilization loop needs to be chopped or auto-
R1
R2 Vos1 CM1
+ + + - +
Vin Va Gm2 Gm1 +
- -
- Vout
-
CH1 Vos3 CH2
+ - - + +
Gm4 Gm3 +
+ - -
LPF
Fig. 2.23 Chopper offset-stabilized amplifier
zeroed. Chopping is more power-efficient than auto-zeroing, therefore chopper

offset-stabilized amplifiers will be discussed in the next section.
2.5.5 Chopper Offset-Stabilized Operational Amplifiers
Figure 2.23 shows a chopper offset-stabilized amplifier. Since Gm4 determines the
low-frequency noise and offset of the overall amplifier, Gm4 is chopped to elim-
inate its 1/f noise and offset. The chopper amplifier composed of chopper CH1,
stabilizing amplifier Gm4, and chopper CH2 senses the offset of the main amplifier
Gm2. A LPF suppresses the chopper ripple due to the chopped offset of Gm4. The
residual offset due to finite gain is expressed by (2.16).
The effects of chopping on the noise of chopper offset-stabilized amplifier are
depicted in Fig. 2.24. The offset and 1/f noise of Gm4 are modulated to the
chopping frequency fch, and then removed by the LPF. For effective suppression of
1/f noise, the bandwidth of the stabilizing loop as well as the chopper frequency fch
should, therefore, be larger than the 1/f noise corner frequency of the main
amplifier. To let the low-frequency noise be dominated by the low-frequency path,
the –3 dB frequency of the LPF should be chosen higher than the 1/f corner fre-
quency of the main amplifier Gm2, and the chopper frequency should be high
enough and thus the chopper ripple can be filtered out properly.
The LPF that filters out the chopper ripple (Fig. 2.23) can be implemented in
several ways, e.g. a continuous-time integrator, a sample-and-hold notch filter, or a
continuous-time notch filter, which will be described as follows.
2.5.5.1 Continuous-Time Integrator
One way to implement the LPF is by using the integrator composed of Gm5, C51
and C52 to filter out the chopper ripple [21], as shown in Fig. 2.25. It depicts a
Noise Noise Noise

PSD PSD LPF PSD
f (Hz) fch f (Hz) fch f (Hz)

(a) (b) (c)
Noise of Gm3 Noise of chopped Noise of chopper offset-
Gm3 and LPF stabilized amplifier
Fig. 2.24 Noise in chopper-stabilized amplifiers
R1
C11
R2 Vos2
+ + - + -
+
Vin Gm2 Gm1
- - + - + Vout
C31 -
C51 C12
Vos4 Cp4 Vos5
+ - + - + -
Gm4 Gm5 Gm3
- + - + - +
CH1 CH2
C32 C52
Fig. 2.25 Chopper offset-stabilization amplifier using an active integrator and multi-path hybrid-
nested Miller compensation
multi-path architecture that employs this technique in combination with hybrid

Miller compensation [22, 23]. The gain stages Gm2 and Gm1 form the high-fre-
quency low-gain path, while the transconductances Gm4, Gm5, Gm3 and Gm1 form
the high-gain low-frequency path. To realize a frequency response with a smooth
roll-off, the unity gain frequency of both paths should be [23].
Gm4 Gm2
f0dB ¼ ¼ ð2:17Þ
2pC3 2pC1
where C3 and C1 are the values of capacitors C31 (or C32) and C11 (or C12),
respectively.
The modulated offset Vos4 of Gm4 is filtered by the integrator consisting of Gm5,
C51 and C52. However, the integrator needs very large capacitors to obtain a low
cut-off frequency. Hence in reality, a residual chopper ripple in the form of a
triangular wave appears at the integrator output. Furthermore, due to the action of
the chopper CH2, the offset of Gm5 (Vos5) appears as a square-wave voltage over
the capacitor Cp4, charging and discharging this capacitor. The resulting
R1
C11
R2 Vos2
+ + - + -
+
Vin Gm2 Gm1
- - + - + Vout
1 2
-
C51 C53 C31
C12
Vos4 Cp4
+ - + -
Gm4 C3 Gm3
- + - +
CH1 CH2
C54 C32
2 1
C52
Fig. 2.26 Multi-path operational amplifier with an embedded sample-and-hold circuit
alternating current at the output of Gm4 is demodulated by chopper CH1, causing a

residual offset.
To eliminate this residual offset, the parasitic capacitor Cp4, or the offset Vos5 of
the integrator, must be minimized. The parasitic capacitor Cp4 can be minimized
by choosing small dimensions for the transistors that are connected to the output
terminals of Gm4. Furthermore, Cp4 can be minimized to ensure a fully symmetric
and balanced layout. To reduce offset Vos5, auto-zeroing can be used [21].
However, it increases the complexity and power consumption of the design. Since
the focus of this thesis is to design a low- noise amplifier with good noise-power
efficiency, the topology with Gm5 auto-zeroed will not be elaborated further.
2.5.5.2 Sample-and-Hold Notch Filter
An alternative solution for implementing the LPF in Fig. 2.23 are to use a swit-
ched-capacitor (SC) sample-and-hold circuit to sample the chopper ripple at the
output of the integrator, as shown in Fig. 2.26. It shows an operational amplifier
with multi-path hybrid-nested Miller compensation. A LPF is implemented with a
SC notch filter consisting of the switches driven by U1 and U2 and the capacitors
C53 and C54 [24]. The switches sample the chopper ripple at the zero-crossing
points. The notch positions of the Sinc filter are located at multiples of the chopper
frequency, and thus are accurately determined by the chopping clock.
This notch filter acts as a passive integrator. To compensate for the extra pole
introduced by the notch filter, the capacitors C51 and C52 are introduced for the
same reason as the hybrid-nested Miller compensation [23]. Capacitors C31 and
C32 are theoretically not needed, but they help to maintain local loop stability. The
capacitor C3 helps to limit the bandwidth of the low-frequency path so that the
delay caused by the notch filter does not cause instability [24].
However, this technique still involves sampling, and so still incurs a certain
noise-folding. More importantly, the sample-and-hold filter exhibits a Sinc
R1
C3
R2
+ +
+
Vin Gm5 - Gm4
- - Vout
-
C1a +
A1
CH1 CH2
+
+ - +
Continuous-time
+
Gm1 Gm2 notch filter Gm3
- + - -
-
C1b
Fig. 2.27 Multi-path operational amplifier with an embedded continuous-time notch filter
filtering response, creating a significant phase shift at the chopping frequency.

Overcoming these phase shift complicates the frequency compensation in the
amplifier.
2.5.5.3 Continuous-Time Notch Filter
As a counterpart of SC implementation, the notch filter also can be implemented in

a continuous-time (CT) fashion. Figure 2.27 shows a CT notch filter incorporated
in the low-frequency path to filter out the chopper ripple [25] in a multi-path
operational amplifier. Additionally, a buffer A1 is used to allow feedback through a
compensation capacitor C1a for the low-frequency path, thus preventing the
chopper ripple at the output of CH2 from feeding forward through C1a to the output
of the amplifier.
The disadvantage of this approach is that the notch filter suppresses the ripple in
an open-loop structure. This implies that the chopping frequency derived from the
time constant of the on-chip relaxation oscillator needs to be well-matched to the
notch position of the CT filter, which is also determined by the product of certain
resistors and capacitors. This frequency tracking is not an issue for the SC notch
filter [24], since the notch positions are precisely determined by the chopping
clock. Furthermore, if a tunable chopping frequency is desired, a phase-locked
loop (PLL) is then required to track the external chopping frequency to ensure that
the RC time constant of the CT notch filter closely tracks the locking frequency.
Otherwise, trimming is required to tune the notch location in the CT notch filter.
2.5.5.4 Auto-Correction Feedback Loop
Another way to suppress the chopper ripple is to use an auto-correction feedback

(ACFB) loop to null the offset in a chopper multi-path operational amplifier [26, 27].
R1
C1
R2
+ +
+
Vin Gm6 - Gm5
- - Vout
-
C32
CH1 CH2
+ - +
Gm1 Gm2
- + -
C31
- + - +
Gm4 Gm3
+ - + -
SC Notch CH3
Filter
Fig. 2.28 Multi-path operational amplifier with an auto-correction feedback loop
Its block diagram is shown in Fig. 2.28. Unlike the designs presented in [24, 25]
which employ notch filters in the signal path to suppress the chopper ripple, this
approach uses a feedback loop outside the signal path. Therefore, it does not cause
any phase shift in the signal path. However, the stability of the feedback loop itself
needs to be taken care of. This is because the notch filter creates a significant phase
shift at the chopping frequency, and the unity gain frequency of the ACFB loop must
occur well below the chopping frequency to ensure the loop stability. Increasing the
unity-gain frequency speeds up the settling of the loop. However, a higher chopping
frequency is then required, thus increasing the charge injection and the offset.
Moreover, since the sensing points of the loop are the virtual ground of Gm2,
they are relatively ‘‘quiet’’. The DC gain of the loop is limited because of the small
ripple excitation. A ripple reduction of only 43 dB [26] is achieved. Furthermore,
the SC notch filter (NF) causes sampling noise at DC. This noise is modulated by
CH2 and creates a peak output noise PSD around the chopping frequency.
2.6 Conclusions
Table 2.2 summarizes the performances of precision operational amplifiers that

achieve lV-level offset. They apply different techniques to suppress chopper
ripple. One technique is to use auto-zeroing to reduce the initial offset of the
amplifier [6, 21]. However, the increased low-frequency noise due to noise folding
leads to a noise penalty, i.e. extra power dissipation is needed to meet a given
noise specification (NEF = 21.8 and 153). A band-pass filter can be implemented
between the choppers to suppress the DC offset, so as to eliminate the output
2.6 Conclusions 47
Table 2.2 Comparison of precision operational amplifiers

Tang [6] Witte [21] Burt [24] Kusuda [27] Belloni [20]
Year 2002 2006 2006 2011 2010
Noise PSD (nV/HHz) 20 15 55 5.9 37
Chopping frequency 15 32 125 200 500
(kHz)
Offset (lV) 3 1.5 3 0.78 1.94
Input bias current (pA) 40 – 70 72 –
GBW (kHz) 2500 1370 350 4000 260
Supply current (Iq)(lA) 800 700 17 1470 14.4
Die area(mm2) 0.67 3.6 0.7 1.26 1.14
NEF [28] 21.8 153 8.7 8.7 5.5
GBW/Iq (kHz/lA) 3.1 2 20.6 2.7 18
Rail sensing capability No Yes Yes Yes Yes
Ripple reduction Auto- Auto- SC notch Auto Chopper-
technique zeroing zeroed filter feedback loop CDS
integrator
chopper ripple [16]. However, the chopping frequency needs to track the center
frequency of the band-pass filter, which requires significant amount of extra
circuitry.
A switched-capacitor [24] or a continuous-time notch filter [25] can be embedded
in a multi-path offset stabilized operational amplifier to reduce the chopper ripple.
However, the SC notch filter [24] involves sampling thus causing noise folding. The
issue associated with the CT notch filter [25] is that the notch filter suppresses
the ripple in an open-loop structure. To effectively suppress the chopper ripple, the
notch frequency of the CT filter needs to closely track the chopping frequency,
which could be limited by the RC spread in the CT notch filter. Another technique
uses an auto-correction feedback loop [26] to suppress the chopper ripple. However,
since the ripple sensing points are at the ‘‘quiet’’ virtual grounds of the output stage,
the limited loop gain restricts the ripple suppression ratio.
In addition, the notch filters generate excess phase shift, meaning that the
chopper clock frequency must be relatively high ([ 125 kHz) to maintain stability
in the signal path [24, 25] or in the feedback loop [26]. Such a high chopping
frequency increases charge injection errors, and hence increases input offset, given
the same noise level and process parameters. The chopper-CDS scheme [20] uses
an AC-coupled capacitor to block the offset, thus generating no chopper ripple.
However, this technique also necessitates a high chopping frequency of 500 kHz,
resulting in a relatively low input impedance and a high input bias current.
Therefore, innovative solutions need to be explored to eliminate the chopper
ripple without causing the above-mentioned issues: noise aliasing, frequency
tracking, limited loop gain, and excess phase shift (high chopping frequencies).
To counteract these problems, a new ripple reduction technique will be proposed
in Chap. 4.
References
1. Witte JF (2008) Dynamic offset compensated CMOS amplifiers. Ph.D. Thesis, Delft
University of Technology, The Netherlands
2. Razavi B (2001) Design of analog CMOS integrated circuits. McGraw-Hill Companies, Inc.,
New York
3. Sansen WMC (2006) Analog design essentials. Springer, Dordrecht
4. Enz CC, Temes GC (1996) Circuit techniques for reducing the effects of op-amp
imperfections: auto-zeroing, correlated double sampling, and chopper stabilization. Proc.
IEEE 84(11):1584–1614
5. Poujois R, Borel J (1978) A low drift fully integrated MOSFET operational amplifier. IEEE J
Solid-State Circuits 13:499–503
6. Tang ATK (2002) A 3 lV-offset operational amplifier with 20nV/HHz input noise PSD at
DC employing both chopping and auto-zeroing. In Proceedings of the IEEE ISSCC, Dig.
Tech. Papers, pp 386–387
7. Enz CC (1989) High precision CMOS micropower amplifiers. Ph.D. Thesis, Thesis no. 802,
Ecole Polytechnique Federale de Lausanne (EPFL). Available at: http://library.epfl.ch/en/
theses/?nr=802,1989
employing ping-pong auto-zeroing and chopping. In Proceedings of the IEEE ISSCC, Dig.
employing ping-pong auto-zeroing and chopping. IEEE J Solid-State Circuits 45(10):
2044–2056
10. Wegmann G, Vittoz EA, Rahali F (1987) Charge injection in analog MOS switches. IEEE J
Solid-State Circuits, SC-22(6):1091–1097
11. Witte JF, Makinwa KAA, Huijsing JH (2004) The effect of non-idealities in CMOS chopper
amplifiers. In Proceedings of the ProRISC 2004, pp 616–619
12. Fan Q, Huijsing JH, Makinwa KAA (2011) Analysis of input impedance, input bias and offset
current of a chopped multi-path current feedback instrumentation amplifier. In Proceedings of
the IWASI, pp 61-66
13. Bakker A, Thiele K, Huijsing JH (2000) A CMOS nested-chopper instrumentation amplifier
with 100-nV offset. IEEE J Solid-State Circuits 35(12):1877–1883
14. McCartney D, Sherry A, Sherry et al (1997) A low-noise low-drift transducer ADC. IEEE J
15. Enz CC, Vittoz EA, Krummenacher F (1987) A CMOS chopper amplifier. IEEE J Solid-State
Circuits, SC-22(3):335–342
16. Menolfi C, Huang Q (1999) A fully integrated, untrimmed CMOS instrumentation amplifier
with submicrovolt offset. IEEE J Solid-State Circuits 34(3):415–420
17. Bilotti A, Monreal G (1999) Chopper-stabilized amplifier with a track-and-hold signal
demodulator. IEEE Trans Circuits Syst I:490–495
18. Huang Q, Menolfi C (2001) A 200nV offset 6.5nV/HHz noise PSD 5.6 kHz chopper
instrumentation amplifier in 1 lm digital CMOS. In Proceedings of the IEEE ISSCC, Dig.
19. van der Meer JC, Riedijk FR, van Kampen E, Makinwa KAA, Huijsing JH (2005) A fully
integrated CMOS hall sensor with a 3.65 lT 3r offset for compass applications.
In Proceedings of the IEEE ISSCC, Digital technical papers, pp 246–247
20. Belloni M, Bonizzoni E et al (2010) A micropower chopper-CDS operational amplifier. IEEE
J Solid-State Circuits 45(12):2521–2529
21. Witte JF, Makinwa KA, Huijsing JH (2007) A CMOS chopper offset-stabilized opamp. IEEE
J Solid-State Circuits 42(7):1529–1535
22. Huijsing JH, Fonderie MJ, Shahi B (2007) Frequency stabilization of chopper-stabilized
amplifiers. United States Patent Nr. 7,209,000, Apr. 2007
References 49
23. Huijsing JH (2011) Operational amplifiers theory and design, 2nd Edition, Springer, Berlin
24. Burt R, Zhang J (2006) A micropower chopper-stabilized operational amplifier using a SC
notch filter with synchronous integration inside the continuous-time signal path. IEEE J.
25. Luff GF (2010) Chopper stabilized amplifier. United States Patent, US 7724080:B2, May
2010
26. Kusuda Y (2010) Auto correction feedback for ripple suppression in a chopper amplifier.
IEEE J Solid-State Circuits 45(8):1436–1445
27. Kusuda Y (2011) A 5.9nV/HHz chopper operational amplifier with 0.78 lV maximum offset
and 28.3nV/C offset drift. In Proceedings of the IEEE ISSCC, Dig. Tech. Papers,
pp 242–243
28. Steyaert MSJ, Sansen WMC, Chang Z (1987) A micro power low-noise monolithic
instrumentation amplifier for medical purposes. IEEE J Solid-State Circ SC-22(6):1163–1168
Chapter 3
Current-Feedback Instrumentation
Amplifiers and Gain Accuracy
Improvement Techniques
As discussed in Chap. 1, compared to other instrumentation amplifier topologies,

the current-feedback instrumentation amplifier (CFIA) is more suitable for bridge
read-out because of its high CMRR [1, 2], rail-sensing capability [1], high input
impedance and power efficiency [3, 4].
The main disadvantage of a CFIA is its limited gain accuracy, which is
determined by the mismatch between the input and feedback transconductors
[1, 5]. These also limit its input range and linearity. Although the limited input
range is often not a problem for bridge readout applications, where the input signal
is in the millivolt range, it limits the usefulness of the CFIA in other applications.
This section will review several previous precision CFIAs. It will be shown that
the gain accuracy and linearity of a CFIA can be improved by using resistor-
degeneration and dynamic element matching (DEM).
3.1 Current-Feedback Instrumentation Amplifier
Figure 3.1 shows a simplified block diagram of a current-feedback instrumentation

amplifier (CFIA). The input transconductor Gm2 and feedback transconductor Gm3
convert the input and feedback voltages into the corresponding currents. Their
difference is then nulled by the gain of Gm1. If the loop gain is high enough, the
overall feedback ensures that the output currents of Gm2 and Gm3 cancel out each
other and thus the closed-loop gain of the amplifier is given by:
Gm2 R1 þ R21 þ R22
Gain ¼ ð3:1Þ
Gm3 R1
In the CFIA, the signal is carried in the current domain. This is a very important
point, because summation or subtraction in the current domain is much easier than that
in the voltage domain. It opens various possibilities for compensation or calibration of
some error sources, such as offset; these will be discussed in Chaps. 4, 5 and 6.

52 3 Current-Feedback Instrumentation Amplifiers
Fig. 3.1 Current-feedback

instrumentation amplifier R21 Vout+
Vin+ + - + -
G m2 G m1 R1
Vin- - + - +
R22
Vout+
Vfbk+ + +
G m3
Vfbk- - -
From (3.1), the gain accuracy of the CFIA will be mainly determined by the
mismatch of the input and feedback transconductances, provided that precision
(0.01 %) gain-setting resistors are used. Furthermore, the CFIA’s input range and
linearity will also be limited by these transconductors. Available techniques to
address these issues will be discussed later in this chapter.
Next, let us review the history of CFIAs. They can be classified into two
categories: indirect current-feedback and direct current-feedback instrumentation
amplifiers. Note that the CFIA is also referred to as ‘‘differential differential
amplifier’’ (DDA) in [6] and ‘‘differential difference amplifier’’ (DDA) in [2].
3.1.1 Indirect Current-Feedback Instrumentation Amplifier
The first CFIA [7] was reported in 1971 and uses indirect current-feedback
(Fig. 3.2). Two high-gain voltage amplifiers A1 and A2 force the current Vin/R1
flowing though the degeneration resistors R1 to also flow through R2. The
amplitude of the latter current is equal to Vfbk/R2. Therefore, the gain is mainly
determined by the resistor ratios:
Vout R2 R3 þ R4
¼ ð3:2Þ
Vin R1 R4
Since the input and feedback transconductances isolate the input and feedback CM
voltages, a high CMRR ([120 dB) can be achieved. However, the drawback of this
topology is that the feedback loop is rather complex, consisting of the input trans-
conductance, voltage amplifiers A1 and A2, and the feedback transconductance. The
multiple stages within the feedback loop leads to stability issues and slow settling.
To overcome the stability issue, the next two topologies use only one high-gain
amplifier. Figure 3.3 shows an indirect CFIA topology [1] with one feedback
amplifier A1. The current sources I1, I2, I3 and I4 are nominally equal, as are I5 and
I6. The loop amplifier A1 forces the differential current flowing through R1 to also
flow through R2 by applying the required voltage Vfbk to the inputs of Q21 and Q22.
3.1 Current-Feedback Instrumentation Amplifier 53
Fig. 3.2 CFIA using two VDD

feedback amplifiers
I3 I3 I3 I3
+
A1 Vout
-
+
A2
-
R3
Q21 Q22
Vin Q11 Q12
R2 +
R1
Vfbk R4
-
Vref
VSS
Fig. 3.3 Indirect current- VDD

feedback instrumentation + Vout
A1
amplifier I1 I2 I3 I4
-
R3
R1 R2 +
Vin Vfbk R4
Q11 Q12 Q21 Q22 -
I5 I6
VSS
Due to the nonlinearity of the differential pairs, the currents flowing through
Q11 and Q12 and of Q21 and Q22 are not linear functions of Vin and Vfbk,
respectively. From (3.1), this non-linearity is cancelled to first order when the
differential pairs are nominally identical, i.e. the transistors match and R1 = R2.
The closed-loop gain is then given by
Vout R3 þ R4
¼ ð3:3Þ
Vin R4
An important advantage of the indirect CFIA topology is that by using PNP (or
PMOS) differential pairs as the input and feedback transconductors, the common-
mode (CM) input range includes ground [1].
3.1.2 Direct Current-Feedback Instrumentation Amplifier
Another category of CFIA uses direct current feedback, as shown in Fig. 3.4.
Here, the input and feedback transconductances share the same bias current, thus
Fig. 3.4 Direct current- VDD

feedback instrumentation
amplifier I3 I4
+ Vout
A1
-
Vin Q11 Q12
R1 R3
R1 +
Q21 Q22 Vfbk R4
-
Vref
I3 I4
VSS
the input and feedback signal currents directly compensate each other. This is in
contrast with the indirect CFIA of Fig. 3.3, in which the two differential pairs do
not share the same bias current. Because of this, the direct CFIA has been used in
low-power biomedical applications where low power consumption is critical [9].
This topology was later implemented in CMOS [10]. However, the stacking of two
transconductance stages increases the minimum supply voltage and reduces the
input CM range, which cannot include ground [1]. Furthermore, since I3 and I4 are
equal, the currents flowing through Q11 and Q12 are also equal. However, the
current flowing through Q21 and Q22 is signal-dependent, which leads to signal
dependent nonlinearity.
In summary, the direct CFIA is suitable for low-power design, while the indirect
CFIA is better for low-voltage designs and for applications that require a large
input CM range that may include one of the supply rails. As discussed in Chap. 1,
bridge sensors require read-out electronics to have a CM input range that includes
at least one rail. Therefore, the work presented in this thesis will mainly focus on
the indirect CFIA. For simplicity, we will refer to this as a ‘‘CFIA’’ in the rest of
this thesis.
3.2 Precision Current-Feedback Instrumentation Amplifiers
The CFIA can be seen as a merged version of the two-opamp topology in which
the output stages are shared. Thus, the dynamic offset cancellation techniques
applied in precision operational amplifiers can also be used in CFIAs. This section
reviews several CFIAs that employ chopping, auto-zeroing and offset stabilization
techniques to mitigate low-frequency errors.
3.2 Precision Current-Feedback Instrumentation Amplifiers 55
C31
C21
C11
+ + - + - + - Vout
Vin Gm31 Gm2 Gm1
- - + - + - +
C22 C12
+ + -
Vfbk Gm32
- - +
C51
CH2 Cpar CH1 Vos5
+ - + - + -
Gm61 Gm5 Gm4
- + - + - +
CH3 C51
+ - C32
Gm62
- +
Fig. 3.5 Chopper-stabilized CFIA in a multi-path architecture [11]

Amplifier
The chopper-stabilized operational amplifier (Fig. 2.25) can be transformed into a

chopper-stabilized CFIA [11], as shown in Fig. 3.5. The resulting CFIA has a low-
frequency path (LFP) and a high frequency path (HFP). The low-frequency
characteristics, such as noise, offset and gain accuracy, are determined by the low-
frequency path, while the high-frequency path dominates the noise after the turn
over frequency.
To reduce the offset of the low-frequency path, the input and feedback trans-
conductors Gm61 and Gm62 are chopped, thus giving rise to a chopper ripple.
Although this ripple is partially filtered by the integrator built around Gm5, C51 and
C52, there is still a residual chopper ripple at the integrator output. To eliminate
this, Gm61 and Gm62 are auto-zeroed to reduce their initial offset, as shown in
Fig. 3.6. The auto-zeroing capacitor CA1-CA4 isolates the input and feedback CM
voltage from the CM of the Gm61 and Gm62. However, the disadvantage is that the
input choppers CH2 and CH3 together with the auto-zeroing capacitors CA1-CA4
constitute a switched-capacitor circuit with an impedance of 1/F1CA, where F1 is
the auto-zeroing frequency, and CA is the value of CA1-CA4. Thus, the input
impedance decreases compared to the case when Gm61 and Gm62 are not auto-
zeroed.
The offset of Gm5 also causes a residual input offset [11]. This offset, being
chopped by CH1 (Fig. 3.6), charges and discharges the parasitic capacitors at the
C51
CH2 CA1 Vos7 Cp5 CH1 Vos6
+ + - + - + - + - Vout
Vin Gm61 Gm5 Gm4 Gm2,1
- - + - + - + - +
CA2
C52 Vfbk
CH3 CA3 Vos8

+ + - CA3
Vfbk Gm62
- - + + - + -
CA4 CA4 Gm8 CA5 Gm7
- + - +
FC F1
F1 FC F1 F1 F1 F1
FC
FC
F1
F1
t
Fig. 3.6 The low-frequency path of a chopper offset-stabilized CFIA and associated timing
diagram [11]
output of Gm61 and Gm62, thus appearing as a square-wave voltage there. After
being chopped again by CH2 and CH3, this square wave translates into an extra
input-referred error. To mitigate this problem, Gm5 is auto-zero offset-stabilized, as
shown in Fig. 3.6. Finally, a 2.5 lV offset is achieved. Gm61 (or Gm62) has a
transconductance of 220 lA/V, which corresponds to a noise PSD of 20 nV/HHz.
Since the noise of Gm61 and Gm62 are uncorrelated, the noise PSD increases by a
factor of H2. Furthermore, because they are auto-zeroed half of the time, the noise
PSD increases by another H2 [11]. As a result, the measured noise PSD is 42 nV/
HHz consuming a 325 lA supply current (NEF = 29.2).
3.2.2 Ping-Pong Auto-Zeroed Current-Feedback Instrumentation

Amplifier
The auto-zeroed ping-pong operational amplifier (Fig. 2.19) can also be used to
build a CFIA [5], as illustrated in Fig. 3.7. The input and feedback transconductors
in the ping stage and pong stage are auto-zeroed to reduce the initial offset of the
3.2 Precision Current-Feedback Instrumentation Amplifiers 57
ch
A
ping stage A
ch
+ + - +
Vin Z Gin Gout Vout
- - + -
A slow-settling R1
ch
offset-nulling loop
A Z
+ - - +
Z Gfbk GAZ
- + + - +
CAZ
Vfbk R2
pong stage
- Vref
Fig. 3.7 CFIA employing ping-pong auto-zeroing and chopping [5]
amplifier. To mitigate the noise folding inherent to auto-zeroing, a slow-settling

offset-nulling loop is applied to limit the noise bandwidth to less than half of the
auto-zeroing frequency [5]. The noise is then modulated away from DC by
chopping the input stages at half the auto-zeroing frequency, reducing the low-
frequency noise to the white noise level. However, due to its ping-pong topology,
this amplifier needs four low-noise Gm stages, so its power consumption is rather
high. It consumes a supply current of 1.7 mA with a noise PSD of 27nV/HHz
(NEF = 43).
3.2.3 Conclusions
As shown in Fig. 3.6, the use of auto-zeroing in the input stages of the low
frequency path of the multi-path CFIA increases the noise PSD by a factor of H2
compared to the case without auto-zeroing [11]. This implies that to achieve the
same noise level, it needs to consume twice as much power. In [5], besides the
factor H2 increase in the noise PSD, four low-noise transconductance stages are
used in a ping-pong auto-zeroed topology, which is much less power-efficient than
the chopped CFIA. Table 3.1 summarizes the performance of these two CFIAs.
Chapter 4 will present a CFIA that employs only chopping to eliminate the
1/f noise and offset [3]. To suppress the associated chopper ripple, a continuous-
time technique is proposed that enables a significant improvement in power
efficiency.
Table 3.1 Comparison of prior art CFIAs

Witte [11] Pertijs [5]
Year 2009 2010
Supply voltage 5V 5V
Supply current 325 lA 1.7 mA
Input noise PSD 42nV/HHz 27nV/HHz
CMRR 130 dB 142 dB
PSRR 114 dB 138 dB
Gain error 0.05 % (Untrimmed) 0.1 % (Untrimmed)
Offset \2.5 lV \2 lV
GBW 640 kHz 800 kHz
GBW/Isupply (kHz/lA) 2 0.5
NEF [3.9] 29.2 43
3.3 Gain Accuracy Improvement Techniques
The previous section mainly discussed how to eliminate the offset and 1/f noise in
CFIAs. Besides that, however, gain error is another dominant error source at low
frequencies. Figure 3.8 shows the output error versus input signal amplitude. For
small input signals, offset dominates; while for large input signals, gain error
dominates.
In a CFIA, the gain error is mainly due to the mismatch between the input
and feedback transconductors (Fig. 3.9). This can be as much as 2 % over
temperature and process. For instance, in a CFIA with a closed-loop gain of 100,
a gain error of 2 %, an offset of 2 lV and a 5 V supply, the minimum input
signal X that ensures the gain error is larger than the output referred offset, is
given by
X 100 2 % 2lV 100: ð3:4Þ
Thus, X C 100 lV, meaning that when the input signal is larger than 100lV,
the effect of gain error becomes larger than that of offset. With a 50 mV full scale
input, the output-referred gain error is 100 mV, which is much larger than the
output-referred offset of 200 lV. Apart from limited gain accuracy, the linear
range of the CFIA is also limited to less than 100 mV due to the input and
feedback transconductors (using simple differential pairs).
3.3 Gain Accuracy Improvement Techniques 59
Fig. 3.8 Output gain error Output

and offset vs. input signal error
amplitude Gain error
Offset error
0 Input signal ampitude (V)
Fig. 3.9 Gain error sources

in the CFIA C2
Vin Gm3 Gm2 Vout

R1
Vfbk
Vfbk Gm4 R2
⎛ Gm3 ⎞ ⎛ R1 + R2 ⎞
Gain = ⎜⎜ ⎟⎟ × ⎜ ⎟
⎝ Gm 4 ⎠ ⎝ R2 ⎠
Precision external
Dominates ! !
resistors 0.01%
2%
(or thin film)

with Resistor-Degenerated Input Stages
To reduce transconductor mismatch and increase linear input range, one common
approach is the use of resistor-degenerated differential pairs [5, 12]. The detailed
schematic of a NMOS input resistor-degenerated stage is shown in Fig. 3.10. Its
input CM range includes the positive supply rail. Figure 3.11 shows a PMOS input
resistor-degenerated stage which has ground-sensing capability.
Taking Fig. 3.10 for example, local negative feedback loops maintain the
NMOS input transistors M1 and M4 at a CM independent drain-source voltages
[13]. Within the loops, M3 and M6 keep the drain currents of the input transistors
(M1 and M4) constant. Folded cascode transistors M2 and M5 ensure that the input
voltage includes the supply rail. The input transistors M1 and M4 level-shift the
input common-mode voltage down by a gate-source voltage, and reproduce the
differential input voltages across resistors R1 and R2. Thus, the signal current
flowing through R1 and R2 is given by:
Vinþ Vin
Idgn ¼ ð3:5Þ
R 1 þ R2
Fig. 3.10 Schematic of VDD

composite NMOS resistive-
degenerated differential pair I1 I1
I3
VB1
M2 M5
Vin+ Vin-
M1 R1 R2 M4
M3 M6
Idgn
Vout+
Vout-
I2 R3 R4 I2
VSS
Fig. 3.11 Schematic of VDD

composite PMOS resistive-
degenerated differential pair I1 I3 I1
Vin+ Vin-
R1 R2
M1 M3 Idgn M6 M4
M5
M2 VB2
Vout+
Vout-
R3 R4
I2 I2
VSS
Transistors M3 and M6 feed this current to the output load resistors R3 and R4.
From (3.5), it can be seen that the transconductance of this stage is mainly
determined by the degeneration resistors R1 and R2.
With careful layout, or by trimming, the degeneration resistors can be made to
match better than the transistors. Gain errors of 0.1 % have been achieved by both
auto-zeroed ping-pong CFIA (NEF = 43) [5] and chopper-stabilized CFIA
(NEF = 153) [12]. As can be seen from the NEFs of these designs, the use of
resistor degeneration leads to a significant loss in power efficiency. This is because
of the extra current required to bias transistors M3 and M6 and the cascode tran-
sistors M2 and M5. All these transistors contribute to noise, and the transcon-
ductance of this stage decreases due to degeneration.
C41
CH2 S511 CA1 Vos61 S513 Cp5 CH1 Vos5
+ + - + - + - + -
Vin Gm61 Vout
VCMi5 Gm5 Gm4 Gm2,1
- - + - + - + - +
S512 CA2 S514
+ C42 Vfbk
VCal Ccal
- +
- Gm63
VRef + -
CH3 S521 CA3 Vos62 S523

+ + -
Vfbk VCMi5 Gm62
- - +
S532 CA4 S524
Fig. 3.12 Chopper-stabilized IA with auto-zero sense amplifiers and auto-gain calibration

Amplifier with Auto-Gain Calibration
The different input and feedback CM voltages of a CFIA can be isolated by

capacitors [11], as shown in Fig. 3.6. The auto-zeroing capacitors CA1-CA4 block
the DC CM voltage. The CM voltage at the inputs of Gm61 and Gm62 is, therefore,
set by the auto-zeroing action and so is the same as their output CM voltage,
which, in turn, is regulated by a CM feedback loop [11]. The voltage non-linearity
of CA1 to CA4 gives rise to a gain error that is a function of the CM voltage [11].
However, this error is less than 0.05 % over a CM range from 0 to 3.6 V.
To further improve gain accuracy, the gain of the low-frequency path of a
chopper-stabilized CFIA (Fig. 3.6) also can be calibrated to equalize the trans-
conductance of Gm61 and Gm62, as proposed in [14]. This is shown in Fig. 3.12, for
simplicity, the auto-zero stabilization loop that could be used to eliminate the
offset Vos5 is omitted here.
There are three operation modes for Gm61 and Gm62: offset auto-zeroing,
auto-gain calibration and normal operation. Figure 3.12 shows the amplifier in the
auto-gain calibration mode. In this mode, both inputs of Gm61 and Gm62 are
connected to a calibration voltage VCal and the output currents of these stages are
compared. The difference of the output currents is then integrated by CCal to a DC
voltage. This voltage feeds to Gm63, which converts it into two differential currents
that fine-tune the tail currents of Gm61 and Gm62 to compensate their mismatch.
Thus the suppression ratio for the mismatch is determined by the gain within the
calibration loop.
Note that the stabilization loop accurately controls the gain at low frequencies.
The gain at high frequencies is determined by the difference of the main input
stages Gm31 and Gm32 in Fig. 3.5, which is not auto-calibrated. However, this is not
C41
CA1 Vos51 S511 Vout
+ + - + -
Vin VCMi5 Gm1 Gm6
- - + - +
CA2 S512 -
Vfbk
C42
+
CA3 Vos52 S521 VoRef
+ -
VCMi5 Gm2
- +
+ CA4 S522
Vfbk
-
CA3 Vos52 S531

+ -
Vrefin VCMi5 Gm3
- +
CA4 S532
Fig. 3.13 Ping-pong-pang auto-zeroed CFIA [16]
a problem because at high frequencies the gain is not accurate anyway because of
the drop in overall loop gain.
Since this approach needs an auto-gain correction phase, its continuous-time
operation is interrupted. This implies that it can not compensate for temperature
dependent gain drift.
3.3.3 Ping-Pong-Pang Current-Feedback Instrumentation

Amplifier
Another method to improve the gain accuracy of the CFIA is to use dynamic
element matching (DEM). An example of this is the ping-pong-pang (PPP) auto-
zeroed CFIA [15, 16], whose simplified schematic is shown in Fig. 3.13.
The three transconductances Gm1, Gm2 and Gm3 are dynamically switched
between the input, feedback, and auto-zeroed stages with the algorithm shown in
Table 3.2 [16]. The average effect of DEM modulates the Gm mismatch to the
DEM frequency, thus ensuring good accuracy at low frequencies.
This concept was first implemented in [17], around the same time as the work
presented in Chap. 5. As shown in Fig. 3.14, the three transconductors are alter-
nately switched between input, feedback and auto-zeroing states. While one
transconductor is auto-zeroed, the other two provide the output signal (Fig. 3.14).
Since each transconductor spends equal time in the input and feedback states, their
mismatch is dynamically averaged out.
Table 3.2 Ping-pong-pang algorithm with DEM

1 2 3 4 5 6 1
Vin Gm2 Gm4 Gm4 Gm3 Gm3 Gm2 Gm2
Vfbk Gm3 Gm3 Gm2 Gm2 Gm4 Gm4 Gm3
VCal Gm4 Gm2 Gm3 Gm4 Gm2 Gm3 Gm4
Fig. 3.14 Block diagram of the ping-pong-pang CFIA with the timing diagram [17]
To analysis the effect of DEM on gain error, a simplified block diagram of a

CFIA that applys DEM to its input and feedback transconductors to average out
their Gm mismatch is shown in Fig. 3.15. The Gm ratios corresponding to the two
DEM phases are given by
Gm3 1
Gm ratio1 ¼ ¼ ð3:6Þ
Gm4 1 þ D
Gm4 1 þ D
Gm ratio2 ¼ ¼ ð3:7Þ
Gm3 1
where D is the initial mismatch between Gm3 and Gm4. The Gm ratios will then
toggle between two levels: (1 ? D)/1 and 1/(1 ? D), causing the CFIA output to
toggle between (1 ? D)Videal, out and Videal,out/(1 ? D). The average gain error
after applying DEM is given by [18]
1þD 1
1 þ 1þD D2
Gm ratioavg ¼ 1þ
2 2 ð3:8Þ
ðforD\\1Þ
(1+Δ)Videal,out
Swapper C1 Videal,out
Vin Gm3 Gm1 Vout
R1 Videal,out /(1+Δ)
Vfbk
R2
Vfbk Gm4
Gm4 = Gm3 (1+ )
Fig. 3.15 CFIA with DEM ripple after applying DEM
Equation (3.8) shows that the use of DEM reduces the initial gain error from D
to D2/2. So for a typical Gm mismatch of 2 %, the use of DEM will reduces this to,
a negligible 0.02 %. In a practical CFIA, however, the common-mode (CM)
dependence of Gm3 and Gm4 may still limit gain accuracy. If the input and output
common-mode voltages are not the same, in a typical situation in IA applications,
The actual transconductance of, say, the one that connects to the input CM will
vary by an extra amount mismatch Dcm and this mismatch will always appears in
the input path during the two DEM swap phases. The average gain error, after
applying DEM, is then given by [18]:
1þDþDcm
1 þ 1þD
1þD
cm
D2 D Dcm
jGain Errorj ¼ j1 j þ þ Dcm
2 2 2 ð3:9Þ
ðforD\\1Þ
From (3.9), the CM-dependent mismatch Dcm will not be suppressed by the
DEM. For simplicity, this issue will be neglected in the following analysis.
However, a circuit-level technique to mitigate this problem will be described in
Sects. 5.6 and 6.6.1.
The accurate gain achieved in the low frequency, however, is at the cost of a
DEM ripple at the amplifier output due to the modulated Gm mismatch. It is
assumed that the closed-loop gain of the CFIA is 1/b, where b is the feedback
factor determined by the gain-setting resistors, i.e. b = R2/(R1 ? R2). The CFIA
output amplitude during the two DEM phases is given by
1 1þD
Vout1 ¼ Vin ð3:10Þ
b 1
1 1
Vout2 ¼ Vin : ð3:11Þ
b 1þD
Therefore, the amplitude of the DEM ripple is given by [3.18]
1
VDEM;ripple ¼ Vout1 Vout2 ¼ 2D Vin ¼ 2D Videal;out : ð3:12Þ
b
Before After
trimming trimming
C2
Swapper
Vin Gm3 Gm2 Vout
R1
Vfbk
R2
Vfbk Gm4
Trimming
Fig. 3.16 Trimming in a CFIA employing DEM
where Videal,out is the ideal output voltage of the CFIA. Equation (3.12) indicates
that the DEM ripple is a product of the mismatch D and the output signal. With a
mismatch of 0.5 % and a 4.5 V output signal, the amplitude of the DEM ripple can
be as large as 45 mV, and therefore must be suppressed.
To reduce this signal-dependent ripple, the transconductors are trimmed by a
5-bit current DAC [17]. As depicted in Fig. 3.16, the trimming mechanism
involves monitoring of the output DEM ripple, and then trimming a current-DAC
that fine-tunes the tail current of Gm3 and Gm4 to compensate for their mismatch.
When the output DEM ripple is reduced to zero, trimming is accomplished. The
disadvantage of trimming is the increased production costs. More importantly,
trimming is a one-time calibration, thus it will not compensate for temperature
drift.
Compared to the previous ping-pong CFIA [5] (Fig. 3.7), this amplifier uses
three transconductors rather than four. Furthermore, its transconductors are not
degenerated. Thus, this amplifier achieves a 2 9 improved power efficiency
(NEF = 24). Overall, this CFIA achieves a gain error of 0.04 % and a noise PSD
of 28nV/HHz, while drawing a 480 lA supply current.
3.3.4 Conclusions
Gain accuracy is the major disadvantage of the CFIA, and is caused by the mis-
match between its input and feedback transconductances. Furthermore, its non-
linearity and limited range is limited due to the input and feedback
transconductors. To improve these, three approaches can be taken. The first
method involves the use of resistor-degenerated input stages [5, 12] to shift
transconductance matching to resistive matching. As a result, a gain error of 0.1 %

and improved linearity has been achieved. However, resistor-degeneration leads to
a noise or power penalty. The second approach uses discrete-time auto-gain cal-
ibration in an auto-zeroed amplifier. However, it is not a continuous-time tech-
nique, unless it is applied in a ping-pong topology. The third method is to apply
DEM to the three stages in a ping-pong-pang CFIA. As presented in [17], such
CFIA can achieve a gain error of 0.04 %. To eliminate the signal-dependent DEM
ripple, trimming can be applied [17]. However, it increases the production cost and
cannot compensate for the temperature drift.
Chapter 5 will present a chopper CFIA that employs DEM to eliminate the gain
error. To suppress the associated DEM ripple, a continuous-time technique is
proposed, thus eliminating the need of trimming. These techniques result in a
significant improvement in power efficiency.
References
1. van den Dool BJ, Huijsing JH (July 1993) Indirect current feedback instrumentation amplifier
with a common-mode input range that includes the negative rail. IEEE J Solid-State Circuits
28(7):743–749
2. Chan PK, Ng KA, Zhang XL (2004) A CMOS chopper-stabilized differential difference
amplifier for biomedical integrated circuits. In Proceedings of the The 47th IEEE international
midwest symposium on circuits and systems (MWSCAS), III-33-6, vol 3
amplifier with a 1 mHz 1/f noise corner and an AC-coupled ripple reduction loop. IEEE J
4. Fan Q, Huijsing JH, Makinwa KAA (2010) A 21 nV/HHz chopper-stabilized multipath
current-feedback instrumentation amplifier with 2 lV offset. In Proceedings of the IEEE
5. Pertijs MAP, Kindt WJ (2010) A 140 dB-CMRR current-feedback istrumentation amplifier
employing ping-pong auto-zeroing and chopping. IEEE J Solid-State Circuits 45(10):2044–
2056
6. Huijsing JH (1981) Comparative study of some types of differential-differential amplifiers. In
Proceedings of the European conference on electrotechnics, Eurocon, B 6-8(1)(2), pp 22–26
7. Krabbe H (1971) A high-performance monolithic instrumentation amplifier.In Proceedings of
the IEEE ISSCC, digital technical papers, pp 186–187
8. Hamstra GH, Peper A, Grimbergen CA (1984) Low-power low-noise instrumentation
amplifier for physiological signals. Med Biol Eng Comput 22:272–274
instrumentation amplifier for medical purpose. IEEE J Solid-State Circuits SC-22(6):1163–
1168
10. Yazicioglu RF, Merken P, Puers R, van Hoof C (May 2007) A 60 lW 60 nV/HHz readout
front-end for portable biopotential acquisition systems. IEEE J Solid-State Circuits
42(5):1100–1110
11. Witte JF, Huijsing JH, Makinwa KAA (2009) A chopper and auto-zero offset-stabilized
CMOS instrumentation amplifier. In Proceedings of the IEEE symposium on VLSI circuits,
pp 210–211
References 67
with 5 lV offset for bidirectional high-side current-sensing. IEEE ISSCC, digital technical
papers, pp 74–75
13. Huijsing JH, Shahi B (2007) Accurate voltage-to-current converters for rail-sensing current-
feedback instrumentation amplifiers. United States Patent, Nr. 7,2022,738, Apr. 2007
14. Boucher RE, Huijsing JH (2008) Auto-gain correction and common mode voltage
cancellation in a precision amplifier. United States Patent, application no. 12/253620, Oct.
2008
15. Huijsing JH (2008) Instrumentation amplifiers developments. In Proceedings of the AACD
workshop, pp 105–126
16. Witte F (2008) Dynamic offset compenated CMOS amplifiers. PhD Thesis, Delft University
of Technology, The Netherlands
instrumentation amplifier with 0.04 % gain error. In procedings of the IEEE symposium on
VLSI circuits, pp 60–61
18. Wu R, Huijsing JH, Makinwa KAA (Dec. 2011) A current-feedback instrumentation
amplifier with a gain error reduction loop and 0.06 % untrimmed gain error. IEEE J Solid-
State Circuits 46(12):2794–2806
Chapter 4
A Chopper Instrumentation Amplifier
with Offset Reduction Loop
This chapter discusses the design and implementation of a chopper current-feed-

back instrumentation amplifier (CFIA). This amplifier can be used in stand-alone
sensor read-out systems that need to drive an external analog-to-digital converter
(ADC).
Firstly, the requirements on the amplifier are described. It is targeted for
thermistor read-out applications in wafer steppers (see Chap. 1). The design of the
CFIA is then discussed. Both the input and intermediate stages of the CFIA are
chopped to achieve a low 1/f noise corner. To reduce chopper ripple, a continuous-
time (CT) offset reduction loop (ORL) is proposed. Due to its CT nature, it does
not cause noise folding, thus offering improved power efficiency over auto-zeroed
amplifiers. It will be shown that this ORL can be applied to both general-purpose
instrumentation amplifiers and operational amplifiers.
Measurement results show that the CFIA achieves a 1 mHz 1/f noise corner at a
noise PSD of 15 nV/HHz while consuming only 230 lA supply current (NEF =
8.8). To our knowledge, this represents the best LF noise performance ever
reported for a stand-alone CMOS instrumentation amplifier.
4.1 Amplifier Requirements
In Chap. 1, a challenging interface application was described: the read-out of a

precision thermistor bridge used in wafer steppers. The output of the thermistor
bridge is only ±24 mV, thus it needs to be amplified before it is digitized or
processed further. This requires the use of a low-noise instrumentation amplifier
(IA). The challenges associated with the design of such an amplifier are summa-
rized below.

70 4 A Chopper Instrumentation Amplifier with Offset Reduction Loop
Table 4.1 Key requirements Specifications Value

of the precision
instrumentation amplifier for Supply voltage 5V
thermistor bridge readout Gain tolerance 0.5 %
Input referred noise PSD 15 nV/HHz
1/f corner frequency 1 mHz
Offset 5 lV
Offset drift 20 nV/C
CMRR [120 dB
PSRR [120 dB
Input common-mode 0–3 V
Output common-mode 2.5 V
Input impedance [20 MX
Supply current \300 lA
The first challenge is to achieve low thermal and 1/f noise. The thermistor
bridge used in the wafer stepper application has a noise PSD of 1m nV/HHz, and
so the amplifier’s noise PSD was chosen to be at roughly the same level, i.e.
16 nV/HHz. As mentioned in Chap. 1, to achieve high power efficiency, the
amplifier’s noise should be white in the bandwidth of interest, which means that
the amplifier should have a 1/f noise corner of a few mHz. To justify such low
noise specifications, the amplifier must also have high CMRR ([120 dB) and
PSRR ([120 dB).
Secondly, in addition to low thermal and 1/f noise, the read-out electronics must
achieve low offset and good gain accuracy to maintain system accuracy over
temperature. The offset and offset drift aim to achieve less than 5 lV and 20 nV/C.
Since the gain error of the thermistor bridge is about 0.5 %, the amplifier’s gain
accuracy does not need to be much better. Good gain drift can be achieved by using
careful layout to match the input and feedback transconductors of the CFIA. This is
our first approach.
The third challenge is the need for the amplifier to accommodate different input
and output common-mode (CM) voltages. The input CM is at 0.61 V since the
bridge is biased at 1.22 V. The output CM is at 2.5 V, since the amplifier’s output
is to be digitized by an Analog-to-Digital converter (ADC) with a 0–5 V input
range.
The fourth challenge is to minimize the amplifier’s power consumption to
reduce self-heating. This should not be larger than that of the bridge (a few
hundreds of lA) to restrict local self-heating errors.
Finally, the amplifier should have a high input impedance ([20 MX), so as not
to attenuate the sensor signal, and it should be able to drive a 50 pF load capacitor,
just like most general-purpose amplifiers. The target specifications of the amplifier
are summarized in Table 4.1.
4.2 Amplifier Architecture 71
Fig. 4.1 Simplified block

diagram of CFIA Vin Vout
Gm2 Gm1
R1
Vfbk
Vfbk
Gm3
R2
4.2 Amplifier Architecture
As discussed in Chap. 1, a CFIA has significant advantages over the classic three-
opamp IA because of its better power efficiency, high CMRR and rail sensing
capability. Moreover, it can easily handle different input and output CM voltages.
Compared with a switched-capacitor IA, a CFIA provides continuous-time signal.
Compared to a capacitively-coupled IA, a CFIA has higher input impedance and
does not produce glitches. These properties make a CFIA more suitable for use as
a stand-alone IA for bridge readout.
Figure 4.1 shows the block diagram of a CFIA. The input transconductor Gm2
and feedback transconductor Gm3 convert the input and feedback voltages into the
corresponding currents. Their difference is then nulled by the gain of Gm1. If the
loop gain is high enough, the overall feedback ensures that the output currents of
Gm2 and Gm3 cancel out each other and thus the closed-loop gain of the amplifier is
given by:

Gm2 R1 þ R2
Gain ¼ ð4:1Þ
Gm3 R2
Gain Accuracy
From (4.1), it can be seen that the amplifier’s gain accuracy will depend on the
open-loop gain of the CFIA and on the matching between the input and feedback
transconductance stages Gm2 and Gm3.
As mentioned in Chap. 1, the output range of the thermistor bridge is 48.6 mV.
To optimally map this range to the 0–5 V range of the ADC, the amplifier should
have a rail-to-rail output with a gain of 183. For an accuracy of 0.5 % at the
intended closed-loop gain of 183, the open-loop gain of Gm1 must be in excess of
90 dB, which is easily achievable with two stages of amplification. By restricting
the input and output CM voltages to the 0–3 V range, the transconductors can be
realized by PMOS differential pairs. With careful layout on the transconductors, it
should be possible to realize a CFIA that achieves better than 0.5 % gain accuracy.
Chopping Strategy
Since chopping is a continuous-time modulation technique that does not cause
noise folding, it is employed here to reduce the offset and 1/f noise of the amplifier.
Simulations show that the unchopped 1/f noise corner of the CFIA in the
C2
C1
f1 140dB f1 f2 50 dB f2
Cpar1
Vin Vout
Gm3 Gm2 Gm1
CH1
f1
CH3 CH4 CH5 R1
Vfbk Vfbk
Gm4
CH2 Cpar2
R2
Fig. 4.2 Three-stage CFIA with the input and intermediate stages chopped
implemented 0.7 lm CMOS technology is 10 kHz. To completely remove 1/

f noise, the chopping frequency should be well above the 10 kHz 1/f noise corner.
However, it should not be too high due to charge injection and clock feed-through
considerations. A chopping frequency of 30 kHz is chosen in this work. The goal
of this design is to achieve a 1 mHz 1/f noise corner.
To drive a 50 pF load, the output stage of the CFIA is implemented in class-AB
fashion. The large signals present in the class-AB stage means that it can not be
easily chopped. Its 1/f noise must therefore be suppressed by the effective gain of
the preceding stages at the chopping frequency. After setting the unity-gain
bandwidth (UGB) of the amplifier to be 800 kHz, the effective gain at the chop-
ping frequency can be translated to a DC gain requirement of the preceding stages.
Simulations show that the CFIA requires two chopped gain stages to provide
sufficient gain to shift the class-AB stage’s 1/f noise corner below 1 mHz. Thus, a
three-stage topology was chosen, in which the input and intermediate stages are
chopped to suppress their 1/f noise. To provide sufficient loop gain and to suppress
the noise from the intermediate stage, the input stages are designed to have a DC
gain of 140 dB. The intermediate stage is designed to have a DC gain of 50 dB.
To confirm the validity of this approach, simulations were done using the
periodic steady-state (PSS) and periodic noise analysis (PNOISE) tools of Spectre
RF [1]. Figure 4.2 shows the overall topology of the three-stage CFIA. The
amplifier’s simulated input-referred noise spectrum without chopping, with the
input stage chopped and with the input and intermediate stages chopped, is shown
in Fig. 4.3. Without chopping, the 1/f noise below 10 kHz is clearly visible. With
only the input stage chopped, a 1/f noise corner of 0.3 Hz was observed, which is
still too high for the targeted 1/f noise corner of 1 mHz. When both the input and
intermediate stages are chopped, the resulting noise spectrum has a 1/f noise corner
of 1 mHz.
Frequency Compensation
Nested Miller frequency-compensation [2] is used to maintain the stability of
the three-stage amplifier. It is designed to be stable for a closed-loop gain of [20
with a 50 pF capacitive load, since unity-gain stability is not required in this
application. The simulated open-loop frequency response of this amplifier is
4.2 Amplifier Architecture 73
1mHz 0.3Hz 10kHz
3mHz – 50mHz
Fig. 4.3 Simulated input noise PSD
Fig. 4.4 Simulated open-loop frequency response

shown in Fig. 4.4. It has an open-loop gain above 250 dB, which means that gain
errors due to finite DC loop gain are negligible.
Input Impedance
Configured at a gain of 100, the gain setting resistors R1 and R2 of the CFIA are
about 300 X and 30 kX, respectively (Fig. 4.2). Due to the action of input
choppers CH1 and CH2, the parasitic capacitors Cpar1 and Cpar2 at the input of Gm3
and Gm4 appear as switched-capacitor impedance at the input and feedback nodes
of the CFIA. Assuming Cpar1 = Cpar2 = 0.6 pF and chopping frequency f1 =
30 kHz, the input impedance is given by 1/(2f1Cpar1) = 28 MX, which is much
larger than the equivalent resistance of the feedback network (300 X). Since the
thermistor bridge has a resistance of 10 kX, the input impedance of the CFIA
causes a gain error of 0.036 %, which is negligible compared to the gain error of
the bridge itself (0.5 %).
4.3 Offset Reduction Loop
As discussed above, the input and intermediate stages of the CFIA are both
chopped to suppress their 1/f noise corner below 1 mHz. However, the amplitude
of the resulting chopper ripple can be quite large (*800 mV), limiting the
headroom. Thus, this ripple must be suppressed.
Several ripple reduction techniques were reviewed in Chap. 2. The discrete-
time sampling techniques [3, 4] involve noise folding, thus incurring a certain
noise penalty. Furthermore, the discrete-time [4] and continuous-time [5] notch
filters all generate excess phase shift in the signal path. In the auto-correction
feedback (ACFB) loop [6], the notch filter generates excess phase shift within the
feedback loop. To maintain stability, the ripple reduction techniques [4–6] need a
high chopping frequency to ease the frequency compensation, thus increasing the
charge injection and the offset. Moreover, the ACFB loop [6] suffers from limited
DC loop gain since the ripple sensing point is at the virtual ground of an amplifier.
In this work, an AC-coupled offset reduction loop (ORL) [7] is proposed that
avoids the foregoing issues.
4.3.1 Basic Concept
Figure 4.5 shows the conceptual diagram of the CFIA with the ORL. The ORL
synchronously demodulates the chopper ripple from AC to DC, averages it to
obtain a DC measure of the offset and then uses it to null the offset, hence the
ripple.
The ORL consists of sense capacitor C4, chopper CH6, integrator Gm6 with Cint
and compensation transconductance Gm5. In the start-up condition, the sense
4.3 Offset Reduction Loop 75
C2
C1
f1 f1 f2 f2 Initial
Vin Vos ripple
Gm3 Gm2 Gm1
CH1
V out
CH3 CH4 CH 5
f1 R1
Vfbk f1 Vfbk
Gm4 C4
CH 2 Gm5
Vo
∫ IDC CH6 IAC
R2
Fig. 4.5 Simplified block diagram of a three-stage CFIA with an AC-coupled ripple reduction
loop (ORL) in start-up condition
Chopper
clock
Chopper offset at the

output of CH3
Output ripple voltage
AC current in sense
capacitor C4
Rectified sense current

into integrator Gm6 and C
Fig. 4.6 Waveforms of chopper clock, chopped offset, output ripple voltage, AC current and
rectified sense current in the ORL
capacitor C4 converts the amplifier output ripple Vout,ripple into an AC current IAC.
The current amplitude is proportional to the derivative of Vout,ripple, given by
dVout;ripple
IAC ¼ C ð4:2Þ
dt
This AC current is demodulated by chopper CH6, and the resulting DC current
IDC is integrated by an integrator, generating a DC compensation voltage Vo that is
proportional to the ripple amplitude. This is then fed back via transconductance
Gm5 to the outputs of Gm3 and Gm4, injecting a current that compensates for the
offset between Gm3 and Gm4. Figure 4.6 shows the waveform of the chopper clock,
C2 C1
f1 f1 f2 f2 reduced
Vos ripple
Vin
Gm3 Gm2 Gm1
CH1
V out
CH3 CH 4 CH 5
f1 R1
Cint
Vfbk f1 Vfbk
Gm4 C4
CH 2 Gm5 Gm6 R2
Vo I DC CH6 I AC
Fig. 4.7 Simplified block diagram of a three-stage CFIA with ORL in steady-state condition
the chopper offset at the output of CH3, the output ripple, the sense current IAC
through C4 and the rectified current IDC.
At steady state, as shown in Fig. 4.7, the input offset VOS is precisely cancelled
by the ORL and the chopper ripple is strongly reduced. Since offset and the
compensation current are both DC signals, the offset reduction ratio is determined
by the DC loop gain in the ORL. The integrator in the ORL (Fig. 4.5) is built with
Gm6 and Cint, as shown in Fig. 4.7.
The synchronous demodulator formed by the chopper CH6 and the integrator
behaves like a narrow-band notch filter around the chopping frequency and its
harmonics. As a result, the ORL has little effect at frequencies near DC, and
therefore also little effect on the low-frequency response of the amplifier.
When input signal frequencies are close to the chopping frequency, however,
the output signal is AC-coupled via C4 into the synchronous demodulator, and then
demodulated to DC and fed back to the outputs of Gm3 and Gm4. This creates
notches in the amplifier closed-loop gain at harmonics of the chopping frequency
f1, 3f1, and 5f1…. The notch at f1 is visible in the closed-loop transfer function of
the amplifier [7], while the notches at the higher harmonics of f1 are suppressed by
the low-pass transfer function of the amplifier. Figure 4.8 shows the transfer
function of the amplifier with a notch at f1 due to synchronous demodulation in the
ORL. This notch will affect the amplifier’s step response as some ringing will
occur before the amplifier output settles. However, this is not a problem for bridge
sensor applications, because such sensors typically output millivolt-signals at
frequencies of a few Hz. For wide-band application, this notch can be buried in a
multi-path amplifier topology to ensure a smooth single-pole response, as will be
discussed in Sect. 4.5. The depth of the notch (Fig. 4.8) is determined by the
amount of signal that is fed back through the offset reduction loop, while the width
of the notch is determined by the unity-gain bandwidth of the ORL.
Fig. 4.8 Frequency response Frequency response

of ORL of the ORL
Notch bandwidth 2f0
Notch depth
f1 Frequency (Hz)
C2
Vos
f1 f1 Vout
Vin X C E
Gm3 G m21
CH1 IX − IC CH3 IE
Cint
f1 B C4
D
Gm5 Gm6
Vo Isc4 IB
CH6
Fig. 4.9 Simplified block diagram of a CFIA with an AC-coupled ORL
4.3.2 Transfer Function Analysis
The loop transfer function of the ORL can be derived with the help of the block
diagram shown in Fig. 4.9. In the forward path between nodes C and D (the
components enclosed by the dashed lines in Fig. 4.9), an input current IC is
chopped by CH3, integrated by C2, differentiated by C4, and then chopped by CH6
again.
For simplicity, the nodes D and E are initially considered to be ideal virtual
grounds. The relation between the current IE flowing into integrator C2 and the
current IB flowing into CH6 is then given by:
IB sC4 C4
K¼ ¼ ¼ ð4:3Þ
IE sC2 C2
Fig. 4.10 Equivalent circuit Cint

of the integrator Gm6 and Cint
I sc4
Vo Vi
Gm6
D Z sc4
Since this gain factor is not frequency-dependent, the operations of the two
choppers CH3 and CH6 around the integrator C2 and the differentiator C4 cancel
each other out. Hence, the relation between ISC4 and IC is also K:
ISC4 C4
K¼ ¼ : ð4:4Þ
IC C2
It should be noted that if nodes D and E are not ideal virtual grounds, there will
be a small error in the value of K expressed by (4.4), which will be neglected in
this analysis.
Let H(s) = V0/ISC4 be the transfer function of the integrator built around Gm6
and Cint (Fig. 4.9). If Gm6 has a finite DC voltage gain of A06, node D is no longer
an ideal virtual ground, and then:
V0 ¼ A06 Vi : ð4:5Þ
Since C4 is chopped by CH6, the switched-capacitor impedance ZSC4 looking
into the chopper from the non-ideal virtual ground of the integrator (node D) is
given by:
ZSC4 ¼ 1=f1 C4 ð4:6Þ
where f1 is the chopping frequency of CH6 and the output Vout is assumed to be a
virtual ground. The action of C4 and CH6 can then be modeled by the Norton
equivalent circuit shown in Fig. 4.10.
From Fig. 4.10, the input voltage Vi can be derived as:
Vi ¼ ISC4 ZSC4 þ ðVo Vi ÞsCint ZSC4 ð4:7Þ
By substituting (4.6) into (4.7), the transfer function H(s) of the integrator can
be calculated with:
V0 ZSC4 A06
HðsÞ ¼ ¼ ð4:8Þ
ISC4 1 þ sZSC4 ð1 þ A06 ÞCint
The loop gain L(s) of the ORL can be expressed as
LðsÞ ¼ K HðsÞ Gm5 : ð4:9Þ
Substituting (4.4) and (4.8) into (4.9), the loop gain becomes,
C
Open loop gain L(s ) (dB ) Closed loop gain of (dB )
X
fUGB f (Hz )
L(0) 0
1
0 L(0)
fpole fUGB f (Hz )
(a) (b)
Fig. 4.11 a Open-loop gain L(s) of the ORL b Closed-loop gain of C/X
C4 ZSC4 A06
LðsÞ ¼ Gm5 : ð4:10Þ
C2 1 þ sZSC4 ð1 þ A06 ÞCint
If A06 1 (4.10) can be simplified to
C4 ZSC4 A06
LðsÞ ¼ Gm5 : ð4:11Þ
C2 1 þ sZSC4 A06 Cint
The loop gain L(s) is plotted in Fig. 4.11a. It is a first-order low-pass function
with a dominant pole that is related to the chopping frequency by:
1 f1 C4
fdominant; pole ¼ ¼ ð4:12Þ
2pA06 Cint ZSC4 2pA06 Cint
The DC loop gain L(0) corresponding to the ripple-reduction ratio, i.e. the
offset-reduction ratio, is given by:
C4 A06 Gm5
Lð0Þ ¼ A06 Gm5 ZSC4 ¼ ð4:13Þ
C2 C2 f1
The phase shift within the loop mainly originates from three blocks: integrator
Gm21 and C2, a differentiator C4 and an integrator Cint. In this design, C4 = 5pF,
Cint =80 pF, f1 = 30 kHz, A06 is about 114 dB, so the dominant pole is at around
0.8 mHz. Since the phase shift of the first two blocks is cancelled out, the feedback
loop is a stable first-order system having a phase margin close to 90, i.e. it is
inherently stable.
The closed-loop gain C/X between nodes C and X in Fig. 4.9 is plotted in
Fig. 4.11b. The transfer function is given by
C 1 1 þ sZSC4 A06 Cint

¼
X 1 þ LðsÞ sZSC4 A06 Cint þ C4 ZSC4CA06 Gm5
2
ð4:14Þ
C2 1 þ sZSC4 A06 Cint
¼ ð Þ
C4 ZSC4 A06 Gm5 1 þ s CC2GCint
4 m5
This result indicates that the ORL effectively high-pass filters the offset and
1/f noise of the input stage. In this design, the DC loop gain is about 114 dB,
which, disregarding other contributors to output ripple, means that even the worst-
case ripple amplitude of 0.8 V should be reduced to microvolt levels. As seen in
this approach, the ripple sensing point is at the integrator output, where the ripple
is large.
The unity-gain bandwidth f0 of the loop can be derived from (4.11) by setting
L(s) =1
C4 ZSC4 A06
LðsÞ ¼ Gm5 ¼ 1: ð4:15Þ
C2 1 þ 2pf0 ZSC4 A06 Cint
Thus the unity-gain bandwidth of the ORL is given by:
Gm5 C4
f0 ¼ : ð4:16Þ
2pC2 Cint
Increasing the unity-gain bandwidth of the ORL reduces the settling time of the
loop. The parameter C2 is determined by the frequency compensation, Gm5 is
chosen much weaker than the input transconductance Gm3 to reduce noise con-
tribution from the ORL. The other two parameters C4 and Cint can be flexibly
chosen according to (4.16). In this design, C2 = 80 pF, Gm5 = Gm3/18 = 14 lA/V,
C4 = 5 pF and Cint = 80 pF, so the unity-gain bandwidth is 1.74 kHz. The notch
bandwidth is roughly equal to 2f0, i.e. about 3.5 kHz. It will be shown in Sect. 4.5
that this notch can be buried in a multi-path architecture.
In summary, the inherent stability of the ORL is the key advantage of the ORL
compared to other ripple reduction techniques [4–6]. This implies that the notch
width from (4.16) and the notch location, i.e., the chopping frequency, can be
independently chosen. Compared to a switched-capacitor notch filter [4] and auto
correction feedback [6], a relatively low chopping frequency can then be chosen,
leading to low offset without the stability problems caused by excessive notch-
filter phase shift. Moreover, the ripple sensing point is at the output of the
amplifier, where the ripple is quite large. Thus, the ripple suppression ratio of our
approach can be much larger than that of the ACFB loop [6].
4.4 Other Sources of Chopper Ripple 81
C2
C1
f1 f1 f2 f2
Vout,ripple
Vin Vos
G m3 G m2 Gm1 Vout
CH 1 CH 3 CH4 CH5
f1 R1
Cint
V fbk f1 V fbk
Gm4 C4
CH 2 G m5 Gm6 R2
Vint Vos6 I DC CH 6 B I AC
Fig. 4.12 Simplified block diagram of a three-stage CFIA with an AC-coupled ORL
4.4 Other Sources of Chopper Ripple
4.4.1 Cascode Buffer Isolation
The ORL suppresses chopper ripple originating from the input stage’s offset,
which is the major contributor to the output ripple. There is, however, a second
source of ripple which originates from the offset of the integrator’s amplifier Gm6,
as shown in Fig. 4.12. This is because the offset of the transconductance stage
Gm6, which is chopped by CH6, appears as a square wave voltage at node B. This
square wave appears across C4, and cannot be distinguished from the output ripple.
As a result, the ripple due to offset Vos6 will not be completely cancelled out.
The ripple due to the offset of VOS6 (Fig. 4.12) can be mitigated in two ways:
by autozero-stabilizing Gm6 or by using a current buffer to isolate CH6 from C4.
The auto-stabilization approach is shown in Fig. 4.13. The stabilization loop
eliminates the offset of Gm6 with a stabilization loop consisting of CA3, CA4,
integrator Gm8 and CA5 and transconductance Gm7 [8]. As can be seen, the
implementation becomes rather complicated, thus a cascode buffer is used here, as
shown in Fig. 4.14. Compared to C4, the cascode buffer 1 presents a much smaller
parasitic capacitance Cpar at the right side of chopper CH6. Cpar is around 0.6 pF,
while C4 is 5 pF, resulting in an 8-fold ripple reduction.
The ORL integrator was realized as a passive integrator built around a second
cascode buffer (CB2), because this only requires half the capacitor area required
by an active integrator. To minimize the effect of common-mode interference, the
CFIA was implemented in a fully differential manner, as depicted in Fig. 4.15.
The introduction of the ORL does not significantly affect the noise performance
of the amplifier. This is because CB2 and Gm5 are located between the choppers
CH3 and CH6 (Fig. 4.15) and so their 1/f noise contributions are chopped out.
C21 Vout+
C11 R21
CH1 CH3 CH 4 CH 5
Vin+ + - + + + - Vfbk+
Gm3 Gm2 2 Gm1 R1
Vin- - + - - - + Vfbk-
1st stage 2
nd stage 3rd stage
CH2 R22
Vfbk+ C12
+ + C22 Vout-
Gm4
Vfbk- - -
C51
Vos6 CH6 C 41
Gm5 Gm6
C 42
C52
CA4
- + - +
Gm7
+ - CA5 Gm8 + -
CA3
F1 F1 F1 F1
Fig. 4.13 CFIA with an ORL implemented with auto-zero stabilized integrator
C2
C1
f1 f1 f2 f2
Vin Vos Vout
Gm3 Gm2 Gm1
CH1 CH4
CH 3 CH5
f1 R1
Vfbk Vfbk
Gm4 f1
CH2
D CH6 C4
Cascode Cascode R2
Gm5 buffer2 buffer 1
Cint Cpar
Fig. 4.14 CFIA with an ORL implemented with isolating cascode buffers
4.4 Other Sources of Chopper Ripple 83
Fig. 4.15 Block diagram of the implemented fully-differential CFIA
Although, the cascode buffer 1 (CB1) is not chopped, its 1/f noise does not
affect the CFIA’s input-referred noise. This is because this is modulated to the
chopping frequency by CH6 (Fig. 4.16b), and then filtered out by Cint (Fig. 4.16c).
After passing through CH1, the filtered 1/f noise at the chopping frequency is
demodulated to DC, as depicted in Fig. 4.16d. As long as the capacitor Cint is
chosen large enough to filter out the modulated 1/f noise, the amplifier will still
maintain its extremely low 1/f noise corner. Note that since the residual chopper
ripple across Cint is up-modulated by CH3, even-harmonics of f1 are introduced at
the amplifier output.
4.4.2 Chopper Ripple from the Intermediate Stage
The input and the intermediate stages of the CFIA are both chopped to eliminate
their 1/f noise, thus providing sufficient gain to suppress the 1/f noise of the output
Fig. 4.16 1/f noise from Noise Noise LPF

CB1 passing through CH6, PSD PSD
LPF and CH1
f (Hz) fch f (Hz)

(a) (b)
1/f noise of CB1 After Chopping
Noise Noise
PSD PSD
fch f (Hz) f (Hz)

(c) (d)
After LPF Input referred
after CH1
stage down to 1 mHz. Thus, the chopped offset of the intermediate stage is another
source of ripple.
Compared to the ripple due to the offset in the input stage, the ripple from the
intermediate stage is much smaller, because Gm2 is 13 times smaller than Gm3.
Unfortunately, the ripple from the intermediate stage is only weakly affected by
the presence of the ORL, because it originates within the frequency-compensation
network, and so compared to the input stage ripple, is filtered by a different low-
pass filter. In this design, the ripple associated with the intermediate stage is
suppressed by chopping at a much higher frequency (510 kHz) than in the input
stage. The resulting output ripple is then below 70 lV (at a gain of 200). Note that
the increased frequency of the intermediate stage causes charge injection and clock
spikes to the amplifier output through the Miller-compensation capacitors,
resulting in a slightly increased offset.
4.5 Applying ORL to General Purpose Instrumentation

Amplifiers and Operational Amplifiers
As discussed in Sect. 4.3, the ORL creates a notch at the chopping frequency. This
is not a problem for the thermistor bridge application, because the bandwidth of
interest is a few Hz, which is far below the chopping frequency of 30 kHz. As
shown in [9], however, the notch can be eliminated by using a multi-path
architecture. As shown in Fig. 4.17, a chopper amplifier with an ORL serves as the
4.5 Applying ORL to General Purpose Instrumentation 85
C11
+ + -
Vin Gm11
- - +
+ -
Gm2 Vout
- +
R1
+ + -
Vfbk Gm12 C12 +
- - + R2 Vfbk
C41 -
C31 VRef
CH11 CH2
+ - + - + -
Gm31 Gm4 Gm5
- + - + - +
CH12 C32
+ - C42
Gm32
- + Cint Cs1
- +
Gm6 CB1 CB2
+ -
Cs2
Fig. 4.17 Block diagram of the CFIA in a multi-path architecture
C21 Vout+
C11 R21
CH1 CH3 CH 4 CH 5
Vin+ + - + + + - Vfbk+
Gm3 Gm2 2 Gm1 R1
Vin- - + - - - + Vfbk-
1st stage nd
2 stage 3rd stage
CH2 R22
Vfbk+ C12
+ + C22 Vout-
Gm4
Vfbk- - -
CHint Vos2 Cpar C 41
Cascode Cascode
Gm5 Buffer 1
Buffer 2
CH6 C 42
Fig. 4.18 Block diagram of the implemented fully-differential CFIA
low-frequency path, while the combination of Gm11, Gm12 and Gm2 serves as a high
frequency path. The chopper amplifier’s loss of gain at the notch frequency is
compensated for by the gain of the high frequency path, resulting in a smooth
single-pole response.
25µA 25µA
VDDA
R1 R2
VDDA
50µA 50µA
VDDA
VB3 M13
VCM,in1
M14
VB1 M5 VB1 M11
+ -
- +
VB2 M6 VB2 M12 GBp
M15 M16
Feedback+
In+ M1 M2 M7 M8 Out+
Out-
M3 M4 M9 M10*
* * *
In- M18
Feedback- M17
- +
GBn
+ -
VCM,in2
M20
VB4 M19
R3 R4
GNDA
Fig. 4.19 Schematic of the input stage amplifier
Measurement results show that the multi-path CFIA achieves an offset of 2 lV,
and a noise PSD of 21 nV/HHz with an NEF of 9.6. By connecting the inputs of
the input and feedback transconductors in parallel, it can be configured as a general
purpose opamp which achieves an offset of 1.2 lV, and a noise PSD of 10.5 nV/
HHz with a NEF of 4.18 [9].
4.6 Circuit Implementations
For clarity, the fully-differential block diagram of the CFIA is shown again in
Fig. 4.18 and the detailed implementation of each block will be described below.
4.6.1 The Input Stages
The input and feedback stages Gm3 and Gm4 are the most important parts of a
CFIA, as they determine its gain accuracy and noise. To minimize the power
consumption for a given noise specification, most of the power should be dissi-
pated in the input stage. In this design, the input stage amplifier consumes a
161 lA supply current, which is 70 % of the total supply current dissipated in the
4.6 Circuit Implementations 87
VDDA
M14
M13 VB1
M6 VB2
M5
VCM,in1
out-
in+ in+ out+
M1 M3 M4 M2
C1 C1
M9 M11
M7 VB3
M10 M12
M8 VB4
GNDA
Fig. 4.20 Boost amplifier GBp
VDDA
M11 M12
M9 VB1
M13 M14
M10 VB2
out-
out+
in+ in+
M1 M2
M3 M4 C1 C2
M6
M5 VB3
VCM,in2
M8
M7 VB4
GNDA
Fig. 4.21 Boost amplifier GBn
CFIA. The input stage is implemented as a fully differential folded cascode gain-
boosted topology, providing a high DC gain of 140 dB to suppress the noise and
nonlinearity from subsequent stages.
Figure 4.19 shows the schematic of the input stage amplifier. Figures 4.20 and
4.21 depict the schematics of the boost amplifiers implemented in a fully
differential structure with input-regulated common-mode (CM) feedback [2]. The

boost amplifier GBp fixes the voltage across the current source M13 and M14 to the
voltage of VCM,in1 and the boost amplifier GBn fixes the drain voltages of the input
and feedback differential-input pair to the voltage of VCM,in2, ensuring a very high
output impedance.
Taking the boost amplifier GBp for example (Fig. 4.20), the input differential
pair (M1 and M2) is provided with two extra transistors (M3 and M4) in a common-
source configuration, whose gates are connected to the input CM reference
VCM,in1. Due to the feedback between the input and output of the booster amplifier
via the main cascode transistors, the output common-mode voltage is regulated in
such a way that the input common-mode voltage is equal to VCM,in1. Capacitors C1
and C2, as shown in Figs. 4.20 and 4.21, are added to the output terminals of the
booster amplifier to augment stability.
This folded cascode gain-boosted input stage (Fig. 4.19) achieves a DC gain of
140 dB, which is determined by
Adc ¼ Gm3 r0 ð4:17Þ
where Gm3 is the transconductance of the input Gm stage, r0 is the output

impedance seen from the output terminal of the input stage amplifier, which is
determined by Gm15 r015 AGBp r013 ==½Gm17 r017 AGBn ðr019 ==Gm3 r03 r01 Þ. Gm15, Gm17,
and Gm3 are the transconductances of transistors M15, M17 and M3; r015, r013, r017,
r019, r03 and r01 are the output impedances of transistors M15, M13, M17, M19, M3
and M1. AGBp and AGBn are the DC gain of the gain-boost amplifiers GBn and GBp.
The transconductances of Gm3 and Gm4 of the CFIA (Fig. 4.18) need to be well-
matched for good gain accuracy. Consequently, the dimensions, bias currents and
drain-source voltages of the input transistors should also be as well-matched as
possible. The CM voltages of the two transconductances may differ, and so for
good matching their Gm should be immune to CM voltage variations. Therefore,
the drain-source voltages of the input transistors M1, M2, M7 and M8 were kept
constant by low-threshold cascode transistors M3, M4, M9 and M10, as shown in
Fig. 4.19. For the same reason, the current sources M5 and M11 are also cascoded.
Compared to the resistor-degenerated differential pair incorporated with local
feedback loops (Fig. 3.10 and Fig. 3.11), this simple PMOS differential pair shown
in Fig. 4.19 is much more power-efficient, since the input and cascode transistors
share the same bias current. However, its linear input range is then only ±100 mV,
which is good enough for bridge readout.
The input stage determines the noise of the CFIA, since the noise of the fol-
lowing stages is suppressed by the gain of the input stage. Assuming the degen-
eration resistors in the current sources have not been added, the main noise
contribution is from the input transistors M1, M2, M7, M8, the current sources M19,
M20, M13, M14 and the input choppers. The input-referred noise is thus given by
Gm13 2 G
e2n ¼ 4ðe21 þ e213 ð Þ þ e219 ð m19 Þ2 þ e2in;chopper þ e2fbk;chopper Þ ð4:18Þ
Gm3 Gm3
where e1, e13, e19, ein,chopper, efbk,chopper are the noise voltages from transistors M1,
M13 and M19, the chopper switches in CH1 and the chopper switches in CH2,
respectively. As discussed in Sect. 4.1, the CFIA aims to achieve a noise PSD of
15 nV/HHz, which is equivalent to a noise resistance of 14 kX. Next, each noise
source in (4.18) will be analyzed successively.
The input and low-threshold cascode transistors M1–M4, M7–M10 are all in
weak inversion for better power efficiency (Fig. 4.19). Thus, the transconductance
of the input differential pair is 250 lA/V with a 55 lA bias current. For each input
transistors, the Gm/I ratio is 18. This corresponds to an equivalent noise resistance
Rin_Gm of 4 kX and an input-referred noise PSD of 8 nV/HHz. Considering both
input and feedback stages, the total input-referred noise PSD is 8 nV/HHz 9 H2 =
11.3 nV/HHz.
To reduce the noise contribution from the current sources M19, M20, M13, and
M14, resistive-degeneration is applied. The resulting equivalent noise resistance
RCS_up for the upper current source M13 (or M14) is 8 kX, and the equivalent noise
resistance RCS_down for the bottom current source M19 (or M20) is 2 kX.
The third noise contribution is from the choppers CH1 and CH2 (Fig. 4.18),
which precede the input and feedback Gm stages. Since their 1/f noise is not
chopped, they need to be fairly large. The switch size in the chopper is chosen to
be 12 l/0.7 l as a compromise between 1/f noise and charge injection. The
1/f noise from the choppers CH1 and CH2 was simulated using PSS and PNOISE
tools in Spectre RF [1].
For thermistor read-out application, the input and output CM voltages of the
CFIA are 0.6 V and 2.5 V, respectively. Therefore, the on-resistances of these two
choppers differ due to their different CM, and can be calculated as
1
Ron in;chopper ¼ ¼ 190X ð4:19Þ
ln Cox WL ðVGS Vth Þ
1
Ron fbk;chopper ¼ ¼ 480X ð4:20Þ
ln Cox WL ðVGS Vth Þ
where Ron_in,chopper and Ron_fbk,chopper are the on-resistance of the input and
feedback choppers. It can be seen that the resistance of the feedback chopper is
around 10 times smaller than that of one differential pair (4.2 kX). This means that
the noise PSD from the choppers is more than 3 times lower than that of one
differential pair, and thus is negligible.
Overall, the noise PSD of the input stage is given by:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
R2 R2
inputstage ¼ 4kTð2 Rin Gm þ 2 in Gm þ 2 in Gm Þ
V
RCS down RCS up : ð4:21Þ
pffiffiffiffiffiffi
¼ 15nV= Hz
Fig. 4.22 Schematic of the intermediate and output stages
where Rin_Gm, RCS_up and RCS_down denote the equivalent noise resistances of the
input Gm stage, the upper degenerated current sources and the bottom degenerated
current sources.
4.6.2 The Intermediate and Output Stages
The intermediate stage was implemented using a folded-cascode topology and the
output stage was implemented in a class-AB fashion to achieve rail-to-rail output.
Figure 4.22 depicts the schematic of these two stages.
To save power, the class-AB mesh structures were incorporated into the output
branch of the intermediate stage [2]. The class-AB mesh was also cascoded to
reduce the variation of the drain-source voltages of M5, M9, M8, and M12 due to
supply variation. The bias condition of the class-AB stage is determined by four
translinear loops. For example, one of the translinear loops consists of M15, M16,
M5 and M23, thus
VGS15 þ VGS16 ¼ VGS5 þ VGS23 ð4:22Þ
qffiffiffiffiffiffiffiffiffiffiffi
2I
With VGS ¼ ln Cox WL
þ Vth ;
sffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffi

I15 I16 I5 I23
W þ W ¼ W þ W : ð4:23Þ
ð L Þ15 ð L Þ16 ð L Þ5 ð L Þ23
To maintain the same current-to-dimension ratio expressed by (4.23), the

transistors within the translinear loops are all well-matched in the layout.
With I16 ¼ I5 and ðWL Þ16 ¼ ðWL Þ5 from (4.23), we can get
ðWL Þ23
I23 ¼ I15 : ð4:24Þ
ðWL Þ15
To achieve better settling, the demodulation choppers CH51 and CH52 should be
located at the non-dominant poles of the intermediate stage. Therefore, chopper
CH51 was located at the ‘‘quiet’’ sources of the cascode transistors M3 and M4. The
same applies to chopper CH52. Since the thermal noise of the intermediate stage is
suppressed by the gain of the input stage, the differential pair consisting of M1 and
M2 was biased at only 4 lA, resulting in a Gm of 20 lA/V. The unchopped
cascode transistors M3, M4, M34, and M36 are the main source of residual 1/f noise.
However, this is suppressed by the gain of the preceding stages.
4.6.3 The Cascode Buffers
Figure 4.23 shows a schematic diagram of the cascode buffers. Transistors M23
and M24 serve as current buffer 2 (CB2) to avoid chopping the large compensation
voltage across Cint (around 200 mV). Transistors M25 and M26 act as current buffer
1 (CB1) to isolate the chopper CH6 from the sensing capacitors C41 and C42. This
isolation scheme (seen Fig. 4.18), provides lower capacitances Cpar1 and Cpar2
(compared to C41 and C42) at the right side of CH6, so as to reduce the ripple
caused by CB2’s offset.
The ripple reduction ratio is determined by the DC loop gain of the ORL, as
discussed in Sect. 4.3. A DC gain of 120 dB is required in the cascode buffer.
Therefore, a gain-boosting topology was employed to increase its output impedance
(Fig. 4.23a).
The offset of the booster GBn is chopped by CH6, resulting in a square-wave
voltage appearing across the drains of M25 and M26. This voltage charges and
discharges the parasitic capacitors Cpar1,2, creating an AC offset current IAC1.
Furthermore, this square wave voltage modulates the bottom NMOS current
sources to another AC offset current IAC2 [10]. The sum of these two AC currents
charges and discharges the sensing capacitor C41 and C52, appearing as another
source of ripple at the amplifier output. The same goes for GBp, the offset of which
has a similar effect.
To suppress this ripple, the position of the chopper was modified (Fig. 4.23b) so
that these drain capacitances are located at the virtual grounds established by the
gain-boosting amplifiers [10]. Now the mismatch of the bottom current sources
VDDA
(a) VDDA
(b)
CH 9 CH 9
Vos,Gp
Vos,Gp
+ -
- +
+ -
-
GB p GB p
+
CH 8
M27 M28 M27 M28
Cint Cint Int-
Int+ Int+
Int-
M 23 M 24 M 23 M 24
Cascode buffer 2
CH 7
- +
+ -
GB n
- +
+ -
1 GB n 2
Vos,Gn Vos,Gn
iAC
Cpar3 Cpar4
CH6
CH 6
1 2
Cpar2
Cpar1
M25 VB M 26 Cascode buffer 1 M25 VB M 26
Vout+ C41 C41
Vout+
3 4 3
Vout-
4
Vout-
C42 C42 R1 R2
R1 R2 GNDA
GNDA
Fig. 4.23 Implementation of the gain-boosted cascode buffer
Fig. 4.24 Block diagram of

the weak Gm stage Gm5
R1 R2
Vin-
Vin+
IORL+
I
and the offset of GBn appear as a square wave at nodes 1 and 2 (Fig. 4.23b). This
square voltage charges and discharges Cpar3,4, generating an AC current. To reduce
this AC current, both CH6 and M23, M24 were implemented with minimum-size
devices. Therefore, the residual ripple caused by the offset of GBn and the mis-
match of the bottom current sources is filtered out by the integration capacitor Cint.
The AC current due to the offset of the upper current sources and GBp is mitigated
in the same manner.
Compensation Gm Stage Gm5
To minimize the noise contribution from the ORL and increase the input range
of the compensation stage Gm5 in Fig. 4.18, Gm5 is implemented with a resistor-
degenerated PMOS differential pair, as shown in Fig. 4.24. Its transconductance is
14 lA/V, which is 1/18 of the input stage Gm3.
VDD
M10 M11
Ccomp W4 W4
R
L4 L4
M1 M2 M3 M7
W1 W1
k W1
L1 L1
L1 W3
L3
1 2
M4 M5 M6 M8 M9
W2 W2 W2 W2 W2
L2 L2 L2 L2 L2
ID ID ID ID ID
VSS
Fig. 4.25 Constant-Gm bias generator circuit
4.6.4 Constant-Gm Bias Circuit
A constant-Gm bias generator circuit is used to bias the transconductances in the

CFIA so that they do not depend on the temperature or process variation, but to
first order are only determined by a resistor. Figure 4.25 shows the schematic
diagram of the constant-Gm bias circuit [11].
The currents flowing through M4 and M5 are set equal by using a feedback loop
consisting of M3, current mirrors M4–M6 and M1–M2. The feedback loop requires
a compensation capacitor Ccomp for stability. The feedback ensures that VGS3 =
VGS2 and thus V1 = V2, which minimizes channel length modulation effect in M4
and M5. The equal VDS of M4 and M5 ensure a good power-supply rejection at low
frequencies. At high frequencies, the power-supply rejection is reduced by the
imbalance caused by Ccomp [12]. M7 is a very long transistor that produces a
startup current to M6 to raise its gate voltage, and then turn on the PMOS devices.
The transistors in the PMOS differential pairs in Gm3 and Gm4 operate in weak
inversion for better power efficiency (Fig. 4.18), as well as transistors M1–M3 in
the constant-Gm bias generator circuit. The bias current ID of a transistor in weak
inversion exhibits an exponential dependence on gate-source voltage VGS, as given
by:
VGS
ID ¼ IS enVT ð4:25Þ
where n[1 is a non-ideality factor that determines the weak inversion slope factor
of a PMOS transistors, VT = kT/q and IS is the specific current, as given by:
W
IS ¼ 2nlp Cox VT2 ð4:26Þ
L
We can write VGS1 + ID1R = VGS2, thus

ID ID
n VT ln þ ID R ¼ n VT ln ð4:27Þ
IS1 IS2
As shown in Fig. 4.25, M1 has the same L as M2, but k times larger W, thus
IS1 ¼ k IS2 ð4:28Þ
Substituting (4.28) into (4.27), ID is then derived, which exhibits a PTAT
property, as
n VT ln IIS1 n VT ln k
ID ¼ S2
¼ ð4:29Þ
R R
The PMOS differential pair in the amplifier is biased with a multiple m of this
current. If their current densities and their operation region are maintained the
same as that of M2, their transconductance will be, to first order, only determined
by the resistors R and thus be insensitive to temperature:
mID ln k
Gm ¼ ¼ m ð4:30Þ
nVT R
After the implementation description of the analog blocks in the chopper
amplifier, some other critical concerns about the chopper clock generator, chopper
layout, and clock shielding will be discussed.
4.6.5 Chopper Clock Design and Layout
Chopper Clock Shielding

For chopper amplifiers, it is critical to minimize clock skew so as to reduce its
effect on 1/f noise and offset (see Sect. 2.4.2). Therefore, the chopping clock signal
must exhibit a 50 % duty cycle to ensure perfect offset cancellation. Furthermore,
the transition of the chopper clocks must occur at the same moments to avoid any
overlap or non-overlap between two complementary chopper clocks (Fig. 2.11).
As observed from the PSS and PNOISE simulations in Spectre RF [1], the rise
and fall time of the clock pulse should be less than 3.3 ns for a 30 kHz chopping
frequency (0.01 % of the clock period). Otherwise, the noise level near 0.01 Hz
will increase. This implies that the longer the transition time, the more important
the 1/f noise from the chopper switches becomes [13] since they are not chopped.
Figure 4.26 shows a chopper clock generator using a divider-by-2 D-flipflop to
generate 50 % duty-cycle chopper clocks. To reduce the rise and fall time of the
chopper clocks (clk1 and clk2), two inverter buffers INVA and INVB are parallel
connected to the outputs of the D-flipflop. The resulting chopper clocks with steep
transitions feed to the nearby choppers.
Fig. 4.26 Divider-by-2

chopper clock generator INVA
D Q clk1
Clk_in
CP QN clk2
CD
INVB
reset
Chopper Clock Layout

For chopper amplifiers aiming for lV offset, a careful chopper layout is
essential. As discussed in Chap. 2, due to charge injection and clock feed-through,
the imbalance of parasitic capacitors between DC1 and DC2 causes a residual offset
(Fig. 4.27a). Furthermore, the charge injection due to DC (Fig. 4.27b) results in
two bias currents, which together with the impedance mismatch between R1 and R2
cause another residual offset.
To mitigate these two effects, the absolute capacitive crosstalk from the clock
lines towards one side of the chopper is minimized. Figure 4.28 illustrates the
chopper layout [14]. In the layout, the poly clock lines only run perpendicular to
the input signals line, but there is no overlap over the output signal lines. Thus,
over-lap capacitances between the clock lines and the output signals are avoided.
According to Fig. 4.27, the output terminals of this chopper layout should be
chosen as the output of the input chopper (connected to the input of G1) and the
input terminal of the output chopper (connected to the output of G1) to minimize
the parasitic capacitances.
Due to transistor mismatch, another mismatch exists in the gate-source and
gate-drain capacitance. Two dummy transistors Md1 and Md2 are used to ensure
that transistors M1–M4 have the same surroundings. The gate line driving the
dummy switch Md2 is necessary to make the crosstalk from fCH to inn equal to the
crosstalk of fCH to inp by making the capacitances CA1 + CA2 equal to CA3 + CA4.
To further reduce capacitive clock feed-through, a grounded metal 1 plate is used
to shield the metal 2 clock lines from the signal lines. To conclude, the chopper
layout depicted in Fig. 4.28 has matched switch transistors, minimal capacitive
crosstalk towards the output, and balanced crosstalk from each clock line to the
differential inputs and outputs.
Chopper Clock Shielding
The clock signal controlled the choppers need to be routed from the clock
generation circuitry to the chopper switches. This may lead to clock feed-through
and substrate feed-through.
To minimize the clock feed-through, an on-chip coaxial cable is made to restrict
the clock coupling from the sensitive analog circuitry, as highlighted in Fig. 4.29.
Figure 4.30a shows the cross section of the on-chip coax cables. In a three-metal
process, the coaxial cable is made of three metal layers: metal 1–3. The layout of
the two complementary chopper clocks is using metal 2.
(a) fCH ΔC2

ΔC4
R1 ΔC1 ΔC3
+ +
+ + - + +
Vin Ioffset V1 G1 V2 Vout LPF VLPF
- + - -
- - -
R2
CH1 CH2
(b) fCH ΔC
ΔC
ΔC
R1 Ibias ΔC
+ +
+ + - + +
Vin V1 G1 V2 Vout LPF VLPF
- + - -
- - -
R2 Ibias
CH1 CH2
Fig. 4.27 a Offset current due to clock feed-through in a chopper amplifier. b bias current due to
charge injection in a chopper amplifier
via con act m2 m1 pol
Md1
CA1 M1
CB1
CB2 M2
CA2
CA3 M3
CB3
CB4 M4
CA4
Md2
gnd gnd gnd

fCH fCH inn inp outn outp
Fig. 4.28 Chopper layout where parasitic capacitor are shown

Fig. 4.29 Chip micrograph of the CFIA with on-chip coax cable for chopper clocks
M3 gnd
M2 gnd fCH fCH gnd gnd
M1 gnd gnd fCH fCH gnd
Grounded P-substrate/P-well
(a) (b)
M2 gnd
M1 gnd fCH fCH gnd
Gounded and isolated

N-well
(c)
Fig. 4.30 Cross section of on-chip coaxial clock lines in (a) 3-metal process (b) a 2-metal
process (use grounded P-substrate as the bottom plate) (c) a 2-metal process (use grounded and
isolated N-well as the bottom plate)
In a two-metal process, the coaxial clock shielding can be made in two ways.
Figure 4.30b uses the grounded P-substrate as the bottom plate for the clock
shielding [14]. However, the disadvantage is that substrate coupling due to the
capacitance from the clock lines to ground increases, thus relatively strong digital
Fig. 4.31 Noise

measurement set-up
buffers are needed to boost the clock signals. To avoid the substrate coupling issue,
an isolated and grounded N-well can be used as the bottom plate in the coax cable,
as shown in Fig. 4.30c.
4.7 Measurement Results
The CFIA with an ORL has been implemented in a 0.7 lm CMOS process. This
process has low-threshold transistors, linear capacitors and high-resistivity poly
resistors. The 4.8 mm2 active chip area is shown in Fig. 4.29. The measured
supply current is 230 lA from a 5 V supply voltage.
Noise
To ensure that the CFIA’s 1/f noise is dominant, the noise measurements were
made with the CFIA configured for a closed-loop gain of 6667 and followed by a
low-noise amplifier (LNA) with a gain of 100, as shown in Fig. 4.31.
With these gain settings, the LNA acts as a differential to singled-ended buffer
amplifier for the CFIA, so the contribution of the LNA and the HP3562A spectrum
analyzer to the measured 1/f noise is negligible. Without chopping, the amplifier
has a white noise floor of 15 nV/HHz and a 1/f noise corner of 3 kHz. Chopping
only the input stage resulted in a 1/f noise corner of 0.1 Hz. After chopping both
the input and intermediate stages, the measured noise spectral density remained
flat down to 1 mHz. Since the amplifier’s offset is smeared out by the window
function of the spectrum analyzer (HP3562A), the 1/f noise corner could not be
accurately measured, but it is clearly below 1 mHz, as shown in Fig. 4.32. This
confirms the simulation results shown in Sect. 4.2.
The measured noise spectrum from 10 Hz to 100 kHz is shown in Fig. 4.33. In
this measurement, the CFIA is configured for a closed-loop gain of 200 and the
LNA for a gain of 1. The measured output noise PSD is 3 lV/HHz. Thus, it can be
confirmed that the input-referred noise PSD equals: 3 lV/HHz/200 = 15 nV/HHz.
Closed-Loop Gain and Notch Measurement
As mentioned in Sect. 4.3, the ORL acts as a notch filter at the chopping
frequency due to the synchronous demodulation in the ORL. Since the notch is
quite narrow, it has little effect on the measured closed-loop response of the
amplifier (Fig. 4.34). The zoom-in closed-loop response around the chopping
frequency is shown in Fig. 4.35. The measured width of the notch, roughly 3.4
kHz wide at a gain of 20 and fch1=40 kHz, agrees well with the calculations
presented in Sect. 4.3.
4.7 Measurement Results 99
Fig. 4.32 Measured output

noise spectrum from 200 lHz
to 160 mHz (CFIA’s gain =
6667, LNA = 100)
Fig. 4.33 Measured output

noise spectrum from 10 Hz to
100 kHz (Gain of the CFIA =
200, LNA = 1)
Fig. 4.34 Measured

frequency response of the
CFIA (gain of 20, fch1 =
40 kHz, fch2 = 510 kHz)
Fig. 4.35 Measured

frequency response of the
CFIA around the chopping
frequency
Fig. 4.36 Step response of

the CFIA with the offset
reduction loop
This notch will affect the settling of the amplifier with a step input, as it could
cause some ringing before the amplifier output settles. The measured step response
of the amplifier is shown in Fig. 4.36. The amplifier is configured at a gain of 100.
Its output is a step from 0 to 3 V. It can be seen that the CFIA takes about 700 ls
to settle.
Offset and Gain Error
Without chopping, the initial offset of the CFIA is less than 1.7 mV. Chopping
only the input stage results in a measured offset of less than 1 lV. Chopping both
the input and intermediate stages increases the offset to 5 lV, mainly because of
the relatively high chopping frequency (510 kHz) used in the intermediate stage
and the choppers in this stage could cause some charge injection and spikes
through the Miller-compensation capacitors C11 and C12 to the amplifier output
(Fig. 4.18), thus increasing the offset. The measured offset of 12 samples is shown
Fig. 4.37 Measured offset

histogram of 12 samples
Fig. 4.38 Measured gain

error histogram of 12 samples
in Fig. 4.37. Their measured gain error is shown in Fig. 4.38, and was less than
±0.5 % at a nominal gain of 200.
Output Ripple
The spectrum of the chopper ripple with and without the ORL is shown in
Fig. 4.39 and Fig. 4.40. Since the frequency range of the HP3562A spectrum
analyzer was limited to 100 kHz, the input choppers CH1, CH2, CH3, and CH6
were clocked at fch1=30 kHz in order to observe the 3rd harmonic of the chopping
frequency. Measurements show that the amplitude of the output ripple at fch1 was
reduced by about 60 dB: from 48 mV to 41 lV. However, a larger second har-
monic (78 lV) is also visible. This is due to the chopped mismatch of the
degenerated current sources and the offset of the booster amplifiers (Fig. 4.23b),
which CH3 then up-modulates to the even harmonics of fch1. However, at the
closed-loop gains for which the amplifier was designed ([20), the amplifier
bandwidth is low enough to effectively filter out these harmonics. At a gain of 200,
the amplifier’s bandwidth is 4 kHz and the measured input-referred output ripple
and noise are 0.55 lV(rms) and 0.95 lV(rms), respectively. Depending on the
offset in the intermediate stage, the amplitude of the corresponding output ripple
(at 510 kHz) varies from 0 to 70 lV.
Fig. 4.39 Measured chopper

ripple with the ORL ‘‘off’’
(125 Hz–100 kHz)
Fig. 4.40 Measured chopper

ripple with the ORL ‘‘on’’
(125 Hz–100 kHz)
4.8 Benchmark and Conclusions
To interface a precision thermistor bridge, a three-stage stand alone current-

feedback instrumentation amplifier (CFIA) has been realized. By chopping both
the input and intermediate stages, their 1/f noise was effectively suppressed, while
the 1/f noise corner of the output stage was suppressed by the gain of the preceding
stages down to 1 mHz. A continuous-time offset reduction loop (ORL) was pro-
posed to reduce chopper ripple. Due to its continuous-time nature, it does not cause
noise aliasing, thus resulting in a very power-efficient solution compared to the use
of auto-zeroing. The ORL dynamically compensates for the offset and thus
eliminates temperature dependent offset and achieves low offset drift.
The ORL behaves like a narrow-band notch filter at the chopping frequency and
its harmonics. As a result, the ORL has little effect on the amplifier’s low-
4.8 Benchmark and Conclusions 103
Table 4.2 Performance comparison of low-power chopper amplifiers

This work Yazicioglu Denison Fan Burt [4] Kusuda
[7] [15] [16] [9] in a [17]
two- in a two-
opamp opamp
IA IA
Year of publication 2009 2008 2007 2010 2006 2011
1/f noise corner (Hz) 1m 3 1 3 – –
Noise PSD (V/HHz) 15 n 55.8 n 94 n 21 n 76 n 9.2 n
Offset (lV) 5 – – 2 6 0.78
GBW (kHz) 800 Stable 350 20 900 350 3700
Gain [20
Chopping frequency 30, 510 – 4 30 125 200
(kHz)
CMRR (dB) [120 [120 100 137 130 [150 dB
PSRR (dB) [120 90 – 120 113 [150 dB
Supply current (lA) 230 2.3 1.2 143 34 2940
Ripple suppression 60 dB – – 60 dB 54 dB –
ratio
Input-referred ripple f1 :0.41 lV – – f1 – –
2f1:0.78 lV :0.45 lV
(gain = 100) 2f1:
0.39 lV
(gain =
100)
NEF [18] 8.8 4.1 4.6 9.6 18.4 19.2
GBW/Isupply 3.5 152 17 6.3 10.3 1.25
frequency response, but creates notches in the amplifier’s closed-loop transfer

function around the chopping frequency f1 (and its harmonics). For wide-band
applications, this notch can be eliminated by using a multi-path architecture [9].
Since the ORL is a stable first-order system with a phase shift close to 90, the
notch width (determined by the unity-gain-frequency) and the notch location
(determined by f1) can be independently chosen [7]. Compared to the situation
when a switched-capacitor notch filter [4] or an auto correction feedback is used
[4.6], a relatively low chopping frequency can be chosen in this work. As a result,
low offset can be achieved without the stability problems caused by excessive
notch-filter phase shift. It has been proven that the concept of the ORL can be
applied to general-purpose chopper CFIA or opamps [9].
To sum up, the advantages of the ORL compared to other ripple reduction
techniques are:
• Continuous-time nature, thus incurring no noise folding ? power efficient
• Inherently stable ? lower chopping frequency and lower offset
Measurement results show that the ORL reduces the amplitude of the chopper
ripple by 1100 times, to below the amplifier’s own input-referred noise. Table 4.2
summarizes the measured performance of the CFIA and compared with the state-of-
the-art chopper amplifiers. This CFIA achieves a worst-case 5 lV offset and a
1 mHz 1/f noise at noise PSD of 15 nV/HHz while consuming only a 230 lA supply
current. The NEF of the CFIA is 8.8, which is quite respectable [4, 15, 16–17]. To the
author’ knowledge, this represents the best LF noise performance ever reported for a
stand-alone CMOS amplifier.
References
1. Kundert K (2005) Simulating switched-capacitor filters with spectre RF, The Designer’s
Guide Community. http://www.designers-guide.org/Analysis/sc-filter.pdf
2. Huijsing JH (2011) Operational amplifiers theory and design, 2nd edn. Springer, New York
3. Tang ATK (2002) A 3lV-Offset operational amplifier with 20nV/HHz input noise PSD at
DC employing both chopping and autozeroing. In: IEEE ISSCC, Digest of Technical Papers,
pp 386–387, Feb 2011
4. Burt R, Zhang J (2006) A micropower chopper-stabilized operational amplifier using a SC
notch filter with synchronous integration inside the continuous-time signal path. IEEE J
5. Luff GF (2010) Chopper stabilized amplifier. United States Patent, US 7724080 B2, May
2010
6. Kusuda Y (2010) Auto correction feedback for ripple suppression in a Chopper amplifier.
IEEE J Solid-State Circuits 45(8):1436–1445
amplifier with a 1 mHz 1/f noise corner and an AC-coupled ripple reduction loop. IEEE J.
8. Witte JF, Makinwa KAA, Huijsing JH (2007) A CMOS chopper offset-stabilized opamp.
IEEE J Solid-State Circuits 42:1529–1535
9. Fan Q, Huijsing JH, Makinwa KAA (2010) A 21nV/HHz Chopper-stabilized multipath
current-feedback instrumentation amplifier with 2lV Offset. In: IEEE ISSCC, Digest of
Technical Papers, pp 80–81, Feb 2010
10. Kashmiri SM, Makinwa KAA (2008) A temperature-to-digital converter based on an
optimized electrothermal filter. In: ESSCIRC Digest of Technical Papers, pp 74–77,
Sept 2008
11. Nauta HC, Nordholt EH (1985) New class of high-performance PTAT current sources.
Electron Lett 21:384–386
12. de Langen KJ (1999) Advanced low-voltage and high-speed techniques for BiCMOS, CMOS
and bipolar operational amplifiers. Ph.D. Thesis, Delft University of Technology, The
Netherlands
13. Sanduleanu MAT, van Tuijl AJM (1998) A low noise, low residual offset, Chopped amplifier
for mixed level applications. In: IEEE IECAS, pp 333–336, Sept 1998
14. Witte JF (2008) Dynamic offset compensated CMOS amplifiers. Ph.D. Thesis, Delft
15. Yazicioglu RF et al (2008) A 200lW eight-channel acquisition ASIC for ambulatory EEG
systems. In: IEEE ISSCC, Digest of Technical Papers, pp 164–165, Feb 2008
16. Denison T et al (2007) A 2.2lW 94nV/HHz Chopper-stabilized instrumentation amplifier for
EEG detection in chronic implants. In: IEEE ISSCC, Digest of Technical Papers, pp 162–163
References 105
17. Kusuda Y (2011) A 5.9nV/HHz Chopper operational amplifier with 0.78lV maximum offset
and 28.3nV/C offset drift. In: IEEE ISSCC, Digest of Technical Papers, pp 242–243,
Feb 2011
instrumentation amplifier for medical purpose. IEEE J Solid-State Circuits
22SC(6):1163–1168
Chapter 5
A Chopper Instrumentation Amplifier
with Gain Error Reduction Loop
This chapter describes a stand-alone chopper current-feedback instrumentation

amplifier (CFIA) that has improved performance compared to the one described in
Chap. 4. It maintains the latter’s low noise and low offset, and also obtains high
gain accuracy and low gain drift without trimming. This is achieved by applying
dynamic element matching (DEM) to the input and feedback transconductors so as
to average out their mismatch. To eliminate the resulting DEM ripple, a gain error
reduction loop (GERL) is employed to continuously null the Gm mismatch. The
concept and analysis of DEM and the GERL is presented in Sects. 5.2 and 5.3.
Then the similarities and differences between the offset reduction loop (ORL) and
the GERL are discussed, together with their effects on the input and feedback Gm
transfer functions.
Measurements show that without trimming, this CFIA achieves a gain error of
0.06 % and a maximum gain drift of 6 ppm/°C. It consumes only 290 lA supply
current with an NEF of 11.2. Compared to previous CFIA with resistor-degenerated
input stages, this work achieves a 4 9 improvement in power efficiency.
5.1 Motivation
In instrumentation amplifiers, offset and gain error are the two dominant sources of
error. For small input signals, offset and CMRR errors dominate; while for large
signals, gain error dominates. The chopped CFIA presented in the previous chapter
uses chopping to achieve microvolt-level offset and high CMRR ([120 dB) [1].
However, its gain error is about 0.5 %, even with careful layout (Sect. 4.7), and is
thus the dominant source of residual error for input signals larger than a few
hundred microvolts. For high-end strain gauge applications, the interface elec-
tronics should have a gain error less than 0.02 % [2]. Thus, the gain error of the
CFIA needs to be reduced.

108 5 A Chopper Instrumentation Amplifier
(1+Δ)Videal,out
Swapper C1 Videal,out
Vin Gm3 Gm1 Vout
R1 Videal,out /(1+Δ)
Vfbk
R2
Vfbk Gm4
Gm4 = Gm3 (1+ )
Fig. 5.1 CFIA with DEM ripple after applying DEM
Provided precision feedback resistors are used, the gain error of a CFIA is
mainly determined by the mismatch between the input and feedback transcon-
ductances. To reduce this mismatch, one common approach is to use resistor-
degenerated input stages [3, 4], as discussed in Sect. 3.3. However, the use of
resistor degeneration leads to a significant loss in power efficiency. Furthermore,
trimming the degeneration resistors [4] increases product costs. A better solution is
to apply dynamic element matching (DEM) to average out their mismatch, thus
permitting the use of simple differential pairs.
5.2 Dynamic Element Matching
As shown in Fig. 5.1, DEM can be implemented by swapping the positions of the
input and feedback transconductors in the circuit. This only requires an input
multiplexer, which we will call a ‘‘swapper’’. The use of DEM reduces the initial
gain error from D to D2/2 [see Eq. (3.8)], Thus, a Gm mismatch of 2 % will be
reduced to 0.02 %. In a practical CFIA, however, the common-mode (CM)
dependence of Gm3 and Gm4 will limit the gain accuracy even if DEM is applied.
In other words, the CM-dependent mismatch Dcm will not be suppressed by DEM
[see Eq. (3.9)]. For simplicity, this issue will be neglected in the following anal-
ysis. However, a circuit-level technique to mitigate this problem will be described
in Sect. 5.6.
The improved gain accuracy achieved by DEM comes at the expense of ripple
at the amplifier output due to the modulated Gm mismatch. The amplitude of this
DEM ripple is given by
1
VDEM;ripple ¼ Vout1 Vout2 ¼ 2D Vin ¼ 2D Videal;out : ð5:1Þ
b
where Videal,out is the ideal output voltage of the CFIA. Eq. (5.1) indicates that the
amplitude of the DEM ripple is a product of the mismatch D and the output signal.
One common solution to eliminate DEM ripple is to trim the transconductances
5.2 Dynamic Element Matching 109
Initial state: 2Δ•Videal,out
DEM C2 Steady state:

Swapper
Vin Gm3 Gm2 Vout
R1
Vfbk
Vfbk Gm4 R2
C7
CH10 C5
Gm6 Gm7
Vint,GE IDC IAC
fDEM
Initial state:
Steady state:
Fig. 5.2 Simplified block diagram of a CFIA with gain error reduction loop in the initial and
steady state
[5]. However, trimming increases production costs and will not compensate for
temperature drift. To avoid these issues, an automatic feedback loop—a GERL
[6]—is employed to suppress DEM ripple.
5.3 Analog Gain Error Reduction Loop
5.3.1 Basic Concept
The GERL employs a synchronous detection technique similar to that used in the
ORL. It extracts mismatch information from the amplitude of the DEM ripple, and
then drives this ripple to zero by continuously nulling the Gm mismatch of the
input and feedback transconductors, thus eliminating the need for trimming.
As shown in Fig. 5.2, the sense capacitor C5 converts the amplifier’s output
ripple VDEM,ripple into an AC current IAC whose amplitude is proportional to the
derivative of VDEM,ripple. This current is then demodulated by chopper CH10, and
the resulting DC current IDC is integrated by Gm7 and C7 to generate a DC
compensation voltage Vint,GE proportional to the ripple amplitude. Via transcon-
ductance Gm6, this voltage is then used to cancel the mismatch between Gm3 and
Gm4 by adjusting their tail currents. When the Gm mismatch is exactly compen-
sated by the feedback loop, the output DEM ripple is ideally reduced to zero.
C2
Swapper
Vin Gm3 Gm2 Vout
R1
Vfbk
Vfbk Gm4 R2
Gain error reduction loop
C7 Comparator
CH10 CH9 C5
Gm6 Gm7
VInt,GE
fDEM Polarity Q1
Fig. 5.3 CFIA with gain error reduction loop and polarity control switch
5.3.2 Qualitative Analysis
Unlike the offset reduction loop (see Sect. 4.3), which feeds back an additive offset-
compensating signal, the GERL feeds back a multiplicative gain-compensating
signal, which adjusts the ratio of Gm3 and Gm4. The output DEM ripple is then the
product of the residual mismatch and the output signal, and so the gain of the GERL
will be signal dependent. This means that the loop gain not only depends on the
amplitude of the output signal [shown in Eq. (5.1)], but also depends on the polarity
of the output signal. Therefore, if no measures are taken, the GERL could turn into a
positive feedback loop, causing Vint,GERL to clip, and actually maximizing the
CFIA’s gain error. To maintain negative feedback, a polarity reversing switch is
used to link the polarity of the GERL to that of the output signal (Fig. 5.3). In
practice, this switch is a chopper, which is driven by a quantizer Q1 that periodically
monitors the polarity of the CFIA’s output signal. However, the offset of the
quantizer could cause it to generate incorrect polarity information, giving rise to
large output ripple. To prevent this, an auto-zeroed quantizer is used with an esti-
mated offset of less than 60 lV. At a closed-loop gain of 100, this translates into an
error of 0.6 lV at the input of the CFIA, which, in most applications, is negligible.
5.3.3 Quantitative Analysis
To gain better insight into the operation of the GERL, its loop transfer function can
be derived with the help of the simplified block diagram shown in Fig. 5.4. For
simplicity, the polarity reversing switch CH9, quantizer Q1 (Fig. 5.3), choppers for
offset reduction and ORL are neglected.
5.3 Analog Gain Error Reduction Loop 111
DEM Vout
Vin G F 1
Multiplier AC =
(1+sτ )
·Ibias-Icmp VG SWP1 VF
=
Ibias
Ibias Divider
·I bias
C7 DEM
Icmp J H C5
Gm6 Gm7
Vint Isc5 IH
CH10
Fig. 5.4 Simplified block diagram of a CFIA with an GERL for loop-gain calculation
From (5.1), the output DEM ripple is a product of the input signal Vin, the
mismatch D between Gm3 and Gm4 and the closed-loop gain 1/b. In line with this,
Vin is multiplied in Fig. 5.4 by a normalized gain error e, resulting in a voltage VG,
where e represents the gain error after feedback loop compensation. For maximum
power efficiency, Gm3 and Gm4 are biased in weak inversion and so their mismatch
is proportional to their tail current mismatch. Thus, e can be expressed as
D Ibias Icmp
e¼ ; ð5:2Þ
Ibias
where D is the initial mismatch between the input and feedback Gm stages, Ibias is
the nominal bias current of input (or feedback) Gm stage and Icmp is the com-
pensation current from the GERL.
The voltage VG is then fed to the DEM swapper SWP1, where it is converted to
a square-wave VF. This voltage is then amplified by the CFIA’s closed-loop gain
AC, and appears as an output ripple. This ripple is differentiated by C5 into an AC
current IH, which is then demodulated by CH10 into a DC current ISC5. For a
sufficiently high DEM frequency, the CFIA’s gain AC can be modeled as an ideal
integrator with a pole at DC:
1 1
AC ¼ ; ð5:3Þ
bð1 þ ssÞ sbs
where the time constant s determines the settling behavior of the output DEM
ripple. The relation between the current IH and VF is then given by
IH sC5 C5
K¼ ¼ : ð5:4Þ
VF sbs bs
Since this gain factor is frequency-independent, the actions of the swapper

SWP1 and the chopper CH10 cancel each other, and so:
ISC5 ¼ K VG : ð5:5Þ
From its Norton equivalent circuit, the action of C5 and CH10 can be modeled as
a switched-capacitor impedance ZSC5 = 1/fDEMC5 [1]. Assuming Gm7 has a finite
DC voltage gain of A07, thus the transfer function H(s) of the integrator built
around Gm7 and C7 (Fig. 5.4) can be calculated as
Vint ZSC5 A07
HðsÞ ¼ ¼ : ð5:6Þ
ISC5 1 þ sZSC5 ð1 þ A07 ÞC7
The tail current mismatch DIbias can be seen as the input of the loop since the
feedback loop tries to cancel it to zero. Thus, the loop gain L(s) of the GERL is
derived as
Vin K HðsÞ Gm6
LðsÞ ¼ ð5:7Þ
Ibias
By substituting (5.4) and (5.6) into (5.7) and assuming A07 1, the loop gain
becomes
C5 ZSC5 A07 Gm6
LðsÞ ¼ Vin : ð5:8Þ
bs 1 þ sZSC5 A07 C7 Ibias
Equation (5.8) indicates that the loop gain L(s) is a first-order low-pass function
with a dominant pole that is related to the DEM frequency by
1 fDEM C5
fdominant pole;GERL ¼ ¼ ð5:9Þ
2pA07 C7 ZSC5 2pA07 C7
The phase shift within the loop mainly originates from three blocks: an inte-
grator made from the single pole closed-loop transfer function AC, the differenti-
ator C5 and the integrator C6. In this design, C5 = 10 pF, C7 = 40 pF,
fDEM = 8 kHz, A07 is 120 dB, so the dominant pole is at around 0.32 mHz. Since
the phase shift of the first two blocks cancels out, the feedback loop behaves as a
stable first-order system with a phase margin close to 90°.
The DC loop gain L(0) corresponding to the DEM ripple reduction ratio is
given by
C5 Gm6 Vin A07 Gm6
Lð0Þ ¼ Vin ZSC5 A07 ¼ ð5:10Þ
bs Ibias bsfDEM Ibias
As seen from (5.10), the DEM ripple reduction ratio depends on the magnitude
of the input signal Vin and the DEM frequency fDEM. In this design, fDEM is 8 kHz,
Gm6 = 0.56 lA/V and Ibias = 55 lA. s is determined by the unity-gain bandwidth
of the CFIA and the feedback factor b, as given by
5.3 Analog Gain Error Reduction Loop 113
1 1 2pC2 1 C2
s¼ ¼ ¼ : ð5:11Þ
2pfUGB;CFIA b 2p Gm3 b Gm3 b
where Gm3 = 270 lA/V, C2 = 80 pF and b = 1/100, thus s = 30 ls. With an

input signal of 30 mV and the worst case 2 % mismatch, the DEM ripple
amplitude without the GERL would be 120 mV at a closed-loop gain of 100.
According to (5.10), however, the ripple reduction ratio is around 113 dB. This
means that neglecting other contributors to the DEM ripple, even the worst-case
DEM ripple should be reduced to microvolt levels.
Since the initial DEM ripple and the ripple reduction ratio L(0) are both pro-
portional to the ideal output signal Videal,out (= Vin/b), the amplitude of the residual
DEM ripple is signal-independent. This is an important advantage of GERL
compared to trimming, because trimming can not eliminate the signal dependency
in the DEM ripple, unless it is trimmed at all the input levels. With a signal-
dependent residual ripple at the CFIA output (after trimming), the sampling
moments of the succeeding ADC then become very critical and must locate at the
zero-crossing points of the ripple. Otherwise, a signal-dependent sampling error
occurs. While with a signal-independent residual ripple (with GERL), the sam-
pling error only causes an offset, which can be easily calibrated out.
The unity-gain bandwidth f0 of the loop can be derived from (5.8) by setting
L(s) = 1
C5 ZSC5 A07 Gm6
LðsÞ ¼ Vin ¼1 ð5:12Þ
bs 1 þ 2pf0 ZSC5 A07 C7 Ibias
Vin C5 Gm6
f0 ¼ ð5:13Þ
2pbsC7 Ibias
In this design, with Vin = 30 mV, C5 = 10 pF, Gm6 = 0.56 lA/V, b = 1/100,
s = 30 ls, C7 = 40 pF and Ibias = 55 lA, the unity-gain-frequency f0 is 41 Hz.
To verify the above analysis of the GERL and the analysis of the ORL
described in Sect. 4.3, the settling times of the ORL and GERL were simulated.
The CFIA is configured at a gain of 100 and a step input from 0 V to 40 mV is
applied. Figure 5.5 shows the simulated waveforms of the step input, the step
output of the CFIA, the integrator output of the GERL and the integrator output of
the ORL.
It can be seen that the integrator outputs of the ORL and the GERL both settle
in an exponential manner, confirming that these two loops are stable first-order
systems. The time constant of the ORL is about 600 ls, while the time constant of
the GERL is about 3 ms. The time constant of the GERL is thus in good agreement
with the prediction of (5.13). During the transition time of the step input, the
periodically sampled polarity-control quantizer is not fast enough to give the
correct polarity information, thus the output of the GERL’s integrator first goes in
the opposite direction before settling to the correct mismatch compensation
voltage.
CFIA Input
40mV
Long zero input
0 3 6 9 12 15
Time (ms)
ORL settles
CFIA Output (V) within 0.8ms
4
Long zero input 1

0
0 3 6 9 12 15
Time (ms)
GERL integrator
output (mV)
300
150 Analog GERL

Long zero input
0
0 3 6 9 12 15
Time (ms)
-150
ORL integrator
output (mV)
200
100 ORL
Long zero input
0
0 3 6 9 12 15
Time (ms)
Fig. 5.5 Simulated settling time of the ORL and the GERL with a step input
5.4 Digitally-Assisted Gain Error Reduction Loop
From (5.8), the loop gain of the GERL is proportional to the input signal and so is
zero for zero input. In this case, leakage causes the integrator output Vint,GE to drift
with a time constant of several seconds and eventually clip. The GERL will then
5.4 Digitally-Assisted Gain Error Reduction Loop 115
Analog GERL
C7 Vout
IGE+
CH10 CH9 C5
Gm6 Gm7
V int,GE Q1
IGE-
fDEM
Polarity
Comparator
Digitally – assisted GERL Vout
10bit f DEM/8 Q
IGE+ 3
CH10 CH9
Counter
C5
Gm6
DAC EN UP/DN!
IGE- fDEM Q2
Polarity
CNT_EN
Comparator
Fig. 5.6 Comparison between the analog GERL and digitally-assisted GERL
need to resettle whenever a finite input signal re-appears: within 14 ms for a

40 mV step input at a closed-loop gain of 100 and the loop parameters given
above. To avoid the need for resettling, a digitally-assisted approach can be
employed to store the mismatch information in the digital domain. The result is the
digitally-assisted GERL shown in Fig. 5.6. For comparison, the analog imple-
mentation of the GERL is also shown. Here, the analog integrator is replaced by a
comparator (Q3), an up-down counter, a 1-bit first order DR DAC.
The comparator Q3 determines the polarity of the demodulated DEM ripple. Its
output then increments or decrements the up-down counter, whose output, in turn,
drives the DAC. The DAC output drives a transconductance Gm6 that, as in the
analog GERL, generates the differential currents required to compensate for the
Gm mismatch of the input and feedback transconductors. The function of the three-
level quantizer Q2 in Fig. 5.6 is to control the polarity of the loop via CH9 for
relative large input signal and also ensures that the integrator state is ‘‘frozen’’ for
small input signals. In this way, the state of the loop will not drift in the presence
of small input signals.
The threshold voltage to define the small inputs is chosen at a level when the
output gain error equals the output-referred offset. For instance, assuming
the closed-loop gain is 100, the gain error is 0.1 % and the offset is 2 lV, then the
minimum input signal X, which ensures that the error induced by gain error is no
less than the output offset error, is given by
X 100 0:1 % 2lV 100: ð5:14Þ
Thus, when the output signal becomes larger than 200 mV, the gain error
dominates. Figure 5.7 shows the decision thresholds of the three-level quantizer.
When the absolute amplitude of the CFIA output is smaller than 200 mV,
Fig. 5.7 Decision threshold

of the three-level quantizer
+ A 200mV
B 0
Frozen
- C -200mV
the digital counter freezes the mismatch information, while when the absolute
output amplitude is larger than 200 mV, the loop’s polarity is controlled by the
quantizer’s polarity decisions.
At steady state, the output of the DAC will toggle between two LSBs, giving
rise to a limit cycle within the loop. 10-bit resolution is enough to ensure that the
resulting input-referred tone is well below the CFIA’s noise level and also pro-
vides enough dynamic range to suppress the DEM ripple. Since the 10-bit counter
is updated at a rate of fS = fDEM/8 = 1 kHz, the digitally-assisted GERL has a
worst-case start-up time of 210/fs = 1 s. This can be significantly reduced by
adjusting the counter’s state in a successive-approximation fashion.
The usable signal bandwidth of the amplifier does not depend on the settling
time of the GERL, but is determined by the DEM frequency (fDEM = 8 kHz). For
input frequencies higher than fDEM/2, the quantizer Q2 (Fig. 5.6) will not give
correct polarity information since it is clocked at fDEM. Thus, good gain accuracy
is achieved in the frequency band lower than 4 kHz. However, this is not a
problem because at high frequencies, the gain will probably not be accurate
anyway due to the first-order roll-off of the CFIA’s open-loop gain.
5.5 Comparison Between ORL and GERL
The combination of chopping and an ORL is similar to that of the combination of

DEM and a GERL (Fig. 5.8). They both use modulation techniques to modulate
the error (offset or Gm mismatch) to a high frequency, synchronously demodulate
the output ripple, and then use a feedback loop to drive the error to zero. The main
difference between them is that the ORL feeds back an additive offset-compen-
sating current to cancel the offset, while the GERL feeds back a multiplicative
gain-compensating current to fine-tune the tail currents of Gm3 and Gm4 in order to
cancel their mismatch. Therefore, the GERL and the ORL can work independently
of each other, provided that these two loops are clocked at different frequencies,
and that the chopper ripple and gain error ripple are suppressed to a low enough
level. Only prior to the complete settling of these two loops, the unsuppressed
offset chopper ripple and DEM ripple may cause some interaction between these
two loops.
5.6 The Effects of Chopping, DEM and GERL on CFIA Performance 117
C2
Swapper CH1 CH5
Vin Gm3 Gm2 Vout
fchop fchop R1
Offset reduction loop (ORL) Vfbk
CH2 C3
CH8 C4
Vfbk Gm4 R2
Gm5 Gm8
fchop fchop
Gain error reduction loop (GERL)
C7 CH10 C5
Gm6 Gm7
fDEM
Fig. 5.8 Simplified block diagram of a CFIA employing chopping ? ORL and DEM ? GERL
5.6 The Effects of Chopping, DEM and GERL on CFIA

Performance
Since the nonlinearities of the input and feedback Gm stages compensate each
other, the typical nonlinearity of the CFIA is at the 30 ppm level (at a gain of 100).
Figure 5.9a shows the typical transfer functions of Gm3 and Gm4 as a function of
their input range. It is assumed that their transconductances decreases slightly with
input amplitude. The presence of offset causes a horizontal shift between the two
Gm characteristics. Chopping eliminates this offset and thus aligns the resulting
average characteristics of Gm3 and Gm4 (Fig. 5.9b).
However, Gm3 and Gm4 still exhibit a transconductance mismatch of D.
Applying DEM to Gm3 and Gm4 reduces their average gain error from D to D2/2, as
shown in Fig. 5.9c. Compared to the situation without DEM, the use of DEM
moves their average transcondutances Gm3,avg and Gm4,avg closer to each other. As
a result, the nonlinearities of these two Gm stages better compensate each other,
thus improving CFIA’s linearity. The GERL improves matters further, since it
drives the mismatch D to zero. As a result, Gm3,avg and Gm4,avg become even more
closely aligned (Fig. 5.9d), which, in turn, results in a further improvement in the
linearity of the CFIA. The same goes for the gain error and gain drift of the CFIA.
Transistor-level simulations are used to confirm the above analysis. The CFIA
is simulated at a gain of 100 with 10 mV offset and with the offset reduction loop
‘‘on’’. A 2 % mismatch is added between the input and feedback transconductors.
Figure 5.10 shows the simulated INL in four cases: (a) DEM phase 1; (b) DEM
phase 2; (c) DEM only and (d) DEM ? GERL. Without DEM, the CFIA’s INL is
around 90 ppm for both of the swapper phases. The use of DEM improves the
amplifier’s linearity from 90 ppm to 30 ppm. As described above, this is because
DEM reduces the transconductor mismatch from D to D2/2, moving the transfer
Gm Gm
No Chop With Chop
Gm3 No DEM Gm3,ch No DEM
Voffset Δ
Gm4 Gm4,ch mismatch
-Vin 0 Vin -Vin 0 Vin

Input range (V) Input range (V)
(a) (b)
Gm Gm
With Chop With Chop
and DEM DEM+ GERL
Gm3,avg Gm3,avg
Gm4,avg Δ2/2 Gm4,avg Δ=0

mismatch
-Vin 0 Vin -Vin 0 Vin

Input range (V) Input range (V)
(c) (d)
Fig. 5.9 Effects of chopping, DEM and GERL on the transfer functions of Gm3 and Gm4 in four
cases: a No Chop and no DEM; b With chopping and no DEM; c With chop and DEM; d With
chop, DEM and GERL
functions of Gm3,avg and Gm4,avg closer to each other. Turning on the GERL further
improves the INL to 6 ppm, since the mismatch is now reduced to zero. It can also
be seen that the combination of DEM and the GERL extends the CFIA’s linear
input range.
5.7.1 Current-Feedback Instrumentation Amplifier with Analog

Gain Error Reduction Loop
Figure 5.11 shows the block diagram of the implemented CFIA. Similar to the
CFIA in Chap. 4, it consists of three fully-differential gain stages with an open-
loop DC gain in excess of 250 dB. The input and feedback transconductors Gm3,
Gm4 as well as the intermediate stage Gm2 are chopped, and their gain sufficiently
suppresses the 1/f noise of the unchopped class-AB output stage Gm1 down to
1 mHz [1]. As in [1], the chopping frequencies for the input and the intermediate
stages are 32 kHz and 512 kHz, respectively. The chopper ripple caused by the
Fig. 5.10 Simulated INL

with transistor-level circuit in
four cases: a DEM phase 1;
b DEM phase 2; c DEM only;
d DEM ? GERL
offset of Gm3 and Gm4 is suppressed by the ORL, while the higher frequency ripple
of the intermediate stage is suppressed by the Miller-compensation network [1].
To minimize their noise contribution, the DEM swapper and the input choppers are
merged and realized by four NMOS choppers CH1, CH2, CH3, and CH4. The
timing of these choppers is also shown in Fig. 5.11. In DEM swap phase 1,
choppers CH1 and CH4 are active (Fig. 5.12), while in DEM swap phase 2,
choppers CH2 and CH3 are active (Fig. 5.13).
As in the ORL (Sect. 4.4.1), a current buffer before the demodulation chopper
CH10 is used to minimize the DEM ripple due to the offset of Gm7 (Fig. 5.8). To
avoid chopping the large compensation voltage across C7, the current buffer 4
(CB 4) is added to separate the chopper CH10 from C7. The implementation of the
intermediate stage Gm2, the output stage Gm1, the cascode buffer 3 (CB3) and the
cascode buffer 4 (CB4) of the GERL is similar to that in the ORL [1].
Configured at a gain of 100, the gain setting resistors R1, R21 and R22 of the
CFIA are 300 X, 14.8 kX and 14.8 kX, respectively. As discussed in Sect. 4.2, the
switched-capacitor (SC) impedance corresponds to a 26 MX input impedance,
which is much larger than the equivalent resistance of the feedback network
(300 X) and is high enough for most sensor read-out applications. In addition, the
swapping action and the parasitic capacitances Cpar1-4 also lead to switched-
capacitor impedance between the input and feedback nodes of the amplifier. This
impedance value is determined by 1/(2fDEM2Cpar1,2) = 52 MX, which is high
enough and thus will not load the output stage of the amplifier.
5.7.1.1 Design Considerations and Implementation of DEM
To ensure that the GERL and the ORL only respond to the DEM ripple and the
chopper ripple, respectively, the two loops are operated at different frequencies.
Choosing a DEM frequency fDEM higher than the input stage’s chopping frequency
C21 Vout+
R21
Cpar1 C11
CH1 CH 5 CH 6 CH 7
Vin+ + - + - + - Vfbk+
Gm3 Gm2 2 Gm1 R1
Vin- - + - + - + Vfbk-
Cpar2 f2 f2
P1 P1 IGE+
fchop1 fchop1 C12
CH2 Cpar3 R22
C22 Vout-
P2 P2 + - ORL
CH3 Gm4 C 41
- + CH 8
Cpar4 IGE- Cascode Cascode

Gm5 C3 Buffer 1
Buffer 2
P2 P2 C 42
fchop1
CH4
Vfbk+
Vint,GE CH10 CH 9 C 51 Comparator
Vfbk-
Cascode Cascode
P1 P1 Gm6 C7 Buffer 4 Cpar5 Buffer 3
P1 (CB4) (CB3)
P2 C 52 Q1
fDEM
fchop1
f DEM fchop1= 4f DEM GERL
Fig. 5.11 Block diagram of a CFIA with ORL and analog GERL
C21 Vout+
R21
C11
CH1 CH5 CH6 CH7
Vin+ + - + - + - Vfbk+
Gm3 Gm2 2 Gm1 R1
Vin- - + - + - + Vfbk-
Clk1 f1 f2 f2
Delta_It1
C12
CH2 R22
C22 Vout+
+ -
Clk2
Gm4 C41
CH8
CH3 - +
C3 Cascode Cascode
Gm5
Delta_It2 Buffer 2 Buffer 1
f1 C42
Clk2
CH4 CH10 CH9 C51
Vfbk+ Comparator
C7 Cascode Cascode
Vfbk- Gm6 Buffer 4 Buffer 3
Clk1 f1/4 C52 Q1
Fig. 5.12 CFIA with ORL and GERL in DEM phase 1
fchop1 causes extra switching activity during a chopping period, as shown in

Fig. 5.14. The resulting spikes would increase the CFIA’s residual offset [7]. To
avoid this, fDEM was chosen to be a sub-multiple of fchop1.
The GERL generates a 2nd harmonic of fDEM at the CFIA output. This is
because the offset of CB3 (Fig. 5.13), after being chopped by CH10, gives rise to
C21 Vout+
R21
C11
CH1 CH5 CH6 CH7
Vin+ + - + - + - Vfbk+
Gm3 Gm2 2 Gm1 R1
Vin- - + - + - + Vfbk-
Clk1 f1 f2 f2
Delta_It1
C12
CH2 R22
C22 Vout-
+ -
Clk2
Gm4 C41
CH8
CH3 - +
C3 Cascode Cascode
Gm5
Delta_It2 Buffer 2 Buffer 1
f1 C42
Clk2
CH4 CH10 CH9 C51
Vfbk+ Comparator
C7 Cascode Cascode
Vfbk- Gm6 Buffer 4 Buffer 3
Clk1 f1/4 C52 Q1
Fig. 5.13 CFIA with ORL and GERL in DEM phase 2
Fig. 5.14 The clock timing DEM transition interrupts the

of fchop1 and fDEM (fchop1/ half clock period of fchop1
fDEM = 1/2, 2, 4)
fchop1
fDEM (=2fchop1)
fDEM (=fchop1/2)
fDEM (=fchop1/4)
chopper ripple across C7. This ripple voltage is converted by Gm6 into differential
currents, which modulate the tail currents and consequently the transconductances
of Gm3 and Gm4 at fDEM. This results in a modulated gain mismatch at fDEM, which
the swapper modulates to even harmonics of fDEM. To ensure that the ORL does
not sense these harmonics, we chose fDEM = fchop1/4, making the interference
between these two loops negligible.
The noise contribution from the GERL is negligible because it is attenuated by
the ratio of Gm3/Gm6 = 480 and suppressed by the finite CMRR of the input
stages. Gm6 is implemented as a resistor-degenerated differential pair. Its bias
current is chosen just large enough to cover the maximum expected mismatch
(2 %) between Gm3 and Gm4, thus minimizing the DEM ripple with zero input.
Since Gm3’s tail current is 55 lA, the bias current of Gm6 is around 1.1 lA.
VDDA VDDA
Vb1
Vb1
IGE+
Vb2 IGE+
* Vb2
*
Cpar
In+
M1 M2
VCM
* M3 M4 *
In-
Out + M1 M2
Cpar
Out -
In+ M3 M4
* *
Low Vth PMOS In-
*
Out+
(a) (b) Out-
Fig. 5.15 a Input (or feedback) Gm stage with low-threshold cascodes; b Input (or feedback) Gm
stage with class-AB boot-strap of the back- gate with unity gain buffers
5.7.1.2 Implementation of the Input and Feedback Gm Stages
The use of DEM ensures good gain accuracy, implying that the input and feedback
transconductors can be implemented with simple differential pairs. Figure 5.15a
depicts the input (or feedback) Gm stages made by this simple differential pair.
Since they typically operate at different common-mode voltages, their CMRR was
enhanced by cascoding the input transistors with low-threshold devices [1].
During the DEM transitions, the CM voltages of the transconductors will change
abruptly. As a result, the parasitic capacitances between the substrate and the n-wells
of the input devices, shown as Cpar in Fig. 5.15 a, will be charged and discharged,
causing large CM current spikes in the input stages. To bypass these spikes, two
class-AB boot-strap unity-gain buffers are employed (Fig. 5.15b). Their detailed
schematic is shown in Fig. 5.16. The n-wells of the input transistors M1, M2, and their
cascodes M3, M4 are actively bootstrapped by class-AB buffers consisting of
M5–M10. M7–M9 act as level shifters to accommodate the bias voltage of the class-
AB stages. The source followers M5 and M6 provide a low-impedance path to
ground, while M10 provides a low-impedance path to the supply. As a result, the
bootstrap circuit effectively bypasses the DEM spikes to the supply rails.
5.7.1.3 Polarity Control Quantizer Q1 in Analog GERL
As discussed in Sect. 5.3.2, the polarity of the GERL is linked to the polarity of the
output signal by a low-offset quantizer and the estimated offset of the quantizer
needs to be less than 60 lV. This offset level is achieved by adding an auto-zeroed
preamp [8] preceding the quantizer, as shown in Fig. 5.17.
Fig. 5.16 Schematic VDDA

diagram of the input (or
feedback) Gm stages with Vb1
class-AB bootstrapping of the IGE+
back-gates Vb2
M10
*
M9
M7 M8
* Ispike *
M5 M6
* *
M1 M2
Cpar
In+
* M3 M4 *
In-
Out+
Out-
The auto-zeroed preamp works as follows. In the auto-zero phase, the input
terminals are shorted by clock UAZ, and the output of the cascode connects the
offset nulling loop, which consists of an integrator and an auxiliary stage gm2. The
loop integrates the differential current generated by the cascode until the current
through gm2 nulls the offset current. At the end of auto-zeroing cycle, the charge
injection mismatch errors cause residual offset, as determined by
Vos1 Vos2 gm2
Vres;offset þ þ DVinj ð5:15Þ
gm2 R gm1 R gm1
where Vos1 is the offset of gm1, Vos2 is the offset of gm2, and DVinj is the charge
injection caused by switches which are located at the output of gm1 and controlled
by UAZ. By choosing a ratio of gm1/gm2 = 60, the offset induced by charge
injection errors is attenuated when input-referred. The amplification phase is
controlled by clocks UAMP, in the manner shown in Fig. 5.17. According to
simulations, this auto-zeroed quantizer achieves a less than 60 lV offset. At a
closed-loop gain of 100, this translates into an error of 0.6 lV at the input of the
CFIA, which is small enough for most applications.
Preamp Quantizer
AMP Vos1 AMP Vout+
Vin+ + -
AZ gm1 R
Vin- - +
Vout-
AZ
Vos2 ΔVinj
- +
AZ
gm1 /gm2 = 60 gm2
+ - AMP
Cint
Fig. 5.17 Circuit diagram of an auto-zeroed preamp and quantizer
Digitally – assisted GERL

Vout+ Vout-
Vdec
IGE 10bit fDEM/8
Q3 SW1 C6 C51
Vint,GE CH9 CH10
+
Counter
Gm6 CB4 CB3
DAC
EN UP/DN!
IGE- fCTRL C52
fDEM
FPGA Polarity control
CNT_EN
Q2
TDEM
fDEM
fCTRL 7 TDEM
Integrate
TDEM
Chopper ripple Dump
across C6
Decision
moment of
Quantizer
Fig. 5.18 Block diagram of the digitally-assisted GERL with the associated timing diagram

with Digitally-Assisted Gain Error Reduction Loop
As discussed in Sect. 5.4, the analog GERL needs to re-settle when an input step
re-appears after a long zero-input. To avoid this, the analog integrator can be
replaced by a digital integrator to store the mismatch information in the digital
domain. The result is the digitally-assisted GERL shown in Fig. 5.6 and whose
detailed implementation is depicted in Fig. 5.18.
To mitigate the effect of the comparator Q3’s offset and hysteresis, a chopped
integrate-and-dump pre-amplifier was realized by integrating the demodulated
DEM ripple on C6 for seven DEM periods (Fig. 5.18). The decision of the com-
parator is then made. During the next DEM period, SW1 resets the voltage on C6.
Thus, the gain of the preamp is 16 9 7 = 112 (41 dB), where C51/C6 = 16. This
gain is sufficient to reduce the offset of Q3 to below 60 lV. The timing of fCTRL
ensures that the positive time is seven times the negative time in one DEM period
(Fig. 5.18). The quantizer Q3 makes the decision exactly at the end of the seventh
DEM period, thus it does not sample the residual ripple caused by the chopped
offset in CB3. Since the ripple is filtered out by this discrete-time sampling, the
second harmonic of fDEM is eliminated in the digitally-assisted GERL.
5.7.2.1 10-bit DR DAC Implementation
As discussed in Sect. 5.4, the output of the DAC will toggle between two LSBs,
giving rise to a limit cycle. A DAC with 10-bit resolution is required to push this
limit cycle below the CFIA’s noise level and also provide enough dynamic range
to suppress the DEM ripple. It was decided to implement the digital logic for the
digitally-assisted GERL off-chip, therefore, an over-sampling DR DAC is a good
choice since it eases off-chip implementation. Furthermore, it uses over-sampling
and noise shaping instead of component matching to achieve a high resolution. By
employing a high OSR, the quantization noise can be shaped to a high enough
frequency. Thus, a small and compact on-chip RC filter is enough to filter out the
shaped quantization noise. Moreover, dithering can be employed to de-correlate
the quantization noise and suppress the limit cycles, a first-order digital DR
modulator was found to be good enough.
Figure 5.19 depicts the block diagram of the over-sampling DR DAC. It con-
sists of an interpolation filter, a 1-bit first-order digital DR modulator and an
analog low-pass filter (LPF). The ten-bit counter is updated at fCNT = 1 kHz, thus
the signal bandwidth at the counter output is quite low (around 1 Hz). Therefore,
the interpolation filter can be implemented as a simple zero-order-hold, which is
sufficient to notch out signal images.
The oversampling ratio of the 1-bit first-order digital DR modulator is chosen to
be 8192 in order to push the quantization noise to a high enough frequency of
8192 kHz, which is then suppressed by the succeeding LPF. The quantizer outputs
the most-significant-bit (MSB) of the register. The multiplier of 512 is realized by
wiring the 1-bit modulator output to the tenth bit from the least-significant-bit
(LSB) of a subtractor [9]. In this way, the modulator changes the 10-bit signal from
the counter into a 1-bit digital signal which contains the gain error correction
signal and the spectrally shaped quantization noise. A 2nd-order RC low-pass filter
with a cut-off frequency of 2 kHz is enough to suppress the shaped quantization
noise. The filtered output signal feeds to Gm6 and generates two differential cur-
rents which compensate the mismatch of Gm3 and Gm4. This dither signal is
×8192 Interpolation Filter
10bit digital signal

from counter Sinc
fs
1 8192
BW=1Hz
Sample rate
relative to
input rate •••••
0 fs 2fs 3fs 4fs 5fs
Dither
1bit
10bit MSB DAC
output
Reg.
A[9:0]
B[9]
512
Analog
Digital modulator LPF
Fig. 5.19 Implementation of 1-bit first-order DR DAC
Output
D Q D Q D Q D Q D Q D Q
Fig. 5.20 Pseudo-random generator for the dither signal
generated by a 6-bit pseudo-random generator, as illustrated in Fig. 5.20. It con-

sists of six D-flipflops and an XOR inverter gate.
5.7.2.2 Polarity Control Quantizer Q2 in Digitally-Assisted GERL
Figure 5.21 shows the block diagram of the three-level quantizer. To reduce the
kick-back effect and offset of the quantizer, a pre-amplifier with MOS input is
implemented preceding the quantizer. Its schematic diagram is shown in Fig. 5.22.
The reference voltages Vref+ and Vref– interchange with each other once during
each comparison cycle to generate two threshold levels ±200 mV, as shown in
Fig. 5.22.
Quantizer Vout+
Vin+ + -
Preamp
Vin- - +
Vout-
Gain = 6
Fig. 5.21 Block diagram of the three level quantizer
Fig. 5.22 Schematic VDD

diagram of the preamp
preceding the quantizer
(Vref+-Vref- = ±200 mV)
Vin+ Vin- Vref+ Vref-
Vpre- Vpre+
VCM
The MOS input of the preamplifier offers high input impedance and a low input
bias current. The switching transitions in the quantizer thus have negligible effect
on the amplifier output. To reduce the recovery time of the preamp, four clamping
diodes are connected there. The resulting output swing of this preamp is limited to
the VGS of two diodes, which is about 1.4 V.
The CFIA with the ORL and the GERL was implemented in a 0.7 lm CMOS
process with low-threshold transistors, linear capacitors and high-resistivity poly
resistors. The 5 mm2 chip micrograph is shown in Fig. 5.23. Both the analog and
digitally-assisted GERLs were implemented. For flexibility, the first-order DR
DAC and the counter were implemented in an FPGA. The CFIA with the analog
GERL only consumes a 290 lA supply current (NEF = 11.2) and the CFIA with
the digitally-assisted GERL consumes a 295 lA supply current, including that
which is consumed in the digital circuitry. The ORL and the GERL draw 14 % and
10 % of the total supply current, respectively. Measurements on 30 samples show
that the CFIA achieves 3 lV offset and 15 nV/°C offset drift.
Fig. 5.23 Chip micrograph

of the CFIA with GERL
5.8.1 Noise
In the digitally-assisted GERL, the quantization noise from the 1-bit DR DAC is
sufficiently suppressed by the second-order RC LPF, and therefore it does not
contribute to the input-referred noise PSD. Figure 5.24 depicts the measured noise
PSD of the amplifier with the analog GERL and digitally-assisted GERL,
respectively, showing that they achieve the same noise level of 17 nV/HHz.
5.8.2 Output Ripple Measurement
Figure 5.25 depicts the output ripple measurement of (a) DEM only; (b)
DEM ? analog GERL; and (c) DEM ? digital GERL. All the measurements are
done with the amplifier configured at a gain of 100 and the ORL ‘‘on’’. As seen
from the results, the ORL and the GERL indeed can work independently with each
other. To verify that the digitally-assisted GERL provides sufficient dynamic range
to suppress the DEM ripple, a relatively large input signal of 30 mV is fed to the
amplifier. The analog GERL suppresses the DEM ripple by more than 40 dB from
4.6 mV down to 37 lV, which varies at most 0.52 lV/°C over temperature. The
digitally-assisted GERL suppresses the output DEM ripple below 47 lV. Fur-
thermore, it suppresses the second harmonic of fDEM down to 18 lV, which is
2 9 smaller than that in the analog GERL (35 lV). This confirms the analysis
described in Sect. 5.7.2. The residual chopper ripple amplitude is 32 lV, close to
the residual DEM ripple.
Fig. 5.24 Output noise spectrum of the CFIA with analog and digitally-assisted GERL (from
1 Hz to 100 kHz)
Figure 5.26 shows the output ripple measurement with an input CMV at 1.25 V
and a feedback CMV at 2.5 V: (a) DEM only; (b) DEM ? analog GERL;
(c) DEM ? digital GERL. Due to the different CM voltages, the residual DEM
ripple at fDEM increases to 60 lV, and ripple at other harmonics also increases a
bit. The CM dependence of the input stages results in some interaction between the
ORL and the GERL, thus the residual chopper ripple increases to about 90 lV.
However, at a gain of 100, the total input-referred ripple (*2 lV) is still low
enough for most applications.
5.8.3 INL
Figure 5.27 shows the typical integrated nonlinearity INLmeasurement results in

three cases: (a) without DEM; (b) with only DEM; and (c) DEM ? GERL. The
initial INL of the amplifier is 25 ppm. The use of DEM improves the amplifier’s
linearity from 25 ppm to 6 ppm (at a gain of 100). This is because that the use of
DEM shifts the transfer functions of Gm3,avg and Gm4,avg closer to each other
(Fig. 5.9c). Thus, their nonlinearity improves with DEM. Turning on the GERL
further reduces the INL to 4 ppm (Fig. 5.27). This is because that the GERL nulls
the mismatch to zero, after which the transfer functions of the two Gm stages
become completely aligned with each other (Fig. 5.9d). As a result, the linearity
improves further. The same explanation goes for the effects of DEM and the
Fig. 5.25 Output ripple (a)

measurement with GERL
‘‘on’’ and ‘‘off’’ (at a gain of
100, fDEM = fchop1/
4 = 8 kHz). a DEM only;
b DEM + analog GERL;
c DEM + digital GERL
(b)
(c)
Fig. 5.26 Output ripple (a)

measurement with GERL and
input CM at 1.25 V and
feedback CM at 2.5 V (at a
gain of 100, fDEM = fchop1/
4 = 8 kHz). a DEM + analog
GERL; b DEM + digital
GERL
(b)
GERL on gain accuracy and gain drift. Since the use of DEM and GERL improves
INL, the input range of the CFIA also increases. Without DEM, the INL is 25 ppm
within ±30 mV. Turning on the DEM and GERL, the input range extends
to ±120 mV with an INL of 10 ppm.
5.8.4 Gain Accuracy and Gain Drift
The use of DEM reduces the CFIA’s worst case gain error from 0.6 % to 0.01 % when
Gm3 and Gm4 are at the same CM voltage of 2.5 V. Under these conditions, the use of
DEM reduces the maximum gain drift from 300 ppm/°C to 9 ppm/ °C (Fig. 5.28) (11
Fig. 5.27 Measured INL of

the CFIA. a No DEM;
b DEM only; and
c DEM ? GERL
Fig. 5.28 Measured gain 4 (a)

drift histogram of the CFIA. 3
a No DEM; b DEM only 2
(output ripple present); and 1
c DEM ? GERL)
0
0 50 100 150 200 250 300
Number of samples
4 (ppm/°C)
(b)
3
2
1
0
0 1 2 3 4 5 6 7 8 9 10
4 (c) (ppm/°C)
3
2
1
0
0 1 2 3 4 5 6 7 8 9 10
Gain drift (ppm/°C)
samples). Turning on the GERL reduces it even further, to 6 ppm/°C (11 samples). Due
to the limited CMRR of the input stages, the gain error increases to 0.06 % when the
input of Gm3 is at 0 V, while the input of Gm3 is at 2.5 V. Other measurement results
include a typical CMRR of 127 dB and a typical PSRR of 130 dB.
CFIA Input
40 mV
Long zero input
0 3 6 9 12 15
Time (ms)
ORL settles
CFIA Output (V) within 0.8ms
4
Long zero input 1

0
0 3 6 9 12 15
Time (ms)
GERL integrator
output (mV)
300
Digital GERL
150 Analog GERL
Long zero input
0
0 3 6 9 12 15
Time (ms)
-150
Fig. 5.29 Settling behavior of the amplifier when a step input re-appears after a long zero input
5.8.5 Settling Behavior of Analog GERL and

Digitally-Assisted GERL
Figure 5.29 shows the measured settling behavior of the CFIA with a 40 mV input
step appears after a long period with zero input. It can be seen that the integrator
output of the analog GERL settles in an exponential manner. In contrast, the state
of the digitally-assisted GERL remains stable, since after start-up, it stores the
mismatch in the digital domain. Therefore, it behaves as a static mismatch com-
pensation loop for zero input. These results confirm the analysis in Sects. 5.3.3 and
5.4.
134
Table 5.1 Performance summary and comparison with state-of-the-art

This work Witte Pertijs Fan Sakunia
[6] [4] [3] [10] [5]
Year 2011 2009 2010 2010 2011
Supply current (lA) 290 (analog GERL) 850 1700 143 480
Input noise PSD (V/HHz) 17n 136n 27n 21n 28n
1/f noise corner (Hz) 1m – – 1 –
CMRR (dB) 127 130 142 137 122
PSRR (dB) 130 114 138 120 128
Gain error (Absolute) Dif CM: 0.06 % (Untrimmed) 0.1 % (Trimmed) 0.1 % (Untrimmed) 0.53 % (Untrimmed) Dif CM: 0.04 % (Trimmed)
Gain drift (ppm/°C) 3 (typ) 6 (max) – 3 (typ) 16 (max) – 2.1 (typ)
Offset (lV) \3 \5 \2 \2 \4 lV
Offset drift (nV/°C) 12 – 3 – 4 (typ)
GBW (kHz) 800 (stable when gain [20) 640 800 900 –
INL 4 ppm – – – –
Die area (mm2) 5 2.5 2.5 1.8 1.48
GBW/Isupply 2.8 0.8 0.5 6.3 –
NEF [11] 11.2 (analog GERL) 152.9 43 9.6 24
5 A Chopper Instrumentation Amplifier
5.9 Benchmark and Conclusions 135
5.9 Benchmark and Conclusions
This chapter addresses the major disadvantage of the current-feedback instru-

mentation amplifier (CFIA): its limited gain accuracy, which is due to the mis-
match between its input and feedback transconductors. To reduce this mismatch,
dynamic element matching (DEM) is applied to modulate the mismatch to the
DEM frequency. Since DEM ensures good gain accuracy, the input and feedback
transconductors can be implemented with simple differential pairs, thus resulting
in an improved power efficiency compared to CFIAs with resistor-degenerated
stages [3, 4]. To suppress the signal-dependent DEM ripple, a gain error reduction
loop (GERL) is proposed to continuously cancel the Gm mismatch, thus elimi-
nating the need for trimming. The residual DEM ripple then becomes signal-
independent. This feature makes the GERL more attractive than trimming [5],
because trimming can not eliminate the signal dependency in the DEM ripple.
Moreover, trimming can not compensate for temperature drift. The use of DEM
and the GERL ensures that the transfer functions of the input and feedback
transconductances are aligned, resulting in an improved performance in terms of
gain error, gain drift and linearity.
Since the loop gain of the GERL is signal-dependent and is zero for zero input,
the analog GERL needs to resettle whenever an input step reappears after long
zero input. To avoid resettling, a digitally-assisted GERL is implemented to store
the mismatch in the digital domain and compared with the analog approach.
During the DEM transitions, the abruptly changing CM voltages cause CM
current spikes in the input stages. To avoid overloading, the back-gates of the input
and feedback transconductors are actively class-AB boot-strapped, so as to bypass
the spikes to the supply rails.
In summary, the advantages of DEM combined with a GERL compared to other
gain error reduction techniques are:
Applying DEM to two simple differential pairs Ô good power efficiency
Dynamic cancellation loop Ô no need for trimming
After suppression by the GERL, the DEM ripple becomes signal-independent
The measured performance of the CFIA is summarized in Table 5.1 and
compared with the state-of-the-art [3, 4, 10]. Without trimming, it achieves a gain
error of less than 0.06 % and a maximum gain drift of 6 ppm/°C in a power
efficient manner (NEF [11] = 11.2). Compared to a CFIA with similar gain
accuracy but which uses resistor-degenerated stages [3], this represents a
4 9 improvement in power efficiency, which is equivalent to a 16-fold improve-
ment in power when the same noise level is achieved [11]. However, it should be
noted that the CFIA with resistor-degeneration stage has a wider input range
(about ±500 mV compared to ±120 mV). Compared to a CFIA with similar
power efficiency [10], this represents a 9 9 improvement in gain accuracy.
Compared to a ping-pong-pang CFIA, it achieves comparable gain error without
trimming and with 2 9 improved power efficiency. These measurement results

confirm that the combination of DEM and a GERL is a power-efficient manner to
improve the gain accuracy, gain drift and linearity of a CFIA.
References

amplifier with a 1 mHz 1/f noise corner and and AC-coupled ripple reduction loop. IEEE J
2. Slattery C, Nie M (2005) A reference design for high-performance, low-cost weigh scales.
Available at http://www.analog.com/library/analogDialogue/archives/39-12/weigh_scale.
html
employing ping-pong auto-zeroing and chopping. IEEE J Solid-State Circuits
45(10):2044–2056
with 5 lV offset for bidirectional high-side current-sensing. In Proceedings of the IEEE
instrumentation amplifier with 0.04 % gain error. Paper presented at the IEEE Symposium on
VLSI Circuits, pp 60–61
with a gain error reduction loop and 0.06 % untrimmed gain error. In Proceedings of the
IEEE ISSCC, digital technical papers, pp 244–245
7. Bakker A, Thiele K, Huijsing JH (2000) A CMOS nested-chopper instrumentation amplifier
with 100 nV offset. IEEE J Solid-State Circuits 35(12):1877–1883
8. Degrauwe M, Vittoz E et al (1985) A micropower CMOS-instrumentation amplifier. IEEE J
Solid-State Circuits SC-20(3):805–807
9. Terada J, Nishimura K et al (2008) A 10.3 Gb/s burst-mode CDR using a DR DAC. IEEE J
10. Fan Q, Huijsing JH, Makinwa KAA (2010) A 21nV/HHz chopper-stabilized multipath
current-feedback instrumentation amplifier with 2 lV offset. In Proceedings of the IEEE
ISSCC digital technical papers, pp 80–81
instrumentation amplifier for medical purposes. IEEE J Solid-State Circuits SC-
22(6):1163–1168
Chapter 6
Read-Out Integrated Circuits
This chapter presents the design and implementation of a read-out IC that is

intended for interfacing thermistor bridge, thermocouple, strain gauge and Hall
sensors. The read-out IC consists of a current-feedback instrumentation amplifier
(CFIA) and an analog-to-digital converter (ADC). The CFIA provides high input
impedance for bridge read-out and also relaxes the ADC’s noise and offset
requirements. The ADC converts the amplified sensor signal into a robust digital
output [1, 2].
The CFIA employs a topology that has been discussed in Chap. 4, thus this
chapter will mainly focus on the design of the ADC and the system-level
collaboration between the CFIA and the ADC. Firstly, the requirements and the
architecture design of the ADC are discussed. This is followed by the system-level
collaboration between the CFIA and the ADC. System-level chopping is employed
to chop the complete read-out chain. The resulting modulated offset is then
averaged out in the digital decimation filter during multiple conversions.
Furthermore, the read-out IC explores the power of digital signal processing
succeeding the ADC to help the CFIA to improve its gain accuracy and gain drift.
Finally, the detailed circuit implementation and the measurement results are pre-
sented. Compared to previous work, our read-out IC achieves state-of-the-art
performance in terms of offset and drift.
6.1 ADC Requirements
Figure 6.1 shows a simplified block diagram of a bridge read-out system with the
proposed read-out IC. It is designed to meet the specifications of typical bridge
sensors and thermocouples (see Sect. 1.6).
As discussed in Sect. 4.2, the CFIA consists of three gain stages, as shown in
Fig. 6.2. The closed-loop gain is defined by precision external resistors. The input

138 6 Read-Out Integrated Circuits
Fig. 6.1 Simplified block V ref+

diagram of a bridge-readout R0
Read-Out IC
system with the read-out IC
Rx R1 Digital
Output
CFIA ADC
R1 Rx
V ref-
and intermediate stages are both chopped, thus providing sufficient gain to sup-
press the 1/f noise of the unchopped output stage down to 1 mHz [3]. The input
referred noise PSD of the CFIA is 15nV/HHz. The 1/f noise corner frequency of
unchopped CFIA is around 10 kHz, and thus the chopping frequency is chosen to
be 30 kHz, which is well above the corner frequency to remove 1/f noise.
To achieve a better power efficiency, the CFIA should dominate the noise,
offset and gain accuracy of the overall read-out IC [4], and the ADC is designed to
maintain the CFIA’s performance. Based on the detailed specifications of the
sensors and the read-out IC (Table 1.3), the requirements of the ADC are derived.
Bandwidth and Resolution
The output of the CFIA is digitized by an incremental DR ADC consisting of a
DR modulator and a decimation filter. It is well known that DR ADCs provide high
resolution with less accurate analog building blocks due to noise shaping and over-
sampling [5]. Here, to maintain the SNR of the CFIA, the target resolution of the
DR ADC is 21-bit with a conversion time less than 0.2 s (BW = 5 Hz). The
reference voltage of the DR ADC is set to be 5 V with 4 V input full scale, which
corresponds to a noise density of 600nV/HHz in 5 Hz bandwidth. Given the
CFIA’s gain of 100, the input referred noise density of the DR ADC is 6nV/HHz,
which is still lower than that of the CFIA (16nV/HHz), as shown in Fig. 6.3.
Compared to the continuous-time (CT) counterpart, switched-capacitor (SC)
DR ADC is chosen here because of its high accuracy [5, 6]. The CT DR ADC uses
resistors or Gm stages to convert the input voltage to a current. Its gain error and
nonlinearity are usually determined by the resistor- or transconductor-matching.
While in the SC DR modulator, its gain error and nonlinearity are determined by
capacitor-matching, with careful layout, which has superior matching over resis-
tors or transconductors.
1/f Noise Corner
As described in Chap. 4, the CFIA already achieves a 1 mHz 1/f noise corner
with multi-stage chopping. The ADC also aims to achieve a 1 mHz 1/f noise
corner so that even when the closed-loop gain of the CFIA reduces in some
applications, the read-out IC can still maintain the 1 mHz 1/f noise corner.
Gain Error
The gain accuracy of the ADC should be much better than the CFIA itself, and
thus the gain accuracy of the read-out IC is not degraded by the ADC. The ADC
6.1 ADC Requirements 139
C2
C1
f1 f1 f2 f2
Vin Vout
Gm3 Gm2 Gm1
CH1
f1
CH3 CH4 CH5 R1
Vfbk Vfbk
Gm4
CH2 R2
Fig. 6.2 Simplified block diagram of CFIA with the input and intermediate stages chopped
R0 Vref+
1.2V Variable gain 5V
11kΩ 11kΩ 5V
Rx R1 Digital
+ Output
Vcm,in Vcm,out
0.6V CFIA 2.5V ADC
11kΩ -
11kΩ Gain=100
R1 Rx
Vref-
Noise PSD Noise PSD Noise PSD

14nV/ Hz 16nV/ Hz 600nV/√Hz
Gain = 100
6nV/√Hz
Fig. 6.3 Noise budget of the total read-out IC
aims to achieve a gain error of less than 20 ppm. The bridge and the ADC are
connected in a ratio-metric structure and thus the accuracy requirement of the
ADC’s reference is much relaxed.
Offset
The gain of the CFIA suppresses the offset of the ADC when input-referred. To
achieve a 1 lV input-referred offset with the CFIA at a gain of 100, the input-
referred offset of the ADC should be less than 100 lV. Table 6.1 summarizes the
specifications of the ADC.
Table 6.1 Specification Topology Second-order

summary of the 21-bit
SC DR modulator Conversion time \0.2 s
Output noise 1.34 lV
Noise PSD 600nV/HHz
Input range ±4 V
Reference voltage 5V
Thermal noise limited SNR 129 dB
Quantization noise limited SNR 129 dB
Number of bits 21bit
Offset \100 lV
1/f noise corner @ 600nV/HHz 1 mHz
Gain error \20 ppm
Supple current \100 lA
Fig. 6.4 Simulated 1

quantization noise over the
N = 1300
(LSB) relative to 20bit
input range
Quantization noise
0.5
-0.5
-1
-4 -3 -2 -1 0 1 2 3 4
Input Range (V)
6.2 Architecture Design of the ADC
6.2.1 Modulator Topology
As discussed above, the DR ADC aims to achieve a quantization noise limited

resolution of 21-bit (129 dB) within a 0.2 s conversion time. A 1-bit quantizer is
chosen because of its inherent linearity. As will be described in Sect. 6.3.2, the
read-out IC is designed in such a way that one decimated digital output is obtained
with four conversions. To achieve 21-bit resolution in the ADC, the resolution of
each ADC conversion should be 20-bit. To achieve such a high resolution, a first-
order incremental DR modulator requires clock cycles larger than 32768 per
conversion [5], corresponding to a sampling frequency fS of 650 kHz. Such a high
sampling frequency implies undesirable high power consumption. Second, the
first-order structure generates periodic sequences, so called limited cycles, for DC
input signals. This issue can become even more severe due to finite DC gain of the
opamp, which causes loss in the integrator and introduces dead-band that could
lock into a limit cycle.
6.2 Architecture Design of the ADC 141
Vin E 1 1 Vout
b1 b2 a3 b1 1/4
z-p1 z-p2
b2 2/3
c1 1
c1 c2 c2 1/2
a3 1
Fig. 6.5 Block diagram of a second-order DR modulator with feedback (FB) topology
To reduce the required clock frequency and errors due to dead band, a second-
order DR modulator is preferable. For a second-order incremental modulator, it
needs 1300 clock cycles per conversion [5]. This is confirmed by Matlab simu-
lation result shown in Fig. 6.4, showing that the quantization noise is less than
±0.5LSB of 20 bit within the entire input range. For one decimated output, four
conversions means 5200 clock cycles. To achieve a conversion time of less than
0.2 s, the minimum sampling frequency is 26 kHz. To leave some margin, a
30 kHz sampling frequency is chosen here since it is the same as the chopping
frequency of the CFIA and furthermore, the ADC and CFIA can be easily syn-
chronized to each other.
Although with a third-order modulator, a sampling frequency fS of 3.2 kHz is
sufficient to achieve a resolution of 20 bit (OSR = 160), there are some other
trade-offs that need to be considered. The chopping frequency of the input stage in
the CFIA has to be higher than the 1/f noise corner of 10 kHz to remove the 1/
f noise. With an ADC’s sampling frequency fS of 3.2 kHz, the under-sampling of
the chopper ripple ([10 kHz) is more severe than that with fS of 30 kHz. Although
a high sampling frequency fS reduces the under-sampling errors, it would consume
more power to meet the settling requirements. Moreover, the increased power in
the ADC does not improve the resolution of the read-out IC, since the noise level
of the CFIA dominates. In summary, a third-order modulator is not an optimum
solution for our applications.
One possible topology for implementing the DR modulator is using the feed-
back (FB) topology shown in Fig. 6.5. In this topology, the entire output signal,
consisting of both the input signal and the quantization noise is fed back to the
input of each integrators. Therefore, the first integrator output must provide a
strong compensation signal. This means that the error signal fed to the first inte-
grator is also relatively large, containing a strong signal component. The opamp in
the first integrator thus needs to consume large power to meet the stringent line-
arity requirement.
To reduce the signal swing and ensure reduced swing with an average of zero at
the first integrator output, a feed-forward path a2 can be applied [7] (Fig. 6.6). This
results in a hybrid feed-forward and feedback (hybrid-FF–FB) topology.
An alternative topology is the input feed-forward (FF) topology [8] shown in
Fig. 6.7. A feed-forward branch (a2) is applied from the first integrator output to the
second integrator output to ensure the stability of the modulator. The feed-forward
path (a1) is inserted to reduce latency in the feedback loop from the input through a1,
a2
DC
b1 1/4
Vin E 1 1 Vout
b1 b2 a3 b2 2/3
DC 0 z-p1 0 0 z-p2 c1 1
DC c2 1/2
a3 1
c1 c2 a2 1/2
Fig. 6.6 Block diagram of a second-order DR modulator with hybrid feed-forward and feedback
(hybrid-FF–FB) topology
a1 0.4
a1 a2 0.8
a3 0.5
b1 0.3
b2 1
a2 c1 1
Vin E 1 1 ++ Vout
b1 b2 a3 +
z-p1 z-p2
c1
Fig. 6.7 Block diagram of a second-order DR modulator with input feed-forward (FF) topology
the quantizer and c1. The coefficients are chosen in such a way that the input signal
and the average of the bitstream feedback signal are well compensated, thus the
resulting error signal E only contains quantization noise, relaxing the linearity and
slew requirements of the opamp in the first integrator.
The output swings of the first and second integrators in the above three
topologies are simulated as a function of the DC input levels (Fig. 6.8). The
coefficients of these three topologies (indicated in the pictures) are chosen for the
same noise transfer function (NTF) and already optimized to reduce swings at
integrators’ output.
It can be seen from Fig. 6.8a that the output swing of the first integrator in the
feedback topology significantly increases when the input signal exceeds 0.6 Vref,
while those in the feed-forward and hybrid-FB–FF topologies remain below
1.1 V. As seen from Fig. 6.8b, the output swing of the second integrator in the
hybrid-FB–FF topology is much larger than that in the FF topology. Since the FF
(a) 2.5 (b) 2.5
Secnd Integrator (V)

Output Swing of the
Output Swing of the
First Integrator (V)
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Vin/Vref Vin/Vref
Fig. 6.8 Simulated maximum peak value of the integrator output swing as a function of the DC
input levels
Fig. 6.9 Pole-zero plot of Z-plane

the NTF 1
Im(z)
0.8
0.6
0.4
0.2
0
Re(z)
-0.2
-0.4
-0.6
-0.8
-1
-1 -0.5 0 0.5 1
topology exhibits the smallest output swings within the whole input range, it is
employed in this work.
STF and NTF of the DR Modulator
From the feed-forward topology in Fig. 6.7, the signal transfer function (STF)
and the noise transfer function (NTF) of the input feed-forward topology can be
calculated as
1
STF ¼ ð6:1Þ
c1
1 ðz p1 Þðz p2 Þ
NTF ¼ ¼
1 þ c1 LðzÞ ðz p1 Þðz p2 Þ þ a2 b1 c1 ðz p2 Þ þ a3 b2 b1 c1 ðz p1 Þ
ð6:2Þ
Table 6.2 Coefficients for b1 0.33

the second-order DR
b2 1
modulator in Fig. 6.7
a1 0.4
a2 0.8
a3 0.5
c1 1
Fig. 6.10 The STF and NTF 3.5dB

0
plot of the DR modulator Noise Transfer Function (dB)
STF
-20
-40
-60 NTF
-80
-100
-120
-140
-160
10-5 10-4 10-3 10-2 10-1 0.5

Normalized frequency (× f s)
It can be seen from (6.2) that the two poles of p1 and p2 in the loop-gain of the
modulator correspond to the two zeros at DC in the NTF. Since the bandwidth of
the modulator is quite narrow (0–5 Hz) and close to DC, the two zeros are chosen
to be at DC to sufficiently suppress the quantization noise near DC, as shown in the
pole-zero plot in Fig. 6.9. Thus, assuming infinite DC gain,
p1 ¼ p2 ¼ 1: ð6:3Þ
To ensure stable second-order noise shaping and reasonable swings at the
integrators’ output, Matlab simulations were done to optimize the coefficients of
the modulator to reduce swings at each integrator output and meanwhile obtain the
targeted SNR. The results are shown in Table 6.2. Figure 6.10 depicts the ideal
NTF and STF of the DR modulator with infinite DC gain. For stability purposes,
the out-of-band gain of the NTF near half of the sampling frequency is chosen to
be a rule of thumb value: 1.5 (3.5 dB) [5].
Figure 6.11 shows the FFT plot of the bit-stream of the modulator. It achieves
an SNR of 140 dB within a 5 Hz bandwidth with a 4 V DC input referenced to
5 V (Fig. 6.11a). The FFT plot of the Matlab simulation with a sine input of 4 V at
0.2 Hz is shown in Fig. 6.11b.
0
0
-20
-50
-40
-60 -100
dBFS
dBFS
-80
-100 -150
-120
-200
-140
-160
-250
-180
-200 -300
10-2 10-1 1 10 102 103 104 10 -3 10-2 10-1 1 10 102 10 3 10 4
Frequency (Hz) Frequency (Hz)
Fig. 6.11 a The noise spectrum of the DR modulator with 4 V DC input 4 V (reference to 5 V).
b The noise spectrum of the DR modulator with 4 V Sine input (reference to 5 V)
6.2.2 Non-Idealities in the DR Modulator
The previous section described the order, architecture, sampling frequency and
coefficients of the DR modulator, which define its quantization noise. This section
discusses other non-idealities of the SC DR modulator, such as kT/C noise finite
DC gain, offset, 1/f noise, and charge injection.
kT/C Noise
In a switched-capacitor DR ADC, its thermal noise is mainly limited by kT/C
noise. The ADC’s input-referred noise density was designed to be 600nV/HHz,
corresponding to a noise voltage Vn of 1.34 lV over a 5 Hz bandwidth. The full input
range is ±4 V, thus the corresponding SNR due to thermal noise is 129 dB. The OSR
is 3000 with a sampling frequency of 30 kHz. To leave some margin, the kT/C noise
is designed to be 1 lV, determined by the sampling capacitor as given by
2kT
¼ Vn2 ¼ ð1lVÞ2 ð6:4Þ
CS OSR
The resulting sampling capacitor CS is 3 pF. The factor of two accounts for the
two input sampling capacitors due to the differential topology.
Leakage
In a first-order modulator, the integrator leakage p limits its effective number of
bits (ENOB). Since the DC gain Adc determines the leakage
1
ADC ¼ ð6:5Þ
1p
So,
1
p¼1 ð6:6Þ
ADC
Vcom P1
(a)
P2 Cint1
P2d
P1d CS1 Vos
+ -
Vin Vx A0 Vout
- +
P1d CS2
P2d P2 Cint2
Vcom P1
Vcom
Qinj1 P1 Cint1
P2
P2d
Sampling
Qinj2 & autozero Integration
P1d CS1 Vos CS1 Vos
+ - + - P1
Vin Vx A0 Vout Vx A0 Vout P1d
- + - +
P1d CS2 CS2
P2
P2d
P2d
P1 P2 Cint2 t0 t1 t2
Vcom
(b) (c) (d)
Fig. 6.12 The operation of an auto-zeroed switched-capacitor integrator in two phases. a Auto-
zeroed Integrator, b auto-zero phase P1, c integration phase P2, d timing diagram
The width of the dead-band Dx in the modulator’s DC characteristic normalized

to its reference is given by [9]
1p 1p
Dxmax ¼ ð6:7Þ
1þp 2
Substituting (6.6) into (6.7), the maximum width of the dead-band is given by
1
Dxmax ð6:8Þ
2ADC
It can be seen from (6.8) that in a first-order modulator, the DC gain of the first
integrator limits the ENOB. A similar phenomenon occurs in a second-order
modulator [5]. To obtain an ENOB of 21bits, a combined DC gain of cascaded
integrators in the order of 123 dB is required, which can be easily achieved by two
integrators. The first and second integrators with a DC gain of more than 90 dB
and 60 dB, respectively, should be sufficient to suppress the error caused by the
integrator leakage.
Offset and 1/f Noise Suppression
The ADC aims to achieve the same 1/f noise corner as the CFIA at 1 mHz. This
is realized by auto-zeroing the first integrator [10] as shown in Fig. 6.12. In the
auto-zeroing phase, the finite DC gain A0 of the opamp in the first integrator results
Fig. 6.13 Simulated input- Input referred noise

referred noise spectrum of the 10-3
PSD (V/ Hz)
1st integrator with and
without auto-zeroing 10-4 Noise of ADC Without
Auto-zeroing
10-5 Noise of CFIA
10-6 53Hz
Noise of ADC w ith
1mHz Auto-zeroing
10-7
1m 10m 100m 1 10 100 1k 10k
Frequency (Hz)
in a non-zero overdrive voltage at the input of the opamp, so Vx is not exactly

equal to Vos. At the end of phase P1, Vx equals
A0
Vx ðt1 Þ ¼ A0 ðVos Vx ðt1 ÞÞ ) Vx ðt1 Þ ¼ Vos ð6:9Þ
1 þ A0
While at the end of phase P2, it equals
Vint ðt2 Þ
Vx ðt2 Þ ¼ Vos ð6:10Þ
A0
Using (6.9) and (6.10), the integrated charge in every period is given by

VOS Vint ðt2 Þ
Q ¼ CS fVin ðt1 Þ þ ðVx ðt2 Þ Vx ðt1 ÞÞg ¼ CS Vin ðt1 Þ þ ð6:11Þ
1 þ A0 A0
Due to the subtraction Vx(t2) - Vx(t1) in (6.11), the auto-zeroing technique
suppresses the offset and 1/f noise and their suppression ratio is determined by the
finite DC gain A0 of the opamp. To obtain a 1 mHz 1/f noise corner, simulation
shows that the opamp in the first integrator must have a DC gain of 160 dB. This
was achieved by using a two-stage Miller-compensated opamp with a gain-boosted
first stage and a class-A second stage.
Figure 6.13 shows the input-referred noise spectrum of the first integrator with
and without auto-zeroing as simulated using the Periodic Steady State (PSS) and
Periodic Noise (PNOISE) tools of Spectre RF [11]. Without auto-zeroing, the first
integrator exhibits a 1/f noise corner of 53 Hz and a noise PSD at 600nV/HHz. With
auto-zeroing, the 1/f noise corner is shifted down to 1 mHz with an increased noise
floor of 600nV/HHz caused by the double-sampling associated with auto-zeroing.
However, the overall noise level of read-out IC is still dominated by the CFIA.
Charge Injection and Clock Feed-Through
With 160 dB DC gain in the first integrator, the integrator leakage and the
residual offset due to finite DC gain, expressed by the third and second term in
(6.11), becomes negligible. The residual offset is then mainly due to charge
injection mismatch.
The auto-zeroed integrator uses a bottom-plate sampling scheme, as shown in

Fig. 6.12. A clocking scheme with delayed falling edges [5] is used to limit the charge-
injection error to that of only four switches: switches controlled by P1, P2. The switches
driven by P2, P1 open first and inject some charge Qinj1 and Qinj2 into CS and Cint,
respectively. Note that since these switches are connected to the virtual ground of the
amplifier, their charge injection can be considered as constant and signal-independent.
When the switches driven by P1d and P2d are open, the capacitors Cs1 and Cs2
are isolated if the parasitic capacitors are ignored for simplicity. The signal-
dependent charges of the switches only flow to the input sources and to ground,
respectively, and do not change the voltage on the capacitors of Cs1 and Cs2. Using
this clocking scheme, the charge injection is reduced to the charge injected by four
switches that are controlled by P1 and P2. The common-mode voltage of the
integrator is set at 2.5 V with a 5 V supply. To accommodate this common-mode
voltage, the switches driven by P1 and P2 are implemented with the minimum size
transmission gate switches composed of PMOS and NMOS switches.
Adding 10 % mismatch in switches and 10 mV offset in the opamp A0
(Fig. 6.12), the simulated input-referred offset of the DR modulator (suppressed by
the CFIA’s gain of 100) is about 1 lV. This residual offset will be removed by
system-level chopping which will be described in the Sect. 6.4.
Gain Accuracy and Linearity
Using the DR modulator topology shown in Fig. 6.14, the same sampling
capacitor CS1 is used to sample both the input and DAC feedback signals [6]. As a
result, the gain accuracy and linearity of the DR modulator is not limited by
component matching. Furthermore, as shown in Fig. 6.1, the ADC’s reference is
connected to the reference of the bridge transducers in a ratio-metric structure and
the gain accuracy of the reference is thus much relaxed. The gain error and
nonlinearity of the DR modulator would be mainly determined by the non-linear
DC gain of the opamp in the first integrator [12]. This nonlinear DC gain can be
modeled with a third-order polynomial approximation [12], given by
jvout j 3
ADC ðvout Þ ¼ ADC ð0Þ ð1 d3 ð Þ Þ ð6:12Þ
dmax
where ADC(0) is the DC gain at the midlevel output, d3 is the gain variation
coefficient, vout is the output swing and dmax is the maximum output swing.
As shown in Fig. 6.15, a DC gain of 100 dB and a d3 \ 0.4 are enough to
achieve a 21 bit INL.
As mentioned before, a DC gain of 160 dB is required in the opamp in the first
integrator to suppress the 1/f noise of the ADC down to 1 mHz. Thus, a two-stage
opamp is implemented with a telescopic gain-boosting first stage and a class-A output
stage to accommodate output swing. Figure 6.16 shows the DC gain characteristic of
the two-stage opamp vs. the output swing. It is clear that the gain is nonlinear and a
maximum gain is obtained when the output voltage is zero and it decreases as the
output voltage increases (d3 = 0.2). According to the simulation results in Fig. 6.15, it
is evident that a typical INL of 1 ppm can be achieved in the DR modulator.
a1
P1 a2
P2 Cint1 ++
P1d Cs1 a3 +
Vin+ Switched
+ - Bs
A1 capacitor
- + second
Vin- integrator a3 +
P1d Cs1 ++
P2 Cint1
P2d
P2d
a2 fS
P1
a1
Ratio-metric
measurement Bs
Vref+
Thermistor Bs
bridge Bs
reference Bs Vref-
Fig. 6.14 A DR modulator topology with the same capacitor for input and feedback sampling
Fig. 6.15 INL simulation 1

with DC gain of 100 dB and 0.8
d3 = 0.4 DC gain = 100dB
0.6 3 = 0.4
0.4
INL (LSB)
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
-4 -3 -2 -1 0 1 2 3 4
Input Range (V)
Settling
The settling at two places need to be taken into account in the design of the
read-out IC: (1) at the interface between the CFIA and the ADC; (2) in the first
integrator of the ADC.
The settling process is composed of two phases, slewing and linear settling thus
the total settling time is given by
tS ¼ tslew þ tsettling ð6:13Þ
Usually, 25 % of the total settling time is allocated to slewing, as typical design
practice [13], thus
Fig. 6.16 DC gain of the 170

first integrator versus output
swing
165
DC Gain (dB)
160
155
150
145
-1.5 -1 -0.5 0 0.5 1 1.5
Output Swing of the First Integrator (V)
tS ¼ 0:25tS þ 0:75tS ð6:14Þ

The slew time depends on the bias current and the charging capacitor, which
defines the slope of the sampled voltage; while the settling time depends on the
time constant s. With a sampling frequency of 30 kHz for auto-zeroing, the total
available settling time is half of the clock period, which is 16.7 ls. The resulting
slew time is 4.2 ls and the linear settling time is 12.5 ls.
To achieve a resolution of N bit (21-bit), the two settling error at the two places
(at the interface between the CFIA and ADC and in the first integrator of the ADC
must be both less than half an LSB. The required time constant s for linear settling
is given by [14]
tsettling
s0 [ ð6:15Þ
ðN þ 1Þ ln 2
with tsettling = 12.5 ls, the required time constant s0 is 0.82 ls.
The settling requirement at the interface between the CFIA and ADC is
determined by the time constant sCFIA of the CFIA, which is equal to Cload/Gm1 =
6pF/250 lA/V = 0.024 ls. Since sCFIA is much smaller than s0, the settling error at
the interface is negligible.
The settling of the first integrator is determined by its opamp since the time constant
of the switches and the sampling capacitor is negligible. For a sampling switch with a
W/L ratio of 5 lm/0.7 lm, the on-resistance is about 1 kX. Thus the associated time
constant is 0.003 ls, much smaller than the required time constant s0.
As discussed before, the opamp employs a two-stage topology, which offers an
optimal compromise between the output swing and the noise. The class-A output
stage accommodates the output swing and provides additional DC gain. The noise
bandwidth of the integrator is determined by the UGB of the opamp and the
feedback factor b [14]. The larger the noise bandwidth, the more noise aliasing
occurs. The total noise from the opamp is determined by the Miller compensation
capacitor CC. To make the opamp’s noise negligible, CC should be chosen larger
than the input sampling capacitor CS. In this design, CC is chosen 11pF large than
CS of 3pF. The bias current of the first stage in the opamp is determined by the
slewing requirement, given by
CC V 11pF 1V
Ibias ¼ ¼ 2:5lA ð6:16Þ
Dt 4:2 ls
where the maximum output swing is 1 V and the slew time is 4.2 ls. With a 2.5 lA
bias current, the transconductance gm of one transistor in the differential pair is
chosen to be 42 lA/V. The time constant of the first integrator is then given by [14]
1 CC
s¼ ¼ ð6:17Þ
2pfUGB b gm b
where b = 0.77(CS = 3pF, Cint = 10pF), gm = 42 lA/V and CC = 11pF, thus s
equals 0.34 ls, which is smaller than that required by the linear settling time of
12.5 ls. It can be seen that the settling is a bit over-designed. However, since the
CFIA consumes 82 % of the power consumption of the read-out IC, further
reducing the power consumption of the ADC will not lead to a significant
reduction in the total power consumption.
6.3 Gain Accuracy Improvement Techniques

in the Read-Out IC
In the read-out IC, the gain accuracy of a SC DR ADC, depending on capacitor

matching, is typically 0.1 %. The gain error of the CFIA, determined by the
mismatch of the input and feedback transconductors (in the order of 1 %), then
dominates gain accuracy of the read-out IC. For some strain gauge applications
[15], a gain error of 0.02 % is required. To achieve this, the same level of
transconductor matching is required. Such accurate matching cannot be expected
solely from precision layout technique. In addition, their gain drift should also be
less than 10nV/°C and 10 ppm/°C, respectively. Such drift performances cannot be
achieved with conventional one-shot trimming during production. This section will
explore the system-level collaboration between the CFIA and the ADC to improve
the CFIA’s gain accuracy and gain drift. In this way, the CFIA and the ADC are
linked or ‘‘tuned’’ to each other to improve the overall performance.
6.3.1 Dynamic Element Matching
Using precision external gain-setting resistors, the gain error of the CFIA is then
mainly determined by the mismatch of the input and feedback transconductors,
Gm3 and Gm4. To average out this mismatch, DEM is applied to swap their inputs
C2
C1
DEM
Vout DT ΔΣ bs
Vin Gm3 Gm2 Gm1 Modulator
Fixed R1
DC input Δ Vfbk
Decimation
Vfbk Gm4 R2 filter & Digital
back-end
GEC path External
VCAL,SET
Gm6 6-bit
DAC
fDEM Δ+Δ CAL,MAX -(Δ+Δ CAL,MAX) Δ-Δ CAL,MAX -(Δ-Δ CAL,MAX)

VCAL,MAX -VCAL,MAX
1conversion 1 decimated output
Vconv1 Vconv2 Vconv3 Vconv4
Fig. 6.17 Read-out IC employing dynamic element matching of Gm3 and Gm4 and the gain error
correction scheme
as shown in Fig. 6.17 [17]. This is implemented by an input multiplexer which is

the so called swapper [18]. After applying DEM, the average gain error becomes
1þD 1
1 þ 1þD D2
jGain Errorj ¼ j1 j ðD\\1Þ ð6:18Þ
2 2
From (6.18), DEM reduces the mismatch from D to D2/2. Thus, for a typical Gm
mismatch of 2 %, DEM reduces the gain error to 0.02 %.
For a CFIA, however, it typically operates with different input and feedback
CM voltage. Thus the input and feedback transconductors exhibit certain
CM-dependency due to finite CMRR, resulting in a CM-dependent mismatch Dcm.
Unlike the mismatch D mentioned in (6.18), this CM-dependent mismatch Dcm,
say, reference to feedback CM voltage, is always located at the Gm stage in the
input path. This implies that DEM will not suppress the Dcm and the resulting gain
error is given by
1þDþDcm
1 þ 1þD
1þD
cm
D2 D Dcm
jGain Errorj ¼ j1 j þ þ Dcm
2 2 2
ð6:19Þ
ðforD\\1Þ
6.3 Gain Accuracy Improvement Techniques in the Read-Out IC 153
From (6.19), the CM-dependent mismatch Dcm will not be suppressed by the
DEM. To improve gain accuracy, the transconductances of Gm3 and Gm4 should be
kept constant over the input CM range. As shown in [3], their CM dependency was
mitigated in a power-efficient manner by cascoding the input transistors with low-
threshold devices M3-M4 (Fig. 4.19). However, the CM-dependent mismatch Dcm,
still in the order of 0.12 % [16], is the main limitation on the final gain accuracy.
Circuit-level techniques will be proposed to address this problem in Sect. 6.6.1.
For simplicity, this issue will be neglected in the following analysis.
Another issue with different common-mode levels is that swapping the inputs of
Gm3 and Gm4 results in large spikes at the CFIA’s output. To avoid digitizing these
spikes, similar to the LF choppers, the multiplexer’s state is altered during the reset
period (1 ms) at the start of every ADC conversion (Fig. 6.17).
6.3.2 Digitally-Assisted Gain Error Correction Scheme
Although the use of DEM reduces the mismatch from D to D2/2, which is small
enough, the mismatch may still vary over temperature, causing a gain drift around
5 ppm/°C. To reduce gain drift further, the mismatch D itself should be minimized,
so as to achieve a better tracking between the input and feedback transconductors
to counteract temperature drift. Therefore, a gain error correction (GEC) scheme is
used to compensate for the static mismatch between Gm3 and Gm4. It is imple-
mented in a digitally-assisted manner with existing ADC to minimize complexity
and area of analog circuitry.
The GEC path, as shown in Fig. 6.17, consists of a decimation filter, digital
back-end and a 6-bit DAC which trims this mismatch by fine-tuning the tail
currents of Gm3 and Gm4 via the transconductor Gm6. With a 20-bit resolution in
the ADC, the output of the digital back-end is sufficiently precise. Assuming the
initial gain error of the read-out IC is 1 % and it needs to be reduced to 0.02 %
(a 34 dB reduction ratio), 6-bit resolution is required for the DAC. For flexibility,
the DAC is implemented with a DR DAC consisting of an interpolation filter, a
digital DR modulator and an RC low-pass filter (LPF).
The GEC scheme uses linear interpolation to find the calibration voltage during
the start-up. Figure 6.18 shows the concept of the linear interpolation. By applying
a fixed DC signal to the CFIA, the appropriate value of VCAL can be determined
within two DEM periods. In the first DEM period, the maximum calibration
voltage VCAL,MAX within the DAC’s output range is applied to the inputs of Gm6.
The decimated results of two conversions within one DEM period are given by
1 þ D þ DCAL;MAX
Vconv1 ¼ Vout;ideal ¼ Vout;ideal ð1 þ D þ DCAL;MAX Þ ð6:20Þ
1
1
Vconv2 ¼ Vout;ideal Vout;ideal ð1 D DCAL;MAX Þ ð6:21Þ
1 þ D þ DCAL;MAX
Fig. 6.18 Linear Gain Error

interpolation for the gain A
Δ+Δ CALMAX
error correction scheme
VCAL,SET
-VCALMAX DAC output
signal
C 0 VCALMAX
B
Δ-Δ CALMAX
where D is the initial mismatch of Gm3 and Gm4, and DCAL,MAX is the extra
mismatch caused by VCAL,MAX. The output-referred mismatch error due to
(D ? DCAL,MAX) can be determined from the difference of these two conversion
results (6.20) and (6.21), as written by
Vout;error;A ¼ 2Vout;ideal ðD þ DCAL;MAX Þ ð6:22Þ
which is noted as point A in Fig. 6.18.

In the second DEM period, the minimum signal -VCAL,MAX within the DAC’s
output range is applied to Gm6. The decimated result of one conversion in half of
the DEM period is expressed as
1 þ D DCAL;MAX
Vconv3 ¼ Vout;ideal ¼ Vout;ideal ð1 þ D DCAL;MAX Þ ð6:23Þ
1
The decimated result of one conversion in the other half of the DEM period is
given by
1
Vconv4 ¼ Vout;ideal Vout;ideal ð1 ðD DCAL;MAX ÞÞ ð6:24Þ
1 þ D DCAL;MAX
The output-referred mismatch error due to (D-DCAL,MAX) can also be deter-
mined by the difference of (6.23) and (6.24), as expressed by
Vout;error;B ¼ 2Vout;ideal ðD DCAL;MAX Þ ð6:25Þ
which is noted as point B in Fig. 6.18. Under the condition that the DAC output
range is linear to the induced mismatch and furthermore the DAC output range is
larger than the worst-case static mismatch of Gm3 and Gm4, the value of VCAL,SET
that minimizes the mismatch error (point C) then can be found by linear inter-
polation (Fig. 6.18). Since the calibration voltage is found within two DEM
periods, one decimated output is also chosen to be the average result of two DEM
periods, i.e. four conversions.
Since Gm3 and Gm4 are biased in weak inversion for maximum power effi-
ciency, and so their mismatch can be tuned by adjusting their tail currents with a
transconductor Gm6. To attenuate the noise contribution of the GEC path, the
transconductor Gm6 is implemented with resistor-degeneration stage to attenuate
6.3 Gain Accuracy Improvement Techniques in the Read-Out IC 155
the noise from the GEC path (Gm3/Gm6 = 480) and thus, the voltage across the
input of Gm6 are linear to the induced mismatch. Since the calibration voltage
VCAL,SET is determined by the ratio of gain errors at points A and B, the value of
the fixed DC input during calibration does not need to be known. After finding the
calibration voltage VCAL,SET in the start-up, the digital word of the DAC freezes
and the read-out IC enters the normal operation.
6.4 Offset and 1/f Noise Suppression Techniques

in the Read-Out IC
For sensor application, the bandwidth of interest is usually several Hz. At low
frequencies, offset and 1/f noise are the dominant error sources. To mitigate these
errors, nested-chopping techniques can be applied in various ways. Two different
possibilities will be described in this Section.
6.4.1 Previous Approach

(Multi-Stage Chopping and System-Level Chopping)
Figure 6.19 shows a simplified block diagram of the read-out IC consisting of a

three-stage CFIA and a SC DR ADC. In previous work, the input and intermediate
stages of the CFIA were both chopped, so as to suppress their own 1/f noise while
providing enough gain to suppress the input-referred 1/f noise of the un-chopped
output stage. However, the resulting up-modulated offset and 1/f noise would give
rise to a few hundred mV of chopper ripple at CFIA output [3].
In principle, the ADC’s clock scheme could be designed such that the ripple
would only be sampled at zero-crossings. However, this is not a robust solution
since the exact location of the zero-crossings will be a function of the CFIA’s
closed-loop frequency response, which, in turn, is a function of its user-selectable
closed-loop gain (Fig. 6.19). As a result, the ripple will not be sampled exactly at
zero-crossings, resulting in residual offset and noise. Moreover, any clock jitter
will also be translated into excess noise. To avoid such issues, chopper ripple must
be suppressed.
In [16], the input stages were chopped at 30 kHz, which was chosen to be
slightly above their 1/f corner frequency and equal to the sampling frequency of
DR ADC. The resulting ripple was removed by a continuous-time offset reduction
loop (ORL) (Fig. 6.20), which synchronously demodulated the CFIA’s output
chopper ripple and used the information to null the offset, and hence the ripple.
The intermediate stage is chopped at a higher frequency (600 kHz), to ensure that
its resulting chopper ripple is sufficiently suppressed by the LPF formed by the
CFIA’s compensation network. However, the associated charge injection and
C2
C1
SC
Vin
CH1 Vos1 CH3 CH4 Vos2 CH5
Vout ΔΣ ADC Dout
Gm3 Gm2 Gm2
R1 Cs1
fch1 fch1 fch2 fch2
Vfbk
CH2
Vfbk Gm4 90° shift Feedback
R2
fs f s DAC
fch1 fs
fch1
Vripple1 Gain = 220
Vripple2 Gain = 100
Fig. 6.19 Simplified block diagram of the read-out IC
clock spikes are observed to increase the CFIA’s residual offset from 1lV to 5lV.
Furthermore, interaction between the high frequency ripple and the shaped
quantization noise of DR modulator is observed to increase low-frequency noise.
In [16], this excess noise is suppressed by chopping the intermediate stage in a
bitstream-controlled (BSC) manner [19], so as to eliminate the correlation between
the chopper ripple and the quantization noise. Since the BSC chopping frequency
fBSC is never higher than fS/2 = 15 kHz, the residual offset is only reduced to 3lV.
The entire read-out IC is then chopped after every two conversions (Fig. 6.21) and
the low-frequency ripple is averaged in the digital decimation filter, resulting in a
worst-case offset of 200 nV.
6.4.2 Proposed Approach

(Input-Stage Chopping Combined with System-Level
Chopping)
In this approach, a nested-chopping scheme that simply combines input stage

chopping (in the CFIA) with system-level chopping (of the whole read-out IC) is
shown to achieve tens of nV-level offset and a sub-mHz 1/f noise corner. In this
scheme, the intermediate stage is not chopped, thus avoiding a potential source of
residual offset and ripple. As in [16], a ORL is applied to reduce the ripple
associated with the chopped input and feedback transconductors.
Applying DEM could cause some offset drift due to different input and feed-
back CM voltage. To average out this drift, the frequency of the system-level
chopping is chosen to be half of the DEM frequency (Fig. 6.21) and one decimated
6.4 Offset and 1/f Noise Suppression Techniques in the Read-Out IC 157
C2
C1
f1 f1 f2 f2
Vin Vos
Gm3 Gm2 Gm1
CH1
V out
CH3 CH4 CH 5
f1 R1
RRL
Vfbk f1 Vfbk
Gm4 C4
CH 2 Gm5 ∫ R2
CH6
Fig. 6.20 Simplied block diagram of the CFIA with ripple reduction loop
fLF
Vref
Vdd
+
+ SC Dout
CFIA Vout 1 + z-1
- Δ Σ ADC CHLF2
2
CHLF1 -
Reset
Read-out IC
fLF_chop
fDEM
Reset
1 conversion
1 decimated output
Fig. 6.21 Read-out IC with system-level chopping and the associated timing diagram
digital output is the average result of four ADC conversions. Instead of using a
‘‘0011’’ pattern for the system-level chopping in four conversions, a ‘‘0110’’
pattern is chosen in this work. Assuming that X is the DC input signal, Vn is the
read-out IC’s low-frequency errors, i.e. offset, drift, and 1/f noise, and Y is the
digital output of one system-level chopping period. When applying system-level
chopping with ‘‘0011’’ pattern, the output Y can be expressed as
ðX þ Vn Þ þ ðX þ Vn Þ z1 þ ðX Vn Þ z2 þ ðX Vn Þ z3

Y¼
4
Vn
¼ X þ ð1 þ z1 Þ2 ð1 z1 Þ ð6:26Þ
4
However, after applying ‘‘0110’’ pattern, the output Y is expressed as
“ 0011” Pattern
-10
“ 0110” Pattern
“ 0011” Pattern
-20
Magnitude (dB)
-30
“ 0110” Pattern
-40
-50
-60
-3 -2 -1
10 10 10
Normalized Frequency ( × π rad/sample)
Fig. 6.22 High-pass filtering effect of system-level chopping on the low-frequency noise with
‘‘0011’’ or ‘‘0110’’ chopping pattern
ðX þ Vn Þ þ ðX Vn Þ z1 þ ðX Vn Þ z2 þ ðX þ Vn Þ z3

Y¼
4
Vn 1 1 2
¼ X þ ð1 þ z Þð1 z Þ ð6:27Þ
4
Figure 6.22 shows the transfer functions of these two chopping patterns on the
1ow-frequency errors. It can be seen that the ‘‘0011’’ pattern exhibits a first-order
high-pass filtering (HPF), while the ‘‘0110’’ pattern exhibits a second-order HPF,
which imposes better rejection over low-frequency errors. More intuitively, it
can be seen that, unlike the ‘‘0011’’ pattern, the ‘‘0110’’ pattern can exactly
compensate for any offset drift that is a linear function of time.
Chopping just the input stage of the CFIA reduces its 1/f noise corner from
10 kHz to 0.3 Hz, Therefore, when the system-level chopping frequency is chosen
to be higher than 0.3 Hz, the CFIA’s 1/f noise will be effectively suppressed. For
one decimated output, it takes 5200 clock cycles (four conversions) with a sam-
pling frequency of 30 kHz, the system-level chopping frequency fsys equals
30 kHz/5200 = 5.8 Hz. Therefore, the low-frequency noise lower than this
frequency will be removed by system-level chopping.
Simulations in periodic steady-state (PSS) and periodic noise analysis
(PNOISE) tools of Spectre RF [11] are made to confirm the validity of above
analysis. Figure 6.23 shows the simulated input-referred noise spectrum of the
read-out IC with various 1/f noise suppression techniques. The input-stage chop-
ping combined with system-level chopping achieves a 1/f noise corner frequency
6.4 Offset and 1/f Noise Suppression Techniques in the Read-Out IC 159
Fig. 6.23 Simulated input Input referred noise

referred noise spectrum of (V/ Hz)
read-out IC with various
1/f noise suppression 10-5 No Chopping
techniques
10-6 Input stage Chop
10-7 Input + Intermedaite

stage Chop
10-8
Input stage +System-level
0.4mHz Chop
-9
10
0.1m 1m 10m 100m 1 10 100 1k 10k 100k
Frequency (Hz)
of 0.4 mHz, which is comparable to the result achieved by the multi-stage

chopping combined with system-level chopping (1/f noise corner less than
1 mHz).
6.5 Error Correction Techniques Summary
Figure 6.24 shows a bridge-readout system based on the proposed read-out IC with
associated timing diagram. Table 6.3 summaries the errors and the associated error
correction techniques applied in the proposed read-out IC. The gain error and gain
drift are reduced by using DEM and the digitally-assisted gain error correction
scheme. The offset and 1/f noise are suppressed by the combination of input stage
chopping in the CFIA and system-level chopping (The second pair of choppers can
be turned off, thus they are displayed grey in the Fig. 6.24). The chopper ripple of
the input stage is removed by the ripple reduction loop.
6.6.1 CFIA Implementation
The Input and Feedback Gm Stages with Enhanced CM Immunity

As discussed in Sect. 6.3.1, with the same CM, DEM reduces gain error of the
CFIA from D to D2/2. However, with different CM, the CM dependency between
Gm3 and Gm4 limits the gain accuracy even with DEM applied. To improve gain
accuracy, the transconductances of Gm3 and Gm4 should be kept constant over the
input CM range. Figure 6.25 shows the conventional input and feedback Gm stages
[3, 16]. Their CM dependency was mitigated in a power-efficient manner by
Read-out IC
Vref+
Resistive 90°
bridge fs
f DEM fch1
fLF_chop Bit-stream
fch2 control
fLF_chop
Gm3
+
21bit SC bs Decimation
CHLP1 CFIA Vout
ΔΣ ADC filter & digital
CHLP2 back-end
-
MUX
Gm4
ΔΣ
Gm6 Vcal
DAC
Vref- GE correction
90° shift Sampling
Vout
fch1
fs
0 0 1 0 1
bs 1 1 1
fch2
f LF_chop
Δ+Δ cal,max -(Δ+Δ cal,max) Δ-Δ cal,max -(Δ-Δ cal,max)
f DEM
Vcal,max -Vcal,max
1 conversion
1 decimated output
Fig. 6.24 Block and timing diagrams of a bridge-readout system based on the proposed read-out IC
Table 6.3 Errors and the associated error-correction techniques employed in the read-out IC
Error Error correction techniques
Gain error and gain drift Dynamic element matching and digital-backend gain error
correction (Sect. 6.3)
Offset and 1/f noise in CFIA Input stage chopping ? system-level chopping or multi-stage
chopping ? system-level chopping (Sect. 6.4)
Offset and 1/f noise in ADC Auto-zeroing in the first integrator (Sect. 6.2.2)
Chopper ripple from the Offset reduction loop and zero-crossing sampling (Sect. 6.4)
input stage
cascoding the input transistors with low-threshold devices M3–M4. Transistors

M1–M4 operate in weak inversion at a large Gm/Ib ratio. The resulting drain-source
voltage VDS of M1 equals the threshold difference of M1 and M3, which is about
0.18 V. This small VDS limits the output impedance of the input differential pair.
Two opposing effects influence the Gm of the input (or feedback) stage in
Fig. 6.25. When the CM level increases, the VDS of M1 also increases due to channel
length modulation, leading to an increased Gm, as shown in curve-2 in Figs. 6.26a
and b. On the other hand, the increased CM level reduces the head-room of the tail
Fig. 6.25 Conventional VDDA

input (or feedback) Gm stage
implemented with PMOS M7
differential pair and low-
M5
threshold cascoding [3]
IGE+
M8 M6
M1 M2
Vin+
* M3 M4 *
Vin-
Cascode
current source M5, hence reducing a tail current and the Gm of the input differential
pair, as depicted by curve 1 in Fig. 6.26a and b. These two opposing effects might
provide an improved CM immunity. However, due to the limited VDS of M1, the
channel-length modulation effect is much larger than the reduced bias-current effect.
The Gm of the input differential pair, therefore, exhibits a limited CM immunity.
Increasing the output impedance of the input pair will reduce the channel length
modulation effect and thus flatten the slope.
Figure 6.29 shows the simulated Gm variation of the input or (feedback) stage
for an input CM range of -0.1 V to 2.8 V (typical process corner). If the current
source in Fig. 6.25 is ideal, the Gm variation is 0.6 % over the CM range, as
depicted by curve 1 in Fig. 6.29. If the cascode current source shown in Fig. 6.25
is used, the two opposing effects partially compensate each other. This is because
the input pair and the current source are cascoded differently (note that the gate
terminals of the input and cascode transistors are connected together in Fig. 6.25).
The CM sensitivities of these two circuits thus differ, resulting in a residual Gm
variation of 0.07 %, as shown in curve 2 in Fig. 6.29. It is observed that Gm
increases with input CM voltage, implying that the CM dependency is limited by
the finite output impedance of the input pair.
To boost the output impedance of the input pair, the threshold difference
between M1 and M3 need to be increased. This can be done in two manners:
reducing the threshold of M3 or increasing the threshold of M1. To reduce the
threshold of M3, a resistor R1 can be added between the source of M1 and the bulk
of M3, as shown in Fig. 6.27a. The voltage drop across R1 introduces a VSB of
0.3 V, so as to reduce M3’s threshold by 0.1 V. However, this approach needs an
extra bias current of I1, leading to an increase in power consumption. A better
alternative is to increase the threshold of M1. As shown in Figure. 6.27b, resistor
R3 is added between the common sources of M1 and M2, and the common bulk of
M1 and M2. The voltage drop across R3 introduces a VBS of 0.29 V to M1,
269
0.5
Gm (μ A /V)
0.4 1) Effective Gm
1) VDS of M5
affected by CS
V DS (V)
2) V DS of M1 268
2) G m of input-pair
0.3 (with ideal CS)
(with ideal CS)
0.2
267
0.1
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Input CM Voltage (V) Input CM Voltage (V)
(a) current source (CS) (b)
Fig. 6.26 a Simulated VDS versus input CM (Fig. 6.25). b Simulated Gm versus input CM
(Fig. 6.25)
VDDA VDDA
-
Vin+
+
M1 M2
R3 VBS
+ R1 -
VSB Vin+
-
Vin- * M3 M4 M1 M2
*
* *
Cascode
Vin- M3 M4
VB1 I1 Cascode
(a) (b)
Fig. 6.27 Input (or feedback) Gm stage. a The threshold of M3 is reduced by body effect, b the
threshold of M1 is increased by body effect
increasing the threshold of M1 from about 0.9 to 1 V due to body effect. As a

result, the VDS of M1 increases, boosting the output impedance of the input pair
from 200 MX to 500 MX. The advantage of this solution is that input and cascode
transistors all share the same bias current, resulting in better power efficiency.
To improve the CM immunity further, the same cascode configuration is
applied in both the input pair and the tail current source (Fig. 6.28), so as to
balance their impedance. This ensures that the two opposing effects have nomi-
nally the same absolute CM sensitivity and so will completely cancel each other.
The combination of the threshold boosting and impedance balancing techniques
reduces the Gm variation from 0.07 to 0.02 % (typical corner) over the input CM
range, as shown in curve 3 of Fig. 6.29. Simulations over all process corners show
Fig. 6.28 Proposed input (or VDDA

feedback) Gm stages with
enhanced CM immunity R1 R2
M7 M5
M8 M6 IGE+
* *
+
R3 V BS
-
Vin+
M1 M2
* *
Vin- M3 M4
Cascode
that this reduction is robust to process variation: the maximum Gm variation is

reduced from 0.089 % (with Fig. 6.25) to 0.033 % (with Fig. 6.28). The only
disadvantage of this approach is that the voltage drops across R2 and R3 reduce the
input CM range from 0–3.7 V to 0–2.8 V.
Compensation Gm6 Stage
Transconductance Gm6 is added to compensate the mismatch of Gm3 and Gm4
by fine-tuning their tail currents. The noise contribution from the gain error cor-
rection path is negligible because it is attenuated by the ratio of Gm3/Gm6 = 480
and suppressed by the finite CMRR of the input stages. Gm6 is implemented as a
resistor-degenerated stage. Its bias current is chosen just large enough to cover the
maximum expected mismatch (2 %) between Gm3 and Gm4. the implementation of
the intermediate stage Gm2, the output stage Gm1, the cascode buffer 2 (CB2) and
the cascode buffer 1 (CB1) (Fig. 4.15) is similar to that with the ORL [3].
6.6.2 ADC Implementation
This section discusses the implementation of the ADC in the ROIC. The ADC is
implemented as a second-order SC DR modulator followed by a decimation filter.
Figure 6.30 shows the block diagram of the modulator with its associated timing
diagram. It is implemented using two fully differential switched-capacitor inte-
grators and a clocked comparator. The modulator uses non-overlapping clocks to
control the timing for switches. Clocks with delayed falling edges (P1d and P1d) are
used to prevent signal-dependent charge injection [5].
Fig. 6.29 Simulated Gm

variations versus input CM a
with ideal current source,
b using Fig. 6.25, c using
Fig. 6.28
cycle
P1
P2d Cf0
P1d P1d
VCM
Vref+ V ref- P2
Bs Bs
Cf1
P2d
P1
Eval Eval
V CM VCM
Cint2 VCM VCM
P2d
P 1d
P2
P1d
Cint1
P2
P1d Cs1 P 2d Cs2 P1 P2 P1

Vin+ P2d Cf2 Bs
+ - + - +
A1 A2
- + - +
Vin- P 1d Cs1 P 2d Cs2 P1 P2d Cf2 P1 Bs
P1d
P2
P1d
Cint1
P2
P 2d
P2
V CM
Cint2 VCM V CM
V CM
Eval
Bs Bs P1
Cf1
Vref+ V ref-
VCM
P1d
P2d Cf0
Fig. 6.30 Block diagram of the second order DR modulator
Topology and Clock Timing

In one DR cycle, when the first integrator is in the sampling and auto-zeroing
phase, as shown Fig. 6.31 (clock P1 and P1d is high), the input signal and the offset
of opamp A1 are sampled on the sampling capacitors CS1. When the clock P2 and
P2d are high, the charge of the feedback voltage controlled by the bit-stream is
subtracted from the charge of the input signal stored on CS1, as shown in Fig. 6.32.
The subtraction result is scaled by the ratio of CS1/Cint1 = 3/10, appearing at the
integrator output.
When the 1st integrator is auto-zeroed, the signal is not available during this
period. Thus the 2nd integrator can only sample the 1st integrator output to the
sampling capacitor CS2 during the integration phase when P2 and P2d are high, as
shown in Fig. 6.32. In the next half clock period (when P1 and P1d are high), the
charge stored on CS2 is transferred to the integration capacitor Cint2.
cycle
P1
P2d Cf0
P1d
P1d
VCM
Vref+ V ref- P2
Bs Bs
Cf1
P2d
P1
Eval Eval
V CM VCM
Cint2 VCM VCM
P2d
P 1d
P2
P1d
Cint1
P2
P1d Cs1 P 2d Cs2 P1 P2
Vin+ P2d Cf2 P1 Bs
+ - + - +
A1 A2
- + - +
P1d
P2
P1d
Cint1
P2
P 2d
P2
V CM
Cint2 VCM V CM
V CM
Eval
Bs Bs P1
Cf1
Vref+ V ref-
VCM
P1d
P2d Cf0
Fig. 6.31 The second-order DR modulator when P1 and P1d are ‘‘high’’
cycle
P1
P2d Cf0
P1d
P1d
VCM
Vref+ V ref- P2
Bs Bs
Cf1
P2d
P1
Eval Eval
V CM VCM
Cint2 VCM VCM
P2d
P 1d
P2
P1d
Cint1
P2
P1d Cs1 P 2d Cs2 P1 P2

Vin+ P2d Cf2 P1 Bs
+ - + - +
A1 A2
- + - +
P1d
P2
P1d
Cint1
P2
P 2d
P2
V CM
Cint2 VCM V CM
V CM
Eval
Bs Bs P1
Cf1
Vref+ V ref-
VCM
P1d
P2d Cf0
Fig. 6.32 The second-order DR modulator when P2 and P2d are ‘‘high’’
In the feed-forward topology, two feed-forward paths from the input and from
the 1st integrator output are introduced. Therefore, an analog adder is required to
sum up the signal at the quantizer input. In this design, this adder is implemented
in a passive structure in two phases, as shown in Fig. 6.33. It is controlled by the
non-overlapping clocks P1, P1d and P2, P2d. The feed-forward coefficients are set
by the capacitor ratios. The quantizer decides near the end of the clock phase P1
when the passive adder has fully settled (Fig. 6.30).
The feed-forward coefficients are determined by the ratio between the feed-
forward capacitors and the total capacitor summation. When the adder is in the
sampling phase (clock P2 and P2d are high), it samples the input signal (Vin +), the
first integrator output (Vint1+) and the second integrator output (Vint2+). Therefore
the amount of charge stored on each capacitor is given by:
Cf0 Cf0
Vin+ VCM
VCM
Vint1+ Cf1 VCM Cf1
Vout+
Cf2 Cf2
Vint2+ VCM
(a) (b)
Fig. 6.33 Passive adder on the positive side a sampling phase, b summing phase
Qff 0 ¼ Cf 0 ðVinþ VCM Þ ð6:28Þ
Qff 1 ¼ Cf 1 ðVint1þ VCM Þ ð6:29Þ
Qff 2 ¼ Cf 2 ðVint2þ VCM Þ ð6:30Þ

When clocks P1 and P1d are high, the adder is in the summing phase with all the
sampling capacitors connected in parallel. The charge redistributes itself among
them and the output voltage of the adder is derived as
Cf 0 ðVinþ VCM Þ þ Cf 1 ðVint1þ VCM Þ þ Cf 2 ðVint2þ VCM Þ
Voutþ VCM ¼
Cf 0 þ Cf 1 þ Cf 2
Cf 0 Cf 1 Cf 2
¼ Vinþ þ Vint1þ þ Vint2þ VCM
Cf 0 þ Cf 1 þ Cf 2 Cf 0 þ Cf 1 þ Cf 2 Cf 0 þ Cf 1 þ Cf 2
ð6:31Þ
Equation (6.28) shows that all the feed-forward coefficients are smaller than
one, the values of which are listed in Table 6.2 (a1, a2, a3).
Implementation Details
The settling behavior and DC gain of the opamp in the first integrator are
important for the overall performance of the modulator. The circuit was therefore
designed for complete settling (i.e. the equivalent settling error is less than
130 dB). To achieve a 160 dB DC gain in A1 for 1/f noise suppression, it is
implemented with a two-stage Miller-compensated amplifier consisting of a gain-
boosted telescopic topology in the first stage and a class-A structure in the second
stage (Fig. 6.34). Overall, this opamp has a UGB of 250 kHz with 5pF load
capacitor and draws only 38 lA.
Since errors introduced by the second integrator are attenuated by the gain of the
first integrator, no offset cancellation or gain boosting are needed here. Thus the
opamp A2 in the second integrator is implemented as a fully-differential two-stage
amplifier consisting of a telescopic first stage and a class-A output stage, as shown in
Fig. 6.35. This amplifier achieves a DC gain large than 100 dB and a UGB of
60 kHz with 6pF load capacitor while consuming only 8 lA.
Fig. 6.34 The schematic VDDA

diagram of the opamp A1 in Vb1 Vb3
the 1st integrator of the DR
modulator shown in Fig. 6.30
Vb2
Vin+ Vout+
Vin-
Vout-
+ -
11p
- +
11p
- +
+ -
Vb4
GNDA
Fig. 6.35 Schematic VDDA

diagram of the opamp A2 in
V b1 Vb3
the 2nd integrator of the DR
modulator shown in Fig. 6.30
V b2
V in+ Vout+
V in-
11p Vout-
Vb4
11p
Vb5
Vb6
GNDA
Fig. 6.36 Chip micrograph

of the ROIC
The read-out IC was fabricated in a standard 0.7 lm CMOS process with low-
threshold transistors, linear capacitors and high-resistivity poly-resistors. The chip
has an active area of 6 mm2, as shown in Fig. 6.36. All the measurement results
are based on the measurements of ten samples.
Noise Spectrum of the ROIC
Figure 6.37a shows the measured output PSD of the read-out IC, which is set
with multi-stage chopping while system-level chopping and DEM are ‘‘off’’. The
noise level is flat from 1 mHz with a noise density of 16.2nV/HHz, corresponding
to a resolution of 20-bits with respect to a full-scale range of ±40 mV in 5 Hz
bandwidth. To eliminate the low-frequency lobe that would occur due to the
interaction between the read-out IC’s residual offset and the Hann window used,
the read-out IC’s offset was subtracted before the FFT was computed. The ADC’s
sampling frequency fS and input stage chopping frequency fch1 is lowered to
10 kHz to display higher frequency resolution (10 kHz/224 = 0.596 mHz). As
seen from the zero-input FFT plot (b) in Fig. 6.37a, the noise level of the read-out
IC is mainly limited by the thermal noise.
Figure 6.37b shows the output noise spectrum of the read-out IC with ORL
‘‘on’’ and ‘‘off’’. It is clearly seen that with ‘‘ORL’’ off, some aliased noise appears
at the signal band below 10 Hz, since the chopper ripple at 10 kHz is being
sampled by a sampling frequency fs of 10 kHz. Furthermore, the uncertainty of the
sampling moment of the chopper ripple due to clock jitter could also induce low
frequency noise. Figure 6.37b shows that the ORL is necessary to avoid any
aliasing error near DC, so as to achieve a flat noise spectrum down to 1 mHz.
To test the effectiveness of system-level chopping on suppressing the 1/f noise,
the choppers in the intermediate stage are turned off, while DEM and system-level
Fig. 6.37 a Measured output (a) 0

spectrum of the read-out IC -20
with DEM and system-level -40 Frequency resolution= 0.596mHz
a fs = 10kHz
chopping off (ORL ‘‘on’’).
PSD (dBFS)
-60
40mV input
b Measured output spectrum -80
of the read-out IC with DEM -100
and system-level chopping -120
off (ORL ‘‘off’’ and ‘‘on’’) zero input
-140
-160
b
1mHz
-180
-200
10-3 10-2 10-1 1 10 102 103 104
Frequency (Hz)
(b) 0
-20
-40
-60
PSD (dBFS)
-80
-100
-120
-140
-160
-180
-200
10-3 10-2 10-1 1 10 102 103 104
Frequency (Hz)
chopping are both on. Figure 6.38 shows the measured noise spectrum with dec-
imated output. It is evident that the 1/f noise corner is suppressed below 0.1 mHz,
showing that input stage chopping combined with system-level chopping can be
used as a better alternative to suppress the 1/f noise.
Noise Spectrum of the ADC
The measured output spectrum of the ADC alone also achieves a flat noise
spectrum from 1 mHz to 10 Hz, as seen from the curve b in Fig. 6.39. This
corresponds to a resolution of 21-bit with a full-scale range of ±4 V (reference to
5 V). This result confirms that the DC gain in the auto-zeroing phase of the first
integrator is sufficient for 1/f noise suppression. The FOM of the ADC (evaluated
by FOM = Power/(2ENOB 9 2fB)) is 10 pJ/Conv.
Gain Error
As discussed before, the gain accuracy of the ROIC is mainly determined by the
mismatch between the input and feedback Gm stages in the current-feedback
instrumentation amplifier. Applying DEM to them reduces the gain error from
0.6 % to ±0.0035 % if Gm3 and Gm4 are at the same CM voltage (2.5 V).
Including the GEC path further reduces the gain error to ±0.00165 %, as shown in
Fig. 6.40. However, due to CM dependency between the Gm3 and Gm4, the gain
error rises to 0.12 % in the worst-case when one of the Gm’s is at 0 V [16]. Using
-140
Frequency resolution = 0.093mHz
PSD (dBFS) -150
-160
-170
-180 0.1mHz
-190
-200
-210
10-4 10-3 10-2 10-1 1
Frequency (Hz)
Fig. 6.38 Decimated output spectrum of the read-out IC with Input stage chopping and system-
level chopping ‘‘on’’ (Choppers in the intermediate stage are ‘‘off’’)
Fig. 6.39 Measured output 0

spectrum of the ADC alone -20
without CFIA -40
4V input
-60
PSD (dBFS)
-80
-100
-120
-140 zero input
-160
-180
-200
-3 -2 -1 2 3 4
10 10 10 1 10 10 10 10
Frequency (Hz)
the bulk-biasing and impedance-balancing techniques (shown in Fig. 6.28), the

typical CMRR of the ROIC is increased from 130 to 140 dB, close to that obtained
with resistor-degenerated stages [20]. Figure 6.41 shows the measured CMRR
histogram of the ROIC compared to [16]. Figure 6.42 shows that the ROIC
achieves a gain error of 0.037 % over a CM range of 0–2.5 V, which is a 3.2X
improvement compared to [16].
Offset
Figure 6.43 shows the measured offset histograms with different chopping
strategies. By employing multi-stage chopping in the CFIA, the ROIC achieves a
maximum offset of 4.6 lV (Fig. 6.43a). Turning on the system-level chopping
reduces offset to less than 200 nV (Fig. 6.43b) [16]. Tuning off the choppers in the
intermediate stage, using the combination of the input stage chopping and system-
level chopping achieves a lower offset of 48 nV (Fig. 6.43c), which confirms the
analysis discussed in Sect. 6.4.2. The measured worst case offset drift is less than
6 nV/°C.
6
No DEM
5
4
3
2
1
Gain Error
0 10 20 30 40 50 60
( 100ppm)
Number of Samples
5 DEM only
4
3
2
1
Gain Error
0 10 20 30 40 50 60 70 80 ( 1ppm)
DEM + GEC
5
4
3
2
1
Gain Error
0 10 20 30 40 50 60 70 80 ( 1ppm)
Fig. 6.40 Measured gain error histograms with the same input and feedback CM
Fig. 6.41 Measured CMRR

5 Measured CMRR CMRR=130dB
and comparison to [16]
[6.16]
showing the effect of bulk 4
biasing the impedance 3
balancing in improving CM
2
immunity
Number of Samples
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4

CMRR (μV/V)
Measured CMRR of
5
this work CMRR=140dB
4
3
2
1
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4

CMRR (μV/V)
INL
The current-feedback instrumentation amplifier is the main source of INL, since
the measured INL of the ADC is around ±1 ppm, as shown in Fig. 6.44. The use
Fig. 6.42 Measured gain

error histograms and 5 Measured gain error [6.16]
comparison to [16] showing 4
3
the effect of bulk biasing the
Number of Samples
2
impedance balancing
1 Gain Error
0 2 4 6 8 10 12 ( 100ppm)
Measured gain error of

5 this work
4
3
2
1 Gain Error
0 2 4 6 8 10 12 ( 100ppm)
(a) 2nd chopper on & LF-chop off

6
5
4
3
2
1
0
-5 -4 -3 -2 -1 0 1 2 3 4 5 (μV)
(b) 2nd chopper on & LF-chop on

6
Number of samples
5
4
3
2
1
0
-250 -200 -150 -100 -50 0 50 100 150 200 250 (nV)
4 (c) 2nd chopper off & LF-chop on

3
2
1
0
-50 -40 -30 -20 -10 0 10 20 30 40 50 (nV)
Fig. 6.43 Measured offset histograms
of DEM improves the INL of ROIC from 35 to 5 ppm. When the gain error
correction path is turned on, the INL is improved slightly, as shown in Fig 6.45.
At low gain settings, saturation of the amplifier’s input stages eventually limits its
linear input range (INL \ 10 ppm) to ±120 mV (with DEM and GEC on), as
depicted in Fig. 6.46.
Fig. 6.44 Measured INL of 5

the ADC without CFIA 4
INL of AD C (ppm of full scale)

3
-1
-2
-3
-4
-5
-2 -1 0 1 2
Input Range (V)
Fig. 6.45 Measured INL of

the ROIC. a No DEM,
b DEM only, c DEM ? GEC
Fig. 6.46 Measured INL of 15

IN L of RO IC (ppm of full scale)
the ROIC with an input range

of ±120 mV 10
-5
-10
-15
-150mV -100mV -50mV 0 50mV 100mV 150mV
Input Range
Fig. 6.47 Measured gain 4

drift histograms of the ROIC. 3 1 ) No DEM
a No DEM, b DEM only, 2
c DEM ? GEC
1
0
0 1 2 3 4 5 6
(ppm/°C)
4
Number of samples
2 ) DEM only
3
2
1
0
0 1 2 3 4 5 6
6 (ppm/°C)
5
4 3 ) DEM + GEC
3
2
1
0
1 2 3 4 5 6
Gain drift (ppm/°C)
Gain Drift
For thermistor and thermalcouple read-out, gain drift of the interface electronics
is essential. The use of DEM improves the gain drift from 6.1 ppm/°C to 4.3 ppm/°C.
Further applying GEC improves the gain drift to 1.2 ppm/°C, as shown in Fig. 6.47.
Thermistor Measurements with the ROIC
To test and compare the performance of the ROIC with other precision instru-
ment, three measurement set-ups (Fig. 6.48) were used to simultaneously measure
the temperature drift in a large (96 cm3) oven-stabilized aluminum block. The first
one used the ROIC to read out a thermistor bridge, the second one used a Keithley
2002 7-1/2 digit multimeter to read-out the same thermistor bridge, the third one
used the same Keithley 2002 multimeter to read-out a Pt-100 reference sensor.
The three measurement systems were set-up for a conversion time of 0.25 s,
which for the ROIC meant that 5000 samples (@ fs = 20 kHz) were decimated by a
sinc3 filter. The measurement results are shown in Fig. 6.49. Due to its lower sen-
sitivity, the resolution of the Pt-100 is much lower than that of the thermistor bridge.
The 0.7 lK (rms) temperature-sensing resolution achieved by the thermistors and
the ROIC is roughly 2X better than that achieved by the thermistors and the Keithley,
despite the fact that the ROIC only draws 270 lA and is much more compact.
The performance of the read-out IC is summarized in Table 6.4 and compared
with other state-of-the-art. The read-out IC achieves a typical gain drift of
0.7 ppm/°C and a 5 ppm INL at a gain of 100. Furthermore, the combination of the
input stage chopping and system-level chopping enable the read-out IC to achieve
Vref Vref
R1 Rx R1 Rx
Keithley Out
ROIC bs
2002
Thermistor Metal foil Rx
R1
Rx resistor
R1
(a) Thermistor + ROIC (b) Thermistor + Keithley
Keithley Out
Pt _100
2002
(c) Pt_100 + Keithley
Fig. 6.48 Three comparison measurement set-ups with thermistors and Pt_100. a Thermis-
tor ? ROIC, b Thermistor ? Keithley, c Pt_100 +Keithley
Fig. 6.49 Temperature drift 20mK

(a) Thermistor + ROIC
as measured by a a (TR = 0.7µK)
thermistor bridge ? ROIC, 15mK
25.12 °C - Temperature (°C)
b thermistor (b) Thermistor + Keithley

bridge ? Keithley, c Pt- 10mK (TR = 1.4µK)
100 ? Keithley (for clarity,
5mK
the curves are offset by
5 mK)
0K
-5mK
(c) Pt_100 + Keithley
(TR = 1mK)
-10mK
-15mK
-20mK
0 200 400 600 800 1000 1200 1400 1600 1800
Time (s)
* Temperature resolution (TR)
0.1 mHz 1/f noise corner, a maximum offset of 48nV and an offset drift of 6nV/°C.
Compared to other work, our work achieves the best gain error of 0.037 %, 20X
better offset and 1.5X better gain drift than [21]. Moreover, it only consumes a
270 lA supply current from a 5 V supply (CFIA 220 lA, ADC 50 lA).
Table 6.4 Performance of the ROIC versus the state-of-the-art

This work AD7193 CS5530 ADS1282 Thomsen
[21] [22] [23] [24]
Year 2011 2011 2009 2007 2000
Supply current 270 lA 4.3 mA 7 mA 4.5 mA 8.2 mAa
Supply voltage 5V 5V 5V 5V 5V
1/f noise corner 0.1 mHz – 25 mHz – 10 mHz
Input range ±40 mV ±39 mV ±78 mV ±263 mV ±28 mV
CMRR 140 dBb 110 dBb 120 dBb 110 dBb –
Input referred noise 16.2 nV/HHz 5nV/HHz 12 nV/HHz 5 nV/HHz 6.2 nV/HHz
density
Gain drift DEM ? GEC: 1 ppm/°C 2 ppm/°C 9 ppm/°C 15 ppm/°C
0.7 ppm/°Cb (typ) (typ) (typ) (typ)
DEM:
3 ppm/°Cb
Gain error 0.037 %a 0.39 %b 1% 1% –
(uncalibrated)
Offset drift 6 nV/°Ca 5nV/°Cbb 10nV/°Cb 20 nV/°Cb 70 nV/°Ca
Offset (uncalibrated) 48 nVa 1 lVa 9.5 lVa 200 lVa –
Nonlinearity 5 ppmb – 30 ppmb 4 ppmb –
Conversion time 0.173 s 0.2 s 0.13 s 0.004 s 0.0083 s
FOM(pJ/Conv) [25] 111 135 9000 172 160
a
Worst case
b
Typical case
6.8 Conclusions
This chapter described the system-level design and implementation of the read-out
IC that consists of a current-feedback instrumentation amplifier (CFIA) and a
switched-capacitor DR ADC. The CFIA provides high input impedance for bridge
read-out and relaxes the noise and offset requirements of the ADC.
To achieve 1 mHz 1/f noise corner, both the input and intermediate stages of
the CFIA are chopped, so as to provide sufficient gain to suppress the 1/f noise
corner of the unchopped output stage down to mHz range. Chopping gives rise to a
chopper ripple at the amplifier output and this ripple will be sampled by the
sampling front-end of the succeeding ADC. To avoid noise aliasing, a continuous-
time offset reduction loop is employed to suppress the ripple from the input stage,
while the ripple from the intermediate stage is chopped in a bit-stream controlled
manner, so as to reduce the correlation between the chopper ripple and quanti-
zation noise.
To reduce offset further to the nV-level, system-level chopping is employed to
chop the entire read-out chain during multiple conversions. The modulated offset
is then averaged out in the decimation filter. It has been found that to suppress 1/
f noise and offset, input stage chopping combined with system-level chopping can
be used as a better alternative to the multi-stage chopping with system-level
6.8 Conclusions 177
chopping. Measurement results show that the former achieves a 0.1 mHz 1/f noise
corner and 48 nV worst-case offset.
The ADC employs a topology whose gain accuracy does not depend on com-
ponent matching. Furthermore, with ratio-metric topology, the gain accuracy of
the ADC’s reference is much relaxed. These two solutions enable the ADC to
achieve a gain error of less than 2 ppm. Thus, the gain error of the read-out IC is
mainly determined by the mismatch between the input and feedback
transconductors.
To eliminate the transconductor mismatch, dynamic element matching (DEM)
is applied to the input and feedback transconductors to average out the mismatch.
However, their CM dependency limits the achievable gain accuracy with DEM
applied. To enhance their CM immunity, bulk-biasing and impedance-balancing
techniques are applied. To reduce gain error and gain drift further, a digitally-
assisted gain error correction scheme is proposed, which explores the power of
digital signal processing succeeding the ADC to improve the gain accuracy and
gain drift of the CFIA. This gain error correction scheme uses a linear interpolation
algorithm to find the correct calibration voltage during start-up. Due to its ratio-
metric property, the input calibration signal does not need to be known. Overall,
GEC path serves as a coarse-trimming to the Gm mismatch, while the DEM acts as
a fine-tuning to compensate for temperature drift.
Measurement results show that the offset and drift of the read-out IC exceeds
the-state-of-art. These qualities make the proposed read-out IC very suitable for
demanding bridge transducer applications, which require low thermal and
1/f noise, high accuracy, low drift, and simultaneously, low power consumption.
References
1. Bakker A, Huijsing JH (2000) High-accuracy CMOS smart temperature sensors. Kluwer

Academic Publishers, Boston 2000
2. Huijsing JH, Riedijk FR, Van der Horn G (1994) Developments in integrated smart sensors.
Sens Actuators 43(1–3):276–288
3. Wu R, Makinwa KAA, Huisjing JH (2009) A Chopper Current-Feedback Instrumentation
Amplifier with a 1 mHz 1/f noise corner and an AC-coupled ripple reduction loop. IEEE J
4. Thomsen A et al (2000) A DC measurement IC with 130 nVpp noise in 10 Hz. In: IEEE
international solid-state circuits conference digest of technical papers, pp 334–335
5. Norsworthy SR, Schreier R, Temes GC (eds) (1997) Delta-sigma data converters: theory,
design and simulation. In: IEEE Press, Piscataway
6. Quiquempoix V et al (2006) A low-power 22-bit Incremental ADC. IEEE J Solid-State
Circuits 41(7):1562–1571
7. Pertijs M (2005) Precision temperature sensors in CMOS technology. Ph.D. thesis, Delft
8. Silva J, Moon U, Steensgaard J, Temes G (2001) Wideband low-distortion delta-sigma ADC
topology. Electron Lett 37(12):737–738
9. Feely O, Chua LO (1991) The effect of integrator leak in RD modulation. IEEE Trans
Circuits Syst 38(11):1293–1305
10. Enz CC, Temes GC (1996) Circuit techniques for reducing the effects of Op-Amp
imperfections: autozeroing, correlated double sampling, and chopper stabilization. Proc IEEE
84(11):1584–1614
11. Kundert K (2005) Simulating switched-capacitor filters with Spectre RF, The Designer’s
Guide Community. Available at: http://www.designers-guide.org/Analysis/sc-filter.pdf
12. Nam K (2005) Design of low-voltage low-power sigma-delta modulators for broadband high-
resolution A/D conversion. Ph.D. Thesis, Stanford University, CA
13. Schreier R, Temes GC (2005) Understanding delta-sigma data converters. IEEE Press and
Wiley-Interscience, NJ
14. Schreier R, Silva J, Steensgaard J, Temes GC (2005) Design-oriented estimation of thermal
noise in switched-capacitor circuits. IEEE Trans Circuits Syst I Regul Pap 52(11):2358–2368
15. Slattery C, Nie M (2005) A reference design for high-performance, low-cost weigh scales, Dec.
Available at: http://www.analog.com/library/analogDialogue/archives/39-12/weigh_scale.html
16. Wu R, Huijsing JH, Makinwa KAA (2011) A 21-bit ±40 mV range read-out IC for bridge
transducers. In: IEEE international solid-state circuits conference digest of technical papers,
pp 110–111, Feb 2011
17. Blanken PG, Menten SEJ (2002) A 10 lV-offset 8 kHz bandwidth 4th-order chopped RD
A/D converter for battery management. In: IEEE international solid-state circuits conference
digest of technical papers, pp 388–389, Feb 2002
with a gain error reduction loop and 0.06% untrimmed gain error. IEEE J. Solid-State
Circuits 46(12):2794–2806
19. Pertijs MAP, Makinwa KAA, Huijsing JH (2004) A sigma-delta modulator with bitstream-
controlled dynamic element matching. In: ESSCIRC Digest of Technical Papers, pp 187–190,
Sept 2004
20. Pertijs MAP, Kindt WJ (2010) A 2010. 140 dB-CMRR current-feedback instrumentation
amplifier employing ping-pong auto-zeroing and chopping. IEEE J Solid-State Circuits
45(10):2044–2056
21. AD7193 datasheet. http://www.analog.com/en/analog-to-digital-converters/ad-converters/
ad7193/products/product.html
22. CS5530 datasheet. http://www.cirrus.com/en/products/pro/detail/P1108.html
23. ADS1282 datasheet. http://focus.ti.com/docs/prod/folders/print/ads1282.html
24. Thomsen A et al (2000) A DC measurement IC with 130nVpp noise in 10 Hz. In: IEEE
international solid-state circuits conference digest of technical papers, pp 334–335, Feb 2000
25. Murmann B, Boser B (2004) Digitally assisted pipeline ADCs: theory and implementation.
Kluwer Academic Publishers, Boston 2004
Chapter 7
Conclusions
This final chapter summarizes the work described in Chaps. 1–6 and provides an
overview of the original contributions and the most important findings presented in
this thesis. It also shows how some of the techniques developed for bridge sensor
readout can also be useful in other applications, and provides an outlook on future
work.
7.1 Original Contributions
The following list provides an overview of the most important original contribu-
tions presented in this thesis. References to the thesis and the appropriate publi-
cations have also been included.
7.2 Chapter 4
• A multi-stage chopping strategy is proposed for a current-feedback instrumen-

tation amplifier (CFIA), in which the input and intermediate stages are chopped
and their gain is made high enough to suppress the 1/f noise of the output stage
down to 1 mHz [1] (Sect. 4.2).
• To suppress the chopper ripple in a CFIA, a continuous-time offset reduction
loop (ORL) is proposed [2, 3]. Due to its continuous-time nature, the ORL does
not cause noise folding, thus reducing the noise power around the chopping
frequency (Sect. 4.3).
• Compared to ripple reduction techniques using continuous-time [4] or discrete-
time [5] notch filters or an auto-correction feedback loop [6], the key advantage
of the ORL is that it is inherently stable. Thus, a low chopping frequency can be

180 7 Conclusions
chosen to achieve low offset without the stability problems caused by excessive
phase shift in the notch-filter (Sects. 2.4.5 and 4.3.2).
• Detailed analysis of the loop gain transfer function of the ORL in a CFIA [3]
(Sect. 4.3.2).
7.3 Chapter 5
• The gain accuracy of a CFIA is mainly determined by the mismatch between its
input and feedback transconductances. To improve this, dynamic element
matching (DEM) can be applied. To suppress the resulting signal-dependent
DEM ripple, a continuous-time gain error reduction loop (GERL) is proposed.
The GERL continuously nulls the mismatch of the input and feedback trans-
conductances, thus eliminating the need for trimming [7, 8] (Sect. 5.3).
• Detailed analysis of the loop gain transfer function of the GERL in a CFIA [8]
(Sect. 5.3.3).
• Since the loop gain of the analog GERL is signal dependent and is zero for zero
input. In this state, leakage causes the integrator output to drift away, thus it
needs to re-settle when a finite input signal reappears. To avoid the need of re-
settling, a digitally-assisted GERL that stores the gain mismatch information in
the digital domain is proposed [8] (Sects. 5.4 and 5.7.2).
7.4 Chapter 6
• To suppress the 1/f noise in a read-out IC that consists of a CFIA and an ADC,
the combination of input stage chopping (in the CFIA) and system-level
chopping (that chops the complete read-out chain) is shown to be a better
alternative to the combination of multi-stage chopping and system-level chop-
ping. Lower offset and comparable 1/f noise corner frequency are achieved in
the former approach (Sect. 6.4).
• To improve the gain accuracy, gain drift and linearity of the read-out IC, DEM
is applied to the input and feedback transconductors of the CFIA. Furthermore, a
gain error correction scheme is proposed in a digital backend to compensate the
Gm mismatch of the CFIA by applying a compensation voltage to an auxiliary
Gm stage. A linear interpolation algorithm, implemented in the digital backend,
is used to determine the appropriate compensation voltage [9, 10] (Sect. 6.3).
• The common-mode (CM) dependency of the input and feedback transconduc-
tors limits the gain accuracy even with DEM applied. To improve CM immu-
nity, threshold boosting and impedance balancing techniques are proposed,
resulting in a 3.2 9 improved gain accuracy without increasing the power
consumption [11] (Sect. 6.6.1).
7.5 Main Findings 181
7.5 Main Findings
The following list summarizes the main findings of this thesis:

• To improve the gain accuracy of a CFIA, DEM can be applied to the input and
feedback transconductors. This reduces the gain error from D to D2/2 and thus
significantly reduces their mismatch. Turning on the GERL improves matters
further, since it drives the mismatch D to zero. As a result, the transfer functions
of these two transconductances become even more closely aligned, which, in
turn, results in a further improvement in gain error, gain drift and linearity
(Sect. 5.6).
• After applying DEM, the residual gain error of a CFIA is not limited by D2/2,
but mainly by the CM dependency of its input and feedback transconductors
(Sects. 3.3.3 and 5.2).
• A power efficient way to increase the output impedance of a PMOS differential
pair is to use a low-threshold cascode transistor to fix the drain-source voltage of
the input transistors. However, the maximum output impedance is limited by the
threshold difference between the normal and low-threshold transistors, which is
determined by the CMOS process used (Sect. 6.6.1).
7.6 Other Applications of this Work
It has been shown in Sect. 4.5 that the ORL can be applied to general purpose
chopper CFIAs [12], chopper operational amplifiers [12] and capacitively-coupled
chopper instrumentation amplifiers [13]. Due to its continuous-time nature, it does
not cause noise folding. The same goes for the GERL. It can be applied in the
high-gain/low-frequency path of a multi-path amplifier, e.g. the one in [12] to
improve its low frequency gain accuracy.
7.7 Future Work
• The offset of the integrator of an ORL gives rise to extra ripple, which can not
be suppressed by the ORL (See Sect. 4.4.1). This offset can be eliminated by a
chopped and gain-boosted current buffer to isolate the demodulation chopper
from the sensing capacitors (Fig. 4.14). It also can be mitigated by autozero-
stabilizing the opamp in the integrator of the ORL (Fig. 4.13). Since chopping is
avoided in this approach, there is no chopper ripple at the integrator output,
which should eliminate the 2nd harmonic of the chopping frequency at the CFIA
output.
182 7 Conclusions
• In the read-out IC, an offset reduction loop is used to reduce the CFIA’s chopper
ripple. This loop also can be implemented in a digitally-assisted manner, i.e. by
using a DAC to generate an offset-compensating current.
• The CM dependency of the input and feedback transconductors limits the gain
accuracy even with DEM applied. To improve CM immunity, bulk-biasing and
impedance-balancing techniques have been applied to two transconductors
made with simple PMOS differential pairs (Sect. 6.6.1). These techniques could
also be applied to the improved CFIA with better gain accuracy, which has been
described in Chap. 5.
References
1. Wu R, Makinwa KAA, Huijsing JH (2008) The design of a chopped current-feedback

instrumentation amplifier. IEEE ISCAS, May
2. Wu R, Makinwa KAA, Huijsing JH (2009) A chopper current-feedback instrumentation
amplifier with a 1 mHz 1/f noise corner and an AC-coupled ripple reduction loop. IEEE
International Solid-State Circuits Conference Digest of Technical Papers, pp 322–323, Feb
3. Wu R, Makinwa KAA, Huisjing JH (2009) A Chopper current-feedback instrumentation
amplifier with a 1 mHz 1/f noise corner and anAC-coupled ripple reduction loop, IEEE J.
4. Luff GF (2010) Chopper stabilized amplifier. United States Patent, US 7724080 B2, 25 May,
2010.
5. Burt R, Zhang J (2006) A micropower chopper-stabilized operational amplifier using a sc
notch filter with synchronous integration inside the continuous-time signal path. IEEE J.
6. Kusuda Y (2010) Auto correction feedback for ripple suppression in a chopper amplifier.
IEEE J. Solid-State Circuits 45(8):1436–1445
with a gain error reduction loop and 0.06 % untrimmed gain error. IEEE ISSCC Dig. Tech.
Papers, pp 244–245, Feb
with a gain error reduction loop and 0.06 % untrimmed gain error. IEEE J. Solid-State
Circuits 46(12):2794–2806
transducers. Submitted to IEEE J. Solid-State Circuits.
transducers. IEEE International Solid-State Circuits Conference Digest of Technical Papers,
pp 110–111, Feb
transducers. IEEE International Solid-State Circuits Conference Digest of Technical Papers,
pp 241–244, Nov
12. Fan Q, Huijsing JH, Makinwa KAA (2010) A 21nV/HHz chopper-stabilized multipath
current-feedback instrumentation amplifier with 2 lV offset. IEEE International Solid-State
Circuits Conference Digest of Technical Papers, pp 80–81, Feb
13. Fan Q, Huijsing JH, Makinwa KAA (2010) A 1.8 lW 1 lV-offset capacitively-coupled
chopper instrumentation amplifier in 65 nm CMOS. ESSCIRC IEEE International Solid-
State Circuits Conference Digest of Technical Papers, pp 170–173, Sept
Summary
This thesis describes the theory, design and realization of precision interface
electronics for bridge transducers and thermocouples that require high accuracy,
low noise, low drift and simultaneously, low power consumption. This thesis is
dedicated to two aspects: (1) the design of precision stand-alone instrumentation
amplifiers (IAs) that can be used to drive an external Analog-to-Digital Converter
(ADC); (2) the design of a read-out IC that combines an instrumentation amplifier
and an ADC. Several new concepts and techniques have been proposed and
verified in CMOS technology.
Chapter 1
An introduction and motivation of the work described in this thesis is given in this
chapter. Precision bridge transducers and thermocouples typically output low-
frequency (LF) signals of a few Hz with millivolt levels. Therefore, they require
instrumentation amplifiers (IAs) with input-referred errors at the microvolt- or
nanovolt- level to boost such signals to levels compatible with the typical input
ranges of subsequent analog-to-digital converters (ADCs). Since sensor output
signals are often either ground-referenced or accompanied by a large common-
mode (CM) voltage, such IAs require ground-sensing capability and a high
common-mode rejection ratio (CMRR) ([120 dB).
Current-feedback Instrumentation amplifiers (CFIAs) are well-suited for bridge
read-out because of their high CMRR, ability to handle different input and output
CM voltages and power efficiency. However, their main disadvantages are limited
gain accuracy due to the mismatch of the input and feedback transconductors and
limited input range due to the non-linearity of these transconductors. This thesis
focuses on the design of improved CFIAs.
Furthermore, the CFIA can be used as a preamplifier and combined with an
ADC to comprise a read-out IC. For instrumentation applications, the incremental

Analog Circuits and Signal Processing, DOI: 10.1007/978-1-4614-3731-4,
184 Summary
DR ADCs are very suitable. The IAs in previous read-out ICs generally employed
switched-capacitor (SC) or two-opamp IA topologies. Neither of these topologies
is particularly power efficient. Since CFIAs are more power efficient, this thesis
presents the design of a read-out IC that combines a CFIA and an ADC, both of
which collaborate to achieve an optimum performance.
As a test-case, the challenging task of developing interface electronics for a
precision thermistor bridge is described. This is intended for use in wafer steppers
where lK-level temperature resolution is required. The resulting interface electronics
is also applicable to other sensors, e.g. strain gauges, thermocouples and Hall sensors.
Chapter 2
Chapter 2 gives an overview of dynamic offset cancellation techniques such as

chopping, auto-zeroing and offset-stabilization. It also shows how to apply these
techniques to operational amplifiers. It is shown that since chopping is a contin-
uous-time modulation technique that does not cause noise folding and is thus
superior than auto-zeroing. However, in stand-alone amplifiers, the modulated
1/f noise and offset need to be suppressed to ensure a ripple free output.
There are numerous ways to eliminate chopper ripple, such as auto-zeroing,
switched-capacitor (SC) or continuous-time (CT) notch filters or the use of an
auto-correction feedback loop. The SC sampling techniques incur a certain noise
folding penalty. The CT notch filter requires a good matching between the period
of the chopping clock and the notch position in the CT filter. Furthermore, all the
above-mentioned techniques suffer from the excess phase shift introduced by the
notch filter. Therefore, a high chopping frequency (or a low unity-gain-bandwidth
of the amplifier) is required to maintain stability. A new ripple reduction technique
is proposed in Chap. 4 that avoids all these issues.
Chapter 3
The use of dynamic offset compensation techniques is extended to precision CFIAs,

since they are well suited for bridge read-out. However, the gain accuracy of a CFIA
is rather limited due to the mismatch between its input and feedback
transconductors. Several techniques can be applied to improve its gain accuracy,
such as resistor-degeneration, auto-gain calibration and dynamic element matching
(DEM). However, resistor-degeneration requires significantly more power and auto-
gain calibration can not maintain a continuous output signal. DEM improves gain
accuracy by modulating the Gm mismatch to the DEM frequency, thus giving rise to
a signal dependent ripple. To eliminate the DEM ripple, trimming can be used.
However, it increases production costs and will not compensate for temperature
drift. A new technique is proposed in Chap. 5 that eliminates the need of trimming.
Summary 185
Chapter 4
The architectural design and implementation of a stand-alone chopper CFIA are

described. It consists of three gain stages, in which the input and intermediate
stages are both chopped at 30 kHz to suppress their 1/f noise and also to provide
sufficient gain to suppress the 1/f noise of the unchopped output stage. To suppress
the chopper ripple due to the offset of the input stage, a continuous-time offset
reduction loop (ORL) is proposed, while the chopper ripple associated with the
intermediate stage was suppressed by chopping it at a much higher frequency
(510 kHz).
The ORL uses a synchronous detection technique to demodulate the ripple, and
then drives the ripple to zero by continuously compensating for the offset. Due to
its continuous-time nature, the ORL does not cause noise folding. Furthermore, the
ORL is inherently stable, which is the key advantage compared to other ripple
reduction techniques using notch filters or auto-correction feedback loop. A low
chopping frequency can thus be chosen for low offset. Other authors have shown
that the ORL can also be applied to general-purpose chopper CFIAs and
operational amplifiers.
Measurement results show that the ORL reduces the amplitude of the chopper
ripple by a factor of 1100, to levels below the amplifier’s own input-referred noise
level. The CFIA achieves 1 mHz 1/f noise corner at a noise PSD of 15 nV/HHz
while consuming only 230 lA supply current (NEF = 8.8), which is quite
respectable compared to previous work. To the authors’ knowledge, this represents
the best LF noise performance ever reported for a stand-alone CMOS
instrumentation amplifier.
Chapter 5
The chopper CFIA described in the previous chapter achieves microvolt-level

offset and a high CMRR ([120 dB). However, its gain error, mainly determined
by the mismatch of the input and feedback transconductors (noted as ‘‘D’’), is
about 0.5 %, which becomes the dominant source of residual error. Thus, the
design and implementation of a CFIA with improved gain accuracy are discussed
in this chapter.
To improve gain accuracy, dynamic element matching (DEM) is applied to the
input and feedback transconductors of the CFIA, so as to average out their
mismatch. DEM modulates the Gm mismatch to the DEM frequency, thus giving
rise to a signal-dependent ripple at CFIA output. To suppress this ripple, a gain
error reduction loop (GERL) is proposed to continuously null the mismatch of the
input and feedback transconductances, thus eliminating the need for trimming.
Unlike the ORL, which feeds back an additive offset-compensating signal, the
GERL feeds back a multiplicative gain-compensating signal, which adjusts the
186 Summary
ratio of the input and feedback transconductances. The output DEM ripple is then
the product of the mismatch and the output signal, and so the gain of the GERL
will be signal dependent. To guarantee negative feedback, a polarity reversing
switch is used to link the polarity of the GERL to that of the output signal.
The loop gain of the GERL is proportional to the input signal and so is zero for
zero input. In this case, leakage causes the integrator output Vint,GE to drift with a
time constant of several seconds and eventually clip. The GERL will then need to
resettle whenever a finite input signal re-appears. To avoid the need for resettling,
a digitally-assisted GERL is implemented to store the mismatch information in the
digital domain in this circumstance. For comparison, the analog implementation of
the GERL is also implemented.
DEM reduces the gain error from D to D2/2, moving the average input and
feedback transconductances closer to each other. This results in a CFIA with
improved gain error, gain drift and linearity. The GERL improves matters further,
since it drives the mismatch to zero. As a result, the average input and feedback
transconductors become even more closely aligned. Finally, the use of DEM and
the GERL also increases the linear input range of the CFIA by a factor of three.
Measurement results show that without trimming, the CFIA achieves a gain
error of less than 0.06 % and a maximum gain drift of 6 ppm/8C in a power
efficient manner (NEF = 11.2). Compared to a CFIA with similar gain accuracy,
but using resistor-degenerated input stages, this represents a 49 improvement in
power efficiency, which is equivalent to a 169 less power when achieving the
same noise level. These measurement results confirm that the combination of
DEM and the GERL is a power-efficient manner of improving the gain accuracy,
gain drift and linearity of a CFIA.
Chapter 6
The CFIA described in Chap. 4 is then combined with a switched-capacitor

sigma–delta (DR) ADC to realize a read-out IC. The system-level design and
implementation of the read-out IC are described. The CFIA provides high input
impedance for bridge read-out and relaxes the noise and offset requirements of the
ADC. The ADC employs a topology whose gain accuracy does not depend on
component matching. Furthermore, the use of a ratio-metric topology means that
the accuracy of the ADC’s reference is much relaxed. These two solutions enable
the ADC to achieve a gain error of less than 2 ppm. Thus, the gain error of the
read-out IC is mainly determined by the mismatch between the input and feedback
transconductors of the CFIA.
To eliminate this mismatch, DEM is applied to the two transconductors to
average out their mismatch. However, the CM dependency of these
transconductors limits the achievable gain accuracy even with DEM applied. To
enhance their CM immunity, bulk-biasing and impedance-balancing techniques
are employed. To reduce gain error and gain drift further, a digitally-assisted gain
Summary 187
error correction (GEC) scheme is applied, which digitally processes the output of
the ADC and feeds back a gain error correcting signal. This improves the gain
accuracy and gain drift of the CFIA. Overall, the GEC path serves as a coarse-trim
of Gm mismatch, while the DEM acts as a fine-trim that compensates for
temperature drift.
To reduce offset to the nV-level, system-level chopping is employed to chop the
entire read-out chain during multiple conversions. The modulated offset is then
averaged out in the decimation filter. It has been found that the combination of
input stage chopping in the CFIA and system-level chopping is a better way to
suppress 1/f noise and offset, compared to the use of multi-stage chopping.
Measurement results show that the former achieves 0.1 mHz 1/f noise corner,
while the latter achieves 1 mHz 1/f noise corner. Furthermore, in the former
approach, the choppers in the intermediate stage of the CFIA are off, thus avoiding
extra offset due to the coupling of charge injection and clock spikes through the
Miller-compensation capacitor. The ultimate residual offset of the read-out IC is
then determined by its resolution and is about 48 nV.
Measurement results show that the read-out IC achieves state-of-art 1/f noise
corner (0.1 mHz), offset (48 nV), gain drift (1.2 ppm/8C), offset drift (6 nV/8C)
and power efficiency (FOM = 111 pJ/Conv). These qualities make the proposed
read-out IC very suitable for demanding bridge transducer applications, which
require low thermal and 1/f noise, high accuracy, low drift, and simultaneously,
low power consumption.
About the Author
Rong Wu was born on November 4, 1981.

She received the B.Eng. degree in
microelectronics from Fudan University,
Shanghai, China, in 2003. After one year
graduation study in Fudan, she started the
M.Sc. study in electrical engineering at
Delft University of Technology, Delft, The
Netherlands, in September, 2004. She
received her M.Sc. degree of TU Delft in
February, 2006 and her M.Sc. degree of
Fudan University in July, 2006, both on
electrical engineering.
In October 2006, she joined the
Electronic Instrumentation Laboratory of
the TU Delft, pursuing her Ph.D degree on
the subject of precision amplifier and
sigma–delta ADCs for sensor read-out. In
December 2011, she received her Ph.D degree. Currently, she is with precision
ADC group of Analog Devices in Wilmington, MA. Her research interests include
sensors, precision analog and mixed-signal interface electronics.

Index
A Clock feed-through, 21, 24, 30–32

1/f noise, 4, 7, 10, 14, 15 CM immunity, 159, 161, 162, 177, 180, 182
CM-dependent mismatch, 64, 108, 152, 153
CMRR, 3, 5, 14, 15, 51, 52, 58, 70, 71, 103, 107,
A 121, 122, 132, 134, 152, 163, 170, 176
AC-coupled, 47, 74, 76 CMOS, 54, 69, 72, 98, 104
ADC, 3, 4, 7, 11 Common-mode (CM), 3, 15, 31, 59, 64
Additive, 110, 116 Common-mode (CM) dependency, 64, 108,
Additive multiplicative, 110, 117 180
Auto-gain calibration, 61, 66 Compensation, 2, 21, 28–30
Auto-zeroed, 24, 26, 30, 37, 38 Continuous-time, 6, 8, 21, 26, 28, 29
Auto-zeroing, 21, 23–26 Conversion time, 138, 140, 141, 174
Automatic offset reduction Correlated double sampling (CDS), 40
Average, 33, 34, 37, 62–64, 107, 108, 117, Current-feedback instrumentation amplifier
137, 141, 142, 151, 152, 154, 156, 177, (CFIA), 107, 118, 124, 135, 137
185, 186 Current-mode, 5, 8
B D
Bias current, 34, 35, 47, 53, 54 Decimation, 11–13, 36, 137, 138, 153, 156,
Bridge transducers, 1, 3, 23, 148, 177 163, 176, 187
Bulk biasing, 170–172, 177, 182, 186 Delta–sigma (DR) modulator, 11
Digitally-assisted, 114–116, 124–128
Dynamic element matching, 10, 18, 51, 62,
C 107, 108, 135, 151, 177, 180,
Capacitively-coupled, 5–8, 71, 181 184, 185
Cascode buffer, 81, 83, 91, 119, 163 Dynamic offset cancellation, 4, 10, 21, 22,
Charge injection, 25, 26, 30, 31, 34 37, 54
Chopper ripple, 18, 21, 28–30, 36, 42–47, 55,
57, 69, 74, 76, 81, 101–103, 116, 118,
119, 121, 128, 129, 141, 155, 156, 159, E
160, 168, 176, 179, 181 Effective number of bits (ENOB), 11, 145
Chopper stabilization, 19 Even-harmonics, 83
Chopping, 7, 13, 21, 23 Excess phase shift, 47, 74

192 Index
F Multiplicative, 110, 116

Feedback, 164, 169, 171, 176, 177, 179 Multi-stage chopping, 138, 155, 168, 170, 176,
Feed-forward, 141–143, 165, 166 179, 180
G N
Gain accuracy, 9, 10, 14, 16–18, 41, 51 Nested-chopping, 155, 156
Gain drift, 11, 14, 15, 174, 177 Noise folding, 6, 12, 18, 27, 29, 46, 71
Gain error, 4, 6, 10, 70, 74, 117 Noise-shaping, 11
Gain error reduction loop, 10, 18, 109, 114, Nonlinearity compensation
118, 180 Notch filter, 42, 44–47, 74, 76, 102, 180
General purpose, 16, 38, 69, 70, 84,
104, 181
Ground-sensing, 59 O
Guard band, 36 Offset, 4, 7, 10, 14, 21, 70, 104,
guard time, 36, 37 156, 182
Offset gain error, 14
Offset reduction loop, 10, 18, 74, 102,
H 176, 182
Hall sensors, 1–3, 16 Operational amplifier, 4, 5, 16, 17, 21, 24, 38,
40, 41, 44–47, 55, 56, 69, 181
Output impedance, 88, 160–162, 181
I
Impedance balancing, 162, 170, 172, 177, 180,
182, 186 P
Incremental, 12, 13, 138, 140, 141 Periodic noise analysis (PNOISE), 72, 158
INL, 117, 118, 129, 131, 148, 171, 174 Periodic steady-state (PSS), 72, 147, 158
Input impedance, 3, 6, 7, 32, 51, 176 Phase shift, 45–47, 74, 79, 80, 103, 112,
Input offset current, 33 180, 184
Input bias current, 34, 35, 47 Phase shift comparator, 45
Integrator leakage, 145–147 Ping-pong, 21, 37, 66, 135
Instrumentation amplifier, 3–5, 9, 11, 15, 51, Ping-pong-pang, 62, 66
52, 54, 69, 102, 107, 137, 169, 171, Power efficient, 5, 9, 42, 153, 159
176, 181 PSRR, 15, 70, 103, 132, 134
K R
kT/C noise, 6, 145 Ratio-metric, 139, 148, 177
Read-out IC, 174, 180, 182
Resistor-degenerated (degeneration), 59, 65,
L 107, 135, 163, 170
Linear interpolation, 153, 177, 180 Resistor-degeneration, 10, 17, 66, 135
Linearity, 5, 8, 12, 14, 51, 136, 138, 148 Ripple reduction, 21, 74, 80, 103, 179
Loop-gain, 144
Low-pass, 28, 36, 76, 79, 112, 125, 153
Low-threshold, 88, 89, 98, 122, 127, S
160, 181 Sample and hold, 7, 42, 44
Low-threshold cascode, 88, 89, 122, 181 Self-heating, 4, 16, 70
Sensor, 1, 2, 7, 16, 137
Settling, 27, 46, 116, 150, 166
M Signal-dependent, 54, 65, 66, 113, 135,
Modulation, 21, 28, 29, 40, 71, 93, 116, 119, 165, 180
160, 181 Sinc2 filter
Multiplexer, 108, 153 Sinc3 filter, 174
Index 193
Start-up, 74, 116, 133, 153, 155 T

Strain gauge, 2, 3, 16, 107, 137 Temperature drift, 14, 23, 66, 109, 135, 153,
Swapper, 108, 112, 117, 152 174
Switched-capacitor, 5, 74, 80, 145 Thermistor bridge, 1, 16, 174
Synchronous demodulator Thermocouple, 2, 4, 16, 23, 137
(demodulation), 76 Three-opamp, 5, 91
System-level chopping, 13, 137, 148, 158, 176 Trimming, 10, 21, 113, 135, 177, 180

Wu2013 PDF

Uploaded by

Copyright:

Available Formats

Wu2013 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wu2013 PDF

Uploaded by

Copyright:

Available Formats

Analog Circuits and Signal Processing

For further volumes:

ISBN 978-1-4614-3730-7 ISBN 978-1-4614-3731-4 (eBook)

Library of Congress Control Number: 2012940238

Ó Springer Science+Business Media New York 2013

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

2 Dynamic Offset Cancellation Techniques

2.4.2 Charge Injection and Clock Feed-Through

3 Current-Feedback Instrumentation Amplifiers and Gain Accuracy

4 A Chopper Instrumentation Amplifier with Offset

4.5 Applying ORL to General Purpose Instrumentation

5 A Chopper Instrumentation Amplifier with Gain Error

6 Read-Out Integrated Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

6.3 Gain Accuracy Improvement Techniques in the Read-Out IC . .. 151

About the Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

R. Wu et al., Precision Instrumentation Amplifiers and Read-Out Integrated Circuits, 1

(a) (b) (c)

1.2 Overview of Read-Out Electronics for Sensors

Fig. 1.3 Read-out IC for Sensor

As the amplifier is used to detect a small differential signal, its input-referred

Fig. 1.4 Bridge

1.3 Instrumentation Amplifier Topologies

Instrumentation amplifiers can be built in several ways. The commonly used

1.3.1 Three-Opamp Topology

The three-opamp IA uses voltage feedback to obtain a gain determined by resistors

1.3.2 Switched-Capacitor Topology

Fig. 1.5 Switched-capacitor Φ1

Fig. 1.6 Capacitively- Cm

1.3.3 Capacitively-Coupled Topology

A recent development is the continuous-time capacitively-coupled (and chopped)

Fig. 1.7 Block diagram of an improved capacitively-coupled analog-to-digital interface [22]

V-to-I Opamp -C S&H Amplifier

Fig. 1.8 A closed-loop capacitive-feedback sampler [22]

Fig. 1.9 Current-mode Current Mirror 1

1.3.4 Current-Mode Topology

Figure 1.9 shows a current-mode instrumentation amplifier [23–26]. The feedback

1.3.5 Current-Feedback Topology

Figure 1.10 shows a current-feedback instrumentation amplifier (CFIA). The input

Fig. 1.10 Current-feedback C1

Gm2 R1 þ R21 þ R22

1.4 Current-Feedback Instrumentation Amplifier

bipolar CFIA used in low-power biomedical applications was reported by Hamstra

Table 1.1 Performance of low-offset CFIAs

1.5 Read-Out ICs

Fig. 1.11 Block diagram of a DR ADC

and a digital-to-analog converter (DAC). The oversampling ratio (OSR) is defined

SC 0 fLF 2fLF 3fLF 4fLF Freq

Fig. 1.12 Block diagram of the read-out IC in [39]

Fig. 1.13 Block diagram of read-out IC in [39]

fslow ffast ffast fslow

Fig. 1.14 Block diagram of hall sensor interface in [46]

results in a high power consumption of 80 mW to achieve an 18-bit resolution with

1.6 Targeted Sensor Applications and Challenges

As a test-case, the challenging task of developing interface electronics for a pre-

Fig. 1.15 Dual thermistor

Metal Foil R1 Rx Thermistor

Table 1.2 Typical sensor specifications and targets of the read-out IC

Precision CFIA & Gain Accuracy (1)