Mit PDF

MASSACHUSETTS INSTITUTE
OF TECHN'_OY
i
NOV 13 2008
LIBRARIES
Package and PCB Solutions for High-Speed

Data Link Applications
by
Asad Kalantarian
Bachelor of Science in Electrical Engineering and Computer Science

Massachusetts Institute of Technology (2007)
Submitted to the Department of Electrical Engineering and Computer Science

in Partial Fulfillment of the Requirements for the Degree of
Master of Engineering in Electrical Engineering and Computer Science
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
August, 2008
©2008 Massachusetts Institute of Technology. All rights reserved.
Author:
Department of Electrical Engineering and Computer Science
August 27, 2008
Certified by:
Nanju Na, Dr.
Senior Engineer, IBM
Thesis Supervisor
Certified by:
/ / Vladimir Stojanovic
Professor of Electrical Engineering
.. Thesis Supervisor
Accepted by:
.- Arthur C. Smith
Professor of Electrical Engineering
Chairman, Department Committee on Graduate Theses
ARCHNIVES
Package and PCB Solutions for High-Speed Data
Link Applications
by
Asad Kalantarian
Submitted to the Department of Electrical Engineering and Computer Science
August 27, 2008
In Partial Fulfillment of the Requirements for the Degree of

Master of Engineering in Electrical Engineering and Computer Science
ABSTRACT
Continual increase in high-speed transfer rates is essential in today's world in order to
keep up with Moor's law scaling and to meet application demands. This increase in
information transfer rates is essentially limited by the bandwidth of the communication
channel. Channel loss, signal crosstalk, and power integrity are the important factors
affecting the bandwidth of the channel and must be fully understood when designing such
high-speed links. Moreover, channel performance optimization requires system level analysis
with co-design of different channel components in order to approach the maximum capacity
of a channel. In this thesis, we focus on the essential aspects of high-speed channel design
with an emphasis on the design of the package area. A technique for reducing discontinuities
in the package is introduced, studied, and simulated on IBM package designs to obtain
improved package performance. The critical Plated-Through-Hole (PTH) via region at the
package and Printed Circuit Board (PCB) interface is also analyzed in a co-study to suggest
design rules for overall improved system performance.
Thesis Supervisor: Nanju Na

Title: Senior Engineer, IBM
Thesis Supervisor: Vladimir Stojanovic

Title: Professor of Electrical Engineering
Acknowledgments
I would like to express my greatest gratitude to the ASICs Packaging team at IBM
Microelectronics in Vermont for their valuable help and support throughout my work at
IBM. I would like to thank my manager Andy Anderson for his support, encouragement, and
valuable discussions on different work and life related subjects. I am especially grateful to
Nanju Na, my supervisor at IBM, who patiently introduced me to electronics packaging field,
and for her willingness to help guide my work despite her extremely busy schedule at IBM.
I further want to thank Carrie Cox, Deborah Zwitter, Erwin Cohen, Eric Tremble, and
Kent Dramstad for their precious help and training throughout the different stages of my
thesis work.
I am especially grateful to my thesis advisor, Vladimir Stojanovic, for his help and
knowledge, and for pulling through for me especially towards the end.
Most importantly, I thank my Creator, the Almighty, Whom has given me beyond what
I deserved and Whom I can never thank enough.
Dedication
To my mom, my dad, my two sisters, and my brother
Contents
Chapter 1 : Introduction ..................................................................................... 17

Chapter 2 : Background ........................................................................................ 19
2.1 Chip ................................................................................................................. 20
2.2 Package ........................................................................................................... 21
2.3 Connector ...................................................................................................... 23
2.4 Printed Circuit Board (PCB) ............................................................................. 24
Chapter 3 : Signal Integrity of High-Speed links................................... ........ 27
3.1 Transmission Line Losses ...................................................... 27
3.1.1 Conductor DC Losses........................................................................ 28
3.1.2 Dielectric DC Losses .......................................................................... 28
3.1.3 Frequency-Dependant Conductor Losses.............................. ....... 28
3.1.3.1 Frequency-Dependant Resistance and Inductance ................... 29
3.1.3.2 Effect of Conductor Surface Roughness ....................................... 29
3.1.4 Frequency-Dependant Dielectric Losses.............................. ........ 30
3.2 Reflections (Impedance Discontinuities and Stubs) ............................................. 32
3.3 Crosstalk (FEXT, NEXT) ............................................................ ........................ 33
3.3.1 Mutual Inductance and Mutual Capacitance ..................................... 34
3.3.2 Types of Crosstalk ..................................................................................... 34
3.4 Power Integrity (digital, analog) ........................................................................... 35
Chapter 4 : Discontinuity Cancellation at the Package Level ....................................... 39
4.1 The Counter-Discontinuity Cancellation Technique Description ...................... 40
4.2 Theory: Distributed Network Behavior of Short Discontinuities ....................... 41
4.3 3D Package Structure Simulations with Discontinuity Improvement .................. 44
4.4 Time Domain System Analysis ................................................... 47
Chapter 5 : PCB Plated-Through-Hole (PTH) Via Design ..................................... 51
5.1 PCB PTH-Via Structure Setup .................................................... 52
9
5.2 Signal Transmission .............................................................. ........................... 56
5.2.1 Time-Domain System Analysis ........................................ .......... 60
5.3 Signal Coupling...................................................................................................... 63
Chapter 6 : Conclusion ............................................................. 69
Chapter 7 : Future W ork ............................................................ ............................ 73
Bibliography............................................................................................................. 75
10
List of Tables
Number Page
Table 3.1: Loss Tangent for Typical PCB insulator. ....................................... ....... 31
Table 4.1: Design variables considered for opposite phase impedance insertion and their
effect on the characteristic impedance ...................................... ....... 46
Table 4.2: Time-Domain simulation of discontinuity cancellation for a typical chip-to-chip
channel ............................................................................................................ 49
Table 5.1: Stubbed PCB Via Configuration ......................... 53
Table 5.2: 11Gbps time-domain eye simulation with FFE ............................................. 62
Table 5.3: 6.2Gbps time-domain eye simulation with FFE ...................................... 62
List of Figures
Number Page
Figure 2.1: Diagram of a high-speed serial link.................................... ............ 19
Figure 2.2: Typical backplane communication channel indicating the different sections of
the signaling path .................................................................................. 20
Figure 2.3: Common Methods of die attachment: (a) wire bond attachment; (b) flip chip
attachm ent...................................................................................................... 22
Figure 2.4: Cross-section of PCB trace configurations defined by parameters w, t, h, and
b: Stripline (left), M icrostrip (right) ....................................... ....... .. 24
Figure 3.1: Cross section of a stripline in a typical PCB, showing surface roughness. ..... 30
Figure 3.2: Crossover between skin-effect and dielectric loss, FR4 8 mil wide and 1 m
long 50 ohm stripline [25] .................................................................... 31
Figure 3.3: Transmission line system with multiple line impedance ............................. 32
Figure 3.4: Example NEXT (blue) and FEXT (red) in a modern package (left) and a modern
1 meter cable (right) ............................................. ............. .......................... 35
Figure 3.5: Power supply circuit including package parasitic model ............................. 36
Figure 3.6: Chip current demand waveform (left) along with the corresponding voltage
supplied through the package analog power nets with reference to on-chip
ground ............................................................. 37
Figure 4.1: High-speed net in a flip-chip package. Solder balls are highly capacitive...... 40
Figure 4.2: Illustration of transmission and reflection on a transmission line with a
discontinuity.................................................................................................... 41
Figure 4.3: Transmission lines with an extra discontinuity inserted (a) at one
discontinuity end and (b) at both discontinuity ends of existing discontinuity
Z1 ................................................ 42
Figure 4.4: Distributed transmission line with a discontinuity represented by lumped
elem ents ............................................. ...................................................... 4 3
13
Figure 4.5: Insertion loss, S21, simulated using lossy ideal transmission lines with varying
line impedances of (a) 0.3mm, (b) 0.5mm, and (c) 1mm for inserted counter-
discontinuity Z2. Z2=502 represents no discontinuity insertion ......... 44
Figure 4.6: Insertion loss of a differential signal, SDD21, through solder balls of 1mm
pitch and PTH vias in a 4-layer EPBGA package where package and board
traces are designed with (a)900 2mm, (b) 70Q 2mm, (c)90 1mm
combined with 502 1mm, (d) 70Q 1mm with 5020 1mm, (e) 902 0.5mm with
500 1.5mm, (f)70n 0.5mm with 50ohm 1.5mm and (g) 50Q 2mm which is a
nominal design structure with no discontinuity insertion, that is the basis for
com paring the other cases. ....................................... .................................. 45
Figure 4.7: Insertion loss of a differential signal, SDD21, through solder balls and
PTH/blind vias in 4 layer FBGA package varying plane voids above balls, via to
via traces and routing trace impedance on package and board. Structure (a}
represents a typical package design with no plane voids, minimum anti-via
diameter and escaping traces of matching impedance and is used as a base
structure to measure performance improvement of other design options.. 47
Figure 4.8: High-speed channel used for time-domain system simulation ................... 48
Figure 5.1: Modeled 16 layer PCB structure on the right displaying the via traces in red;
Dimensions and structure of a differential via pair on the left ................... 52
Figure 5.2: Via configurations with different signal densities ...................................... 54
Figure 5.3: Top and 3D view of the Box, Diagonal, and Staggered via configurations..... 55
Figure 5.4: Insertion loss for differential vias with back-drilled stubs (longer via lengths)57
Figure 5.5: Insertion loss for differential vias with back-drilled stubs (shorter via lengths)58
Figure 5.6: TDR impedance waveform of different length differential vias .................. 58
Figure 5.7: Insertion loss for three differential via pairs with stubs............................. 59
Figure 5.8: High-speed channel used for time-domain simulations......................... 61
Figure 5.9: Insertion loss plots for channels with different PTH-vias. Left plot: worst case
channel for 11Gbps simulation. Right plot: channel with package PTH-via and
6.2Gbps transfer rate simulation.......................... ...... ............... 63
14
Figure 5.10: Eye diagram for 11Gbps simulation of a) Layer 12 PTH-via with worst case
channel, b) Layer 12 PTH-via, and c) Layer 1 PTH-via............................. . 63
Figure 5.11: Near-End crosstalk (NEXT) for the Box, Staggered, and Diagonal via
configurations ............................................................... ............................ 64
Figure 5.12: Far-End Crosstalk (FEXT) for the Box, Staggered, and Diagonal via
configurations ............................................................... ............................ 65
Figure 5.13: Differential NEXT for the diagonal configuration for layer 2 back-drilled
(Blue), layer 5 back-drilled (Green), and layer 5 with stub (Red) ................ 66
Figure 5.14: Differential FEXT for the diagonal configuration for layer 2 back-drilled
(Blue), layer 5 back-drilled (Green), and layer 5 with stub (Red) ................ 66
Figure 5.15: Differential NEXT coupling for different via lengths with back-drilled stubs67
Chapter 1 : Introduction
Higher bandwidth has become more important than ever in today's computing
systems. Personal computers, routers, switches, and game consoles all require higher
bandwidth to meet the increasing performance demand of new applications. Moreover, the
continuous scaling of integrated circuit technology, confirming Moore's prediction, over the
recent years has resulted in massive computational capacity and hence data processing
capability which in turn has created the demand for high-speed communication across
different components in a system [1]. These systems extend to optical communication
networks spanning across the globe, but all come down to chip-to-chip communication in a
single board [2-4]. The massive flux of information in and out of the chip has caused simple
input/output (I/O) drivers to be replaced with sophisticated high-speed circuits which in
turn depend on reliable high bandwidth channels.
Channel design, which was conveniently and justifiably ignored at lower frequencies,
has become a major bottleneck for high-speed communication [5]. The increase in data
rates to the tens of Giga bits per second (Gbps) region has prompted more careful signal
integrity considerations in the design of the channel from the transmitter of one chip to the
receiver on the next [6]. The decrease in wavelength size due to higher frequency signaling
has caused the once short electrical lengths of different components to become significant
due to transmission line delays, loss and signal coupling in these components [7].
Different channel components are traditionally designed separately with the goal of
optimizing each component in isolation, assuming overall channel performance
improvement. Modern work in high-speed channel design, however, requires an overall
system analysis to quantify component level improvements as we show in this thesis.
This thesis presents an in depth analysis of different bottleneck areas in high-speed

channel design with a focus on the package (substrate) and PCB regions. A technique for
reducing inherent discontinuities at the package level is studied and performance
improvement is obtained at relatively low cost. In addition, the critical region of package
and PCB interface is analyzed and system level simulations are performed, in time and
frequency domain, in order to allow for better design rules for package and PCB design.
Chapter 2 provides an overview of the structure of a typical high-speed channel and

the different components which constitute the channel in a chip-to-chip application. The
different causes impacting signal integrity in high-speed channels is outlined in Chapter 3,
providing the theoretical challenges faced in high-speed channel design. The effects of
signal losses, reflection, crosstalk and power integrity on signal integrity is described to
provide the fundamentals for the work in this thesis.
As a step towards package performance enhancement, a concept of discontinuity

cancellation for package design is presented in Chapter 4 and is rigorously analyzed with 3D
electromagnetic structure simulations which are modeled based on IBM package offerings
to quantify performance improvements seen both at the package and channel level along
with the impact on the channel bandwidth.
It is recognized that isolated package design cannot be the solution to overall

optimized channel performance nor the effective way of obtaining minimized cost for
performance and thus a package-board co-study is performed at the critical area of Printed
Circuit Board (PCB) Plated-Through-Hole (PTH) via design. Chapter 5 outlines the methods
used for this study and the impact on both component level and more importantly system
level performance. Different tools are used to quantify this performance impact in high-
speed channels both in the time-domain and in the frequency domain.
Chapter 6 concludes this work with the key understandings obtained from this work
that can be used to achieve better performance/cost optimized design rules for high-speed
package design. Chapter 7 provides an outline for future work in this area.
Chapter 2 : Background
Due to the density constraint on the number of wires between chips in chip-to-chip
communication and the increased demand for higher flow of information, blocks of parallel
data are serialized and transmitted off chip using high-speed links by utilizing Serialize/De-
serialize circuitry. A general structure of a high-speed serial link is given in Figure 2.1. The
channel can contain different components such as the package, connector, PCB, and
possibly the socket.
bl bl
-b2
lb3 - ------ ba
b3
bi
b4 ---- - b4
*
Figure 2.1: Diagram of a high-speed serial link
The use of these high-speed links has become ubiquitous today. Around eighty
percent of Application Specific Integrated Circuits (ASIC) produced by IBM Microelectronics
today contain High-Speed SerDes (HSS) links. SerDes, which stands for
serializer/deserializer, is a circuit technique for rapid transmission of high-speed data [8]. In
the past few years we have witnessed a rapid increase in data transmission rates over
different processing technologies.
At this stage, the bottleneck in achieving higher HSS signaling rates lies in the
limitation of channel bandwidth [9]. In order to address this bandwidth limitation, several
designs started to use transmit pre-emphasis to compensate for the low-pass nature of the
channel [10-12]. Four-level signaling has also been considered as another solution, instead
of binary NRZ signaling, to increase the data rates without increasing the signaling rate
beyond the channel limit [13, 14]. However, such sophisticated equalization and modulation
techniques increase link cost and limit the achievable bandwidth since this complexity
comes with significantly larger power requirements. Thus improving channel performance
with simple solutions can keep the cost of the link low while simplifying the link circuits and
improving their energy-efficiency. For these reasons, this thesis focuses on solutions to
maximizing channel efficiency while keeping the link complexity at a minimum.
The channel bandwidth limitation is highly dependent on the channel design, structure
as well as channel length in the application space. The materials used, the component
physical structure, signal coupling due to compact routing, and power integrity are all
important factors which directly and indirectly impact the bandwidth of the channel. In
order to improve the band-limited channel, we thus need to look at the structure of the
channel as well as the mechanism that produces this limitation.
A typical backplane system depicting the different sections of the signaling path in a
backplane communication channel is displayed in Figure 2.2. The typical channel consists of
transmitter and receiver chips, the packages for the chips, connectors, and mother board
traces. The subsequent sections expand upon each of these components.
Figure 2.2: Typical backplane communication channel indicating the different sections of the
signaling path
2.1 Chip
The chip consists of the silicon circuitry as well as the metal layers for signal and
power redistribution. High-speed transceivers with potential sophisticated channel
equalization circuitry reside in the chip. Channel equalization can compensate frequency-
dependant loss and dispersion of long traces in boards and packages as well as dispersion
due to device loadings [15, 16]. Feed Forward Equalization (FFE) and Decision Feedback
Equalization (DFE) are two commonly employed equalization techniques in the design of the
transceiver, in order to flatten the channel frequency response [17].
The aggregate termination impedance becomes frequency selective because of

parasitic capacitance and does not match the transmission line characteristic impedance of
50 Ohms at all frequencies. Electrostatic Discharge (ESD) protection circuitry is a major
source of this parasitic capacitance which if not accounted for can be a source for signal
degradation [18].
2.2 Package
The package is the holder of the chip and is essentially a space transformer between
the chip and the motherboard. It provides the mechanical, thermal, and electrical
connections necessary to properly interface the circuits in the chip to the rest of the system.
The physical attributes of a package can be divided into three categories: the attachment of
the die (chip) to the package, the on-package connections, and the attachment of the
package to the PCB [19].
2.2.1 Attachment of the Die to the Package
Figure 2.3 depicts the two most common methods of die attachment: wire bond
attachment and flip chip attachment. Wire bond is cheaper than flip chip but is limited in its
performance due to the increased amount of parasitic impedance it introduces. A wire bond
is a very small wire with a diameter on the order of 1 mil (diameter of a human hair is about
3mils, where one mil is a milli-inch). The primary impact of the bond wire is its added series
inductance. The length of a bond wire is on the order of 100mils, which is significant at the
operating frequencies and thus produces transmission line behavior. In addition, wire bonds
also exhibit large amount of crosstalk, which limit the proximity of signal nets to each other
on the wire bonds and thus the number of I/Os that a wire bond package can provide [20,
21].
Die --.... Solder ball
Bond wire
x On-package routing
(a) Chip is mounted (b) Chip is mounted with

with the pads on top the pads on the bottom
Figure 2.3: Common Methods of die attachment: (a) wire bond attachment; (b) flip chip
attachment.
Flip-chip packaging, on the other hand, is almost ideal form an electrical perspective.
As depicted in Figure 2.3, a flip-chip connection is obtained by placing small balls of solder
on the pads of the die and then placing the die upside down on the package substrate. The
series inductance of the flip-chip connection is on the order of 0.1 nH, which is an order of
magnitude less than that of a typical wire bond [19]. Other advantages are that the effect of
crosstalk in the solder balls (also called Controlled-Collapsed-Chip-Connector or C4) is
minimal, and the I/O count on the flip-chip package is larger due to the fact that the solder
ball connections can potentially be placed over the entire die, not just on the periphery.
Thermally and mechanically, however, flip-chip packaging is dismal. The thermal

coefficient of expansion must be similar between the die and the substrate so that heating
of the chip doesn't cause the solder balls to be strained. In addition, cooling of the chip is
more challenging and requires costly methods since the whole side of the die is covered
with solder balls and this reduces heat transfer. For these reasons, flip-chip packaging is
more costly and is mainly used for high performance applications.
2.2.2 Routing of the Signals on the Package
The routing of the signals in the package can be done in a controlled or non-controlled
impedance fashion. High-speed digital designs require the use of controlled impedance
packaging and this isthe type we are concerned with in this thesis. A controlled impedance
package typically resembles a miniature PCB with different layers and power and ground
planes with the die placed at the center. Both wire bond and flip-chip connections can be
used to connect the die for routing through the package. Small dimensioned transmission
lines are used to route the signals from the bond pad on the package to the board
attachment. The different layers are connected through buried, blind, and plated-through-
hole vias for signal routing as well as for power delivery. Controlled impedance routing
requires careful design for maximum signal transmission and minimum coupling between
signal nets. Power delivery through the package also requires minimum inductance to
reduce switching noise.
2.2.3 Attachment of the Package to the Board
Similar to the attachment of the die to the package, there are a variety of methods to
connect the package to the PCB. The most common method used in ASIC packaging is the
Ball Grid Array (BGA) method which is essentially a larger version of flip-chip. The BGA
descended from the older Pin Grid Array (PGA) connection type. Land Grid Array (LGA) has
also recently gained some ground with most of Intel's current processors being mounted
using LGA. LGA uses pads of bare gold-plated copper to connect the package to the socket
on the PCB.
2.3 Connector
Connectors are devices that are used to connect one PCB board to another. An
example of a high-speed connector is the slot connector used to connect the Pentium III
processor to the motherboard. Despite their relatively shorter electrical lengths, connectors
contribute significantly to the loss in a high-speed channel. The primary electrical factors
affecting high-speed performance in connectors are: [22]
1 Mutual inductance - causes crosstalk
2 Series inductance - slows down signal propagation and creates electromagnetic

interference (EMI)
3 Parasitic capacitance - Slows down signal propagation

2.4 Printed Circuit Board (PCB)
A printed circuit board is used to mechanically support and electrically connect
electronic components using conductive pathways, or traces. These traces are etched from
copper sheets laminated onto a non-conductive substrate. PCBs are rugged, inexpensive,
and can be highly reliable.
The typical PCB stackup consists of alternating layers of core and prepregs (short for
pre-impregnated). A core layer is a thin piece of dielectric with copper foil bonded to both
sides. The core dielectric is cured fiberglass-epoxy resin. Prepeg layers are uncured
fiberglass-epoxy resin which cure (i.e. harden) when heated and pressed. The most
common, widely available, and relatively low cost dielectric material used in the PCB
industry is FR-4 (woven glass and epoxy).
PCB traces are used for high-frequency connections within a single PCB and on
backplanes connecting one or more PCBs. The traces in the PCB are generally either
microstrip or stripline. Microstrip traces have a solid reference plane on one side while
striplines lie between two solid reference planes, as shown in Figure 1.5.
Planes
Figure 2.4: Cross-section of PCB trace configurations defined by parameters w, t, h, and b:

Stripline (left), Microstrip (right)
A major piece of the transmission path through a PCB that has a dominant effect on
high-speed signals is the Plated-Through-Hole (PTH) via. PTH vias are means of transporting
signals into interior layers of a multi layer PC board. Depending on how the signal is routed
through the PTH via, the PTH can become very 'visible' to signals with increased frequency
content. The major source of impedance discontinuity in the PTH via occurs when the signal
is routed through the via in such a way that a portion of the via is not used for signal
transmission. This extra portion of the via is sometimes called the 'via stub'. Back drilling is a
technique used to remove this extra via stub, but is expensive and avoided in low cost
applications thus forgoing signal integrity at higher frequencies.
Chapter 5 will present an in depth analysis of the effect of the PTH via and via stub on
signal transmission and signal coupling in the PCB. The analysis is based on 3D
electromagnetic field simulations of 3 differential signal nets in the PCB via escape region. In
order to fully appreciate the simulation results, in the following chapter we will discuss the
physical effects that cause degradation of signal integrity.
Chapter 3 : Signal Integrity of High-Speed links
In order to improve the performance of high-speed links, we must have a good

understanding of the physical effects that cause the signal degradation we wish to avoid. In
this chapter, we present the challenges of attaining signal integrity at higher frequencies in
general interconnects. These tools will be used in our analysis of key performance limiting
areas in the chip-to-chip high-speed channel design in subsequent chapters.
Higher signaling rates and faster edge rates or rise times (defined as the time required
for the signal to change from %10 to %90 of its final level) require increased high frequency
content transmission through the channel. Given this high frequency content, the channels
can no longer be considered as ideal interconnects and frequency dependant losses and
signal interference must be taken into account.
The signal integrity challenges can be divided into the four categories of signal loss,
reflection, crosstalk, and power integrity issues. Signal loss and reflections will prevent the
transmitted signal from reaching its destination with full power while crosstalk and power
supply noise can cause the received signal to become noisy and unreliable.
3.1 Transmission Line Losses
As the frequency of operation increases, and the wavelength (X=c/f, c is the speed of
light, f is the frequency of operation) of the highest frequency component of the signal
approaches the circuit's physical size, the signal travelling through the interconnect is no
longer a simple digital signal going through a lumped Resistor Capacitor (RC) network, it
begins to behave like a wave propagating along a transmission line. Transmission line
behavior typically starts to appear when physicals lengths of the components are larger
than one hundredth the length of the wavelength. This transmission line is modeled as a
distributed series of inductors, capacitors, and resistors. The transmission line is defined by
its characteristic impedance, propagation delay and losses [23].
In modern interconnects smaller dimensions and high-frequency content cause the

resistive losses in the transmission line to intensify and become ever more important.
Resistive losses affect the performance of the digital system by decreasing the signal
amplitude, thus affecting noise margins and slowing edge rates, which in turn affects timing
margins. We thus consider both the frequency dependant and frequency independent
losses of the transmission line.
3.1.1 Conductor DC Losses

The conductor DC loss is essentially the resistive component of transmission lines
which is due to conductors not being perfect conductors. Based on this, the DC loss
depends primarily on two factors: the resistivity of the conductor and the total area in
which the current is flowing. The resistive loss is given by
pL pL
R= = (3.1)
A Wt
where R, in ohms (0), is the total resistance of the line, p is the resistivity of the conductor
material in ohm-meters, L the length of the line, W the width, t the thickness, and A the
cross-sectional area of the signal conductor.
3.1.2 Dielectric DC Losses

Dielectric materials used in different components of the channel are not perfect
insulators and thus there is a DC loss associated with the resistive drop across the dielectric
material between the signal conductor and the reference plane. The dielectric dc loss is,
however, negligible for conventional dielectrics and can thus be ignored [19].
3.1.3 Frequency-Dependant Conductor Losses

At low frequencies the DC loss is sufficient for accurate system simulation; however,
as the frequency increases other phenomena that vary with the spectral content of the
28
digital signals begin to dominate. The most prominent of these frequency-dependant
variables isthe skin effect.
3.1.3.1 Frequency-Dependant Resistance and Inductance

Skin effect manifests itself primarily as resistance and inductance variations. At higher
frequencies, the current tends to flow closer to the surface of the conductors instead of
flowing uniformly throughout the conductor cross section. The measure of the distance
over which the current falls to 1/e of its original value is termed the skin depth [23]. The skin
depth delta is calculated with the equation
= p= meters (3.2)
where p is the resistivity of the metal, w the angular frequency (2ntF), and p the
permeability of free space (in henries per meter). From equation 3.2 it can be seen that as
the frequency increases, the current will be confined to a smaller area, which will increase
the resistance, as seen in the DC resistance equation 3.1.
The total inductance of a conductor is caused by magnetic flux produced by the

current. The total inductance can be divided into internal inductance caused by current
flowing through the center of the wire and external inductance caused by the current
flowing through the skin depth of the conductor. At high frequencies the internal
inductance approaches zero as the current through the center of the wire disappears. Thus
for high-speed digital systems, internal inductance is safely ignored and the external
inductance is used as the conductor inductance.
3.1.3.2 Effect of Conductor Surface Roughness

In reality, due to manufacturing limitations, metal surfaces are rough and this
effectively increases the resistance of the material especially at higher frequencies when
the mean surface roughness is a significant percentage of the skin depth. It has been
empirically shown that surface roughness can account for up to 10-50% increase in loss
compared to the ideal smooth surface case [19]. Figure 3.1 is a cross section of a stripline
displaying surface roughness. In this particular case the surface roughness on the top of the
conductor is approximately 0.2mils (5 pm).
Figure 3.1: Cross section of a stripline in a typical PCB, showing surface roughness.
The roughness of the conductor is dubbed as the tooth structure and the magnitude of
the surface variations is described as tooth size. For example, the measured tooth size in
Figure 3.1 is 5 pm. PCB vendors indicate that typical FR4 boards have a tooth size of 4 to 7
pm.
3.1.4 Frequency-Dependant Dielectric Losses

When a time-varying electric field is impressed onto a material, any molecules in the
material that are polar in nature will tend to align in the direction opposite that of the
applied field. This is termed electric polarization. The classical model for dielectric losses,
which was inspired by experimental measurement, involves an oscillating system of
molecular particles, in which the response to the applied electric fields involves damping
mechanisms that change with frequency [24]. Any frequency variations in the dielectric loss
are caused by these mechanisms.
Frequency-dependant dielectric loss increases linearly with signal frequency:
aD = r f tan 6 (3.3)
where tan 6 is the loss tangent, c is the speed of light and Er is the relative permittivity [22].
Dielectric loss is typically specified with the loss tangent which strongly depends on
the type of insulator material. Table 3.1 shows the loss tangent for some typical PCB
insulators.
Table 3.1: Loss Tangent for Typical PCB insulator.
Material FR4 Polyamide GETEK Rogers4350 Teflon

Loss Tangent 0.035 0.025 0.01 0.004 0.001
Most legacy backplanes use FR4 material, which has the highest loss tangent. Newer
and higher performance backplanes use either Rogers or one of the NELCO materials with
lower loss.
The linear dependence of dielectric loss to frequency causes it to dominate over the
skin-effect at very high frequencies. The crossover frequency of skin effect and dielectric
loss depends on the material properties and dimensions of the trace. Both effects are
illustrated in Figure 3.2 for an FR4 material.
0.8
0
S0.6
S0.4
0.2
0
1.OE+06 1.0E+07 1.OE+08 1.OE+09 1.0E+10
Frequency, Hz
Figure 3.2: Crossover between skin-effect and dielectric loss, FR4 8 mil wide and 1 m long 50
ohm stripline [25]
3.2 Reflections (Impedance Discontinuities and Stubs)
When a light wave passes to a medium different from the original one, some of the
signal gets reflected back to the original medium due to the difference in the refractive
indexes of the two mediums. The same behavior occurs with EM waves passing from one
material to a different one, some of the incident wave's energy passes through the line but
the remaining portion is reflected. The amount of reflection depends on the characteristic
impedance of the transmission line. The characteristic impedance of a lossy transmission
line is given by:
R + jowL
Zo= G+i Ohms (3.4)
where R is the series resistance of the inductor, G is the parallel conductance of the
capacitor, and w is the radial frequency. The reflection coefficient is thus dependent on
frequency when Rand G are not significantly small to ignore.
Z, Z,
Figure 3.3: Transmission line system with multiple line impedance.
If the characteristic impedance of the transmission line matches that of the load or
source, the incident wave is delivered without reflection. Otherwise the mismatch in
characteristic impedance will cause reflections at the mismatch boundary. The Reflection
Coefficient describes the amount of reflection at the boundary where the characteristic
impedance seen by a traveling wave changes causing an impedance discontinuity. For
example, the reflection coefficient at the junction of two transmission line segments with
impedances Z1 and Z2, as seen in Figure 3.3, is given by:
2- z
r= (3.5)
Z2 + Z1
Characteristic impedances of different components are designed to a consistent value when
possible (typically 50 ohms is used for single ended traces and 100 ohm for differential
traces). However, manufacturing variability in the fabrication of the traces results in
variation in characteristic impedance which is beyond the designer's control. In addition,
the unavoidable impedances of the certain components in the channel produce potentially
large impedance discontinuities in the interconnect structure that can severely degrade
performance.
Other sources of reflections are via stubs which have stronger frequency dependence.
Here the stub acts as a capacitor which reflects high frequency energy. Back-drilling the via
stub can help alleviate this problem in high performance designs but is not desirable for low
cost applications.
3.3 Crosstalk (FEXT, NEXT)
The coupling of energy from one line to another called crosstalk occurs when the
electromagnetic fields from different signal nets interact. Crosstalk is very widespread in
digital designs, occurring on the chip, the PCB board, connectors, the chip package, and on
cables with the strongest crosstalk occurring between the signal lines in chip packages and
connectors [26]. Moreover, as applications demand physically smaller and faster products,
the amount of crosstalk in digital systems is increasing dramatically.
Excessive line to line coupling, or crosstalk, can have two detrimental effects: First,
crosstalk will induce noise onto other lines, which may further degrade the signal integrity
and reduce noise margins. Second, and as a less impacting effect, it can change the
performance of the transmission line in a bus by modifying the effective characteristic
impedance and propagation velocity, which in turn affect the system-level timing and the
integrity of the signal. Due to these consequences of crosstalk, system performance
depends heavily on data patterns, line-to-line spacing, and switching rates [19]. We now
look at the mechanisms that cause crosstalk.
3.3.1 Mutual Inductance and Mutual Capacitance

Mutual inductance is one of two ways crosstalk is produced. Mutual inductance Lm
induces current from an aggressor line onto the victim line by means of the magnetic field.
Mutual inductance Lm injects a voltage noise onto the victim proportional to the rate of
change of the current on the aggressor line. The magnitude of this noise is calculated as
VsLm = dlaggressor (3.6)

noise L=m m dt
The induced noise is proportional to the rate of change, and thus the mutual
inductance becomes more significant in high-speed digital applications.
Mutual capacitance is the other mechanism that causes crosstalk. Mutual capacitance
is the coupling of the two conductors via the electric field. Mutual capacitance Cm injects
current onto the victim line proportional to the rate of change of voltage on the aggressor
line:
(3.7)
nose,Cm
noise,cm =
m dVaggressor
Cm
dt
Thus, mutual capacitance also becomes significant in high-speed digital applications [19].
3.3.2 Types of Crosstalk

Crosstalk can be divided into far-end (FEXT) and near-end (NEXT) crosstalk. In FEXT,
the aggressor signal travels in the same direction as the victim signal. NEXT, on the other
hand, is the coupling of the aggressor signal to the victim traveling in the opposite direction,
and is in general much more critical since the strong aggressor signal can couple into an
attenuated victim signal in the connector or package located on the receiver side where the
signal is already attenuated. Since crosstalk is caused by either capacitive or inductive
coupling of two signal lines (more so inductive in modern dense connectors), it is relatively
less at lower frequencies. Due to the low-pass nature of the channel, FEXT is also
34
attenuated at higher frequencies but still remains significant as the signal itself will be
attenuated at the receiving end. As an example, Figure 3.4 depicts the FEXT and NEXT
crosstalk in a modern package trace and cable.
Figure 3.4: Example NEXT (blue) and FEXT (red) in a modern package (left) and a modern 1
meter cable (right)
3.4 Power Integrity (digital, analog)
Figure 3.5 shows a simplified circuit of the power supplied to High-Speed Serdes (HSS)
cores. The current demand at the chip must be supplied through the power nets and from
the 'ideal' voltage regulator on the motherboard. Voltage drops due to the parasitic
resistance in the power nets as well as voltage noise due to parasitic inductance of the
power nets is of major concern for power integrity which in turn leads to signal integrity
[27]. Power supply circuitry for digital circuits generally employs decoupling capacitors
placed in various sections of the power nets to alleviate noise caused by parasitic elements.
Board-Side Caps (BSC) on the motherboard near the Voltage Regulator (VR), Motherboard
caps (MBC) in the socket cavity on the board, Land-Side Caps (LSC) on the cavity in the
BGA/LGA side of the package, Die-Side Caps (DSC) on the top side of the package,
embedded capacitors between the package layers, and on-chip capacitance in the chip are
examples of decoupling capacitors used to alleviate voltage noise to provide a steady
voltage for the circuitry in the silicon at different frequency ranges [28]. Many algorithms
have been developed for efficient decoupling capacitor (de-cap) placement in various
regions of the power net [29].
I~ A_ I_ _
(Chin - Sin nl, rnir racKa ]e model
...
...............
...........
.. .............
................
b V_DC
V_SRC
Vdc=1.5 V
1I
T -
NETT
On-Chip Ground Ideal chip capacitance

Chip current demand waveform
Figure 3.5: Power supply circuit including package parasitic model
Power supplied to analog circuits has thus far not been of much concern due to the
constant power supply nature of Current-Mode Logic (CML) circuits used in HSS cores.
However, the increase in CMOS logic used in HSS cores, to achieve power saving, is
changing the nature of analog power net design in the package as well as in various other
components. Analog power nets must account for voltage noise caused by the increased
current fluctuations to assure a steady voltage supply. Figure 3.6 displays a current demand
waveform for HSS cores along with the associated voltage supplied through analog power
nets in the package.
.0
a.
Figure 3.6: Chip current demand waveform (left) along with the corresponding voltage
supplied through the package analog power nets with reference to on-chip ground.
Figure 3.6 depicts a 60 mV voltage fluctuation due to the package parasitic behavior.
Assuming that this is within the tolerance range of the analog circuits and taking into
account other noise caused by chip and board parasitic behavior, this package net design
can be appropriate for the given current demand waveform, otherwise, more careful design
must be performed for the analog power nets in the package.
This analog net power integrity analysis example demonstrates how power net
parasitic effects can translate into high voltage fluctuations at the circuit resulting in
undesirable signal transmission and thus loss of signal integrity.
In the remaining portion of this thesis we concentrate on challenges in high-speed

channel design, caused by factors outlined in this chapter, that contribute to signal integrity.
Chapter 4 : Discontinuity Cancellation at the Package Level
This chapter discusses a package design technique to enhance high-speed signal

performance by reducing the large impedance discontinuity effects of various regions in the
package such as vias and solder balls. The technique involves inserting an intentional
counter-discontinuity in complementary phase to the existing discontinuity in order to
mitigate this discontinuity. In the next sections, the transmission line behavior of short
multiple discontinuities are analyzed using theoretical approximation and simulation
examples to demonstrate the validity of the mentioned technique. This broadband
impedance matching technique is then applied to package via and solder ball transitions of
high-speed differential nets using 3D simulation to evaluate improvement at target
frequencies and its impact on bandwidth.
Discontinuities in high speed package interconnects are having more significant

impacts as data rates have increased beyond 10Gbps. Vias in thick laminates cores and
solder balls at the package to PCB interface introduce large capacitive discontinuities
(Figure 4.1), while bond wires in wire-bond packages introduce large inductive
discontinuities. In such cases, the behavior of the structure is inherent to the package and
the designer has limited options to control the impedance given the package technology.
The variables available to the designer are in the dependant on the packaging technology
but may include trace width, via to via pitch, via plane opening, and BGA plane opening,
which we will consider in our FBGA 3D model simulation.
Ilder Balls; highly
e
I00
mpedance-controlled
u;4 - Chinlp Pads

Package Design
highly capacitive.
Figure 4.1: High-speed net in a flip-chip package. Solder balls are
4.1 The Counter-Discontinuity Cancellation Technique Description

intentional
In an effort to decrease capacitive discontinuity in the solder ball path,
signal nets have
voids in the ground conductor layers above the solder balls of high speed
ground conductor
been used in recent package design practices [30]. While voiding the
can interfere with
helps reduce capacitance and hence the discontinuity significantly, it
routing space
neighboring signal structures and reduces either the power plane or signal
both also detrimental to signal integrity.
In this chapter we look at an alternative approach for obtaining further improvement
transmission. The technique uses transmission line behavior in short

in signal
an inductive
discontinuities. For example, if the existing discontinuity is capacitive,
to achieve
discontinuity is introduced to compensate for the capacitive discontinuity
inductance
matching impedance as seen by the traveling wave over the short distance. This
devices to
insertion has been used in chip termination of high speed circuits and ESD
compensate for the large parasitic capacitance of chip circuits and pads [31, 32].
40
Contrary to narrowband impedance matching with lumped elements which is used for
RF narrowband applications, insertion of transmission line discontinuities allows for
broadband impedance matching which is desirable for the broadband high-speed
communication we are studying. In addition, we take advantage of the current structure of
the signal net to insert transmission line discontinuities which is much simpler and cost
effective than lumped element matching for the package.
In the next sections, transmission line behavior with short discontinuities is revisited
from a distributed line behavior and a lumped element circuit point of view. Package design
structures in via-solder ball transition are studied for discontinuity compensation and its
effectiveness is discussed using 3D EM (electromagnetic) simulations.
4.2 Theory: Distributed Network Behavior of Short Discontinuities
Consider the lossless transmission line with a short discontinuity of I << X/10 in Figure
4.2. The transmission line is terminated at both ends with matching impedance to Zo.
Assuming negligible phase shift with I << X/10, the transmitted wave arriving at the end of
Z1, at t=O, TO, can be approximated as (1-F12) [33]. For example, 81% of the incident wave
will be transmitted at t=O if Zo=500 and Z1=200.
Total transmission at the line end of Zj: T=TO+T1+T2+...

TO: transmission at tO
T1: transmission of reflected wave after first round trip in
line discontinuity Z1
ST1 T2: transmission of reflected wave after second round trip
) at t=O, (1+T1)(1-F1 )=(1-F' 2)

at t=tl, (1+I 1)(-f 1)( -F1)(1-F 1)=F1 2(1-F 12 )
ST2 at t=t2, (1+f 1)(-F1)(-I7)( -F1 )( -F1) (1-F1)=F14(1- 12 )
Figure 4.2: Illustration of transmission and reflection on a transmission line with a

discontinuity
Next consider two transmission line configurations in Figure 4.3 where an extra
discontinuity Z2 is purposely inserted at the end(s) of the existing discontinuity Z1.
Zo Z, Z2 Zo Zo Z2 Z, Z2 Zo
F -F 9 -IF r 3
F1 -F- -rIF -F2 r F2 -rFI- -F.
p TO j TO
----- T----------------+=----------------------TO=
T=TO+T1+T2+... T=TO+T1+T2+...
2 2
TO=(1+FT)(1+F 2 )(1+F 3 ) TO=(1-F 3 )(1-F2 )(1+ 2 )(1+ 3 )=(1-F 2 )(1- 3)
(a) (b)
Figure 4.3: Transmission lines with an extra discontinuity inserted (a) at one discontinuity
end and (b) at both discontinuity ends of existing discontinuity Z1
A traveling wave encountering physical discontinuity experiences a gradual impedance

transition rather than abrupt impedance change as the field interacts with structures in
both sides of the discontinuity. Impedance seen by the wave is observed as an integrated
characteristic of line segments alongside the discontinuity. Transmission line behavior can
be seen more intuitively using lumped circuit elements if the line length is considerably
smaller than the wavelength. Consider a transmission line with a discontinuity at x=O as in
Figure 4.4. Assuming a lossless transmission line (G=O, R=O), the characteristic impedances
are characterized as F and , respectively. Traveling waves entering at x=O will
experience an integrated characteristic of two discontinuities. Assume line impedance 20Q

with LI=40pH and C1=O.lpF at x<O, and line impedance 125fl with L2=250pH and C2=16fF at
x>O in Figure 4.4. Impedance at the discontinuity would be seen as 500 with L=290pH and
C=116fF near the discontinuity x=O, which effectively offsets each other's large
discontinuity.
Ri,L,,Gi,Ci: transmission line parameters per unit length
Figure 4.4: Distributed transmission line with a discontinuity represented by lumped

elements
The configuration of Figure 4.3 (b) is used in the following analysis to further look into
the behavior of package design options. Simulations of transmission line configuration in
Figure 4.3 (b) were conducted using lossy transmission lines in Agilent ADS with the
discontinuity Z2 varying in length and impedance while maintaining the discontinuity length
of Zi and overall transmission line length remain at 1mm and 5mm. As seen in Figure 4.5,
improved transmission is observed with higher impedance of inserted discontinuity. This
suggests that inserted inductive discontinuity and existing capacitive discontinuity
effectively offset each other over a wide frequency bandwidth. As observed in Figure 4.5,
better insertion loss can be achieved with higher impedance and longer discontinuity lines
in the lower frequency range but this improvement dies out quickly at the higher frequency
range. Moreover, wider bandwidth is achieved with relatively lower impedance and short
discontinuity. While shorter discontinuity lines such as 0.3mm and 0.5mm can be more
efficient in performance and bandwidth as observed in Figure 4.5, much higher impedance
and hence larger inductance in the interconnect is required to achieve the target
improvement and may not be realizable in practical manufacturing space.
4 5mm
Port 1 Zo=500 Z2 Z1=20 Z2 Zo=50 Port 2
(2mm -1 2) 2 1mm 12 (2mm -1 2)
(a) 12=0.3mm [GHz]
[dB
(b) /2=0.5mm [GHz] (c)I,=lmm [GHz]
Figure 4.5: Insertion loss, S21, simulated using lossy ideal transmission lines with varying
line impedances of (a) 0.3mm, (b) 0.5mm, and (c) 1mm for inserted counter-discontinuity
Z2. Z2=50Q represents no discontinuity insertion
4.3 3D Package Structure Simulations with Discontinuity

Improvement
A 3D model of laminate core vias and solder balls of a differential pair was simulated
to study the effectiveness of the discontinuity cancellation technique in a 4-layer Enhanced
Plastic Ball Grid Array (EPBGA) wire-bond package as shown in Figure 4.6 where Plated-
Through-Hole (PTH) vias are connected from the top to bottom layers including the middle
core layer. Part of the package and board traces at escape edges of balls and vias were
designed with higher impedance, 700 and 90M, than the target matching impedance to
offer counter-discontinuity to the capacitive discontinuity of the solder balls and vias while
maintaining an overall 2mm of lossy traces. It is observed in Figure 4.6 that higher gain was
achieved with higher line impedance and longer discontinuity length but at the cost of
further limited bandwidth consistent with the transmission line behavior observed in Figure
4.5. Compared to the simulation results using transmission lines, insertion loss bandwidth of
this package structure is observed to decrease more quickly with longer discontinuity with
little gain at lower frequencies.
100pm pre-prel
150um core
100apm pre-prel
vias
Balls
[GHz]
Figure 4.6: Insertion loss of a differential signal, SDD21, through solder balls of 1mm pitch
and PTH vias in a 4-layer EPBGA package where package and board traces are designed with
(a) 909 2mm, (b) 70M 2mm, (c) 90Q 1mm combined with 5092 1mm, (d) 700 1mm with
50. 1mm, (e) 902 0.5mm with 50Q2 1.5mm, (f) 7092 0.5mm with 50ohm 1.5mm and (g)
5092 2mm which is a nominal design structure with no discontinuity insertion, that is the
basis for comparing the other cases.
A further design experiment was conducted with a 4-layer Fine Ball Grid Array (FBGA)
package stack-up where blind vias are used in the outer thin pre-preg layers. With more
design flexibility available using blind vias, performance improvement of inserting inductive
discontinuities using the following three different areas of the structure was analyzed as is
shown in Figure 4.7: escaping trace impedance on package/board, via transition, and plane
void above balls and via clearance. Table 4.1 summarizes the different variables available in
the FBGA package for impedance matching as well as the effect on the characteristic
impedance of the component as the variable is increased in size.
Table 4.1: Design variables considered for opposite phase impedance insertion and their
effect on the characteristic impedance
Feature Size Characteristic Impedance

Trace width ,
Trace length 1 No Change
Via-via trace width _
Via-via pitch t No Change
Plane void above ball 1 t
Via clearance size 1 1
It is observed that insertion loss and bandwidth vary over a wide range depending on
whether inductive counter-discontinuity is introduced or not. Structure (a) of Figure 4.7
represents a typical design of solder ball-via transition with no plane voids, minimum anti-
via diameter and escaping traces matching to termination. Inductive discontinuities in (b)
and (c) of Figure 4.7 are implemented as part of the escaping trace and at via to via
connecting traces, respectively. In structure (f), inductive discontinuity of escaping trace is
combined with plane voids and larger via clearance. While voids on plane layers above balls
as in (d) and (e) of Figure 4.7 offer the most desirable improvement and bandwidth gain by
reducing existing capacitive discontinuity, those voids reduce routing space and are not
always feasible to implement without interfering with neighboring signal structures.
Although inductive discontinuities added to escaping traces and via to via traces as in (b)
and (c) are not as effective as plane voids with limited bandwidth, those still provide a very
practical solution with improved performance up to 11 to 16GHz, a suitable upper range for
many applications. Combining inductive trace design, plane voids, and larger via clearance
as shown in (f)yields further performance improvement.
Op1m pre-preg
G
100um core
/ 40pm pre-preg
52
(e) S2
(d)
Balls
Board
Ztrace Via to via Plane void Anti-via

I\ rn rn aritive Nnnie 4mnnm
Z (b) 80-1mm capacitive None 4400im

ia to via trace (c) 50 inductive None 400m
(d) 50Q capacitive 4)600m on P 4400gm
"Plane void (e) 50Q capacitive 4600pm on G,P 4 400gm
AI\ Ql_-l1mm rn~nwiithim 1 flnnllm nn P A fnnllm
pI
lV VV I VI I VVVp,
- Trace '-l ll
Figure 4.7: Insertion loss of a differential signal, SDD21, through solder balls and PTH/blind
vias in 4 layer FBGA package varying plane voids above balls, via to via traces and routing
trace impedance on package and board. Structure (a) represents a typical package design
with no plane voids, minimum anti-via diameter and escaping traces of matching
impedance and is used as a base structure to measure performance improvement of other
design options.
4.4 Time Domain System Analysis
A technique to improve the performance of a high-speed links in the package by

demonstrating the improvement in insertion loss of the package net for high-speed
differential signaling has been presented. However, the effects of the given improvement
must be analyzed in light of the system as a whole to be able to fully quantify the
improvement in the channel performance. In this section, the package component along
with discontinuity cancellation optimization is incorporated in a typical channel scenario for
system level analysis. Since ultimately signals are measured at the receiver end in time-
domain, time-domain eye diagram simulations are performed to quantify the
improvements investigated in this chapter.

Note that modern high-speed channels operate at fast enough rates that require
active channel equalization in order to correctly decipher transmitted bits at the receiver
end. However, channel equalization can mask over the component level improvements we
wish to observe in the channel and as such we purposely avoid any equalization in our eye
diagram simulations. We use High-Speed Signal Clock Data Recovery (HSSCDR), an internal
IBM tool, to perform the eye diagram simulations in this section. A 6.2 Giga bits per second
(Gbps) and an 11Gbps high-speed transmitter (TX) and receiver (RX) circuit, based on an in
production IBM circuit design, is used to perform the simulations. As seen in Figure 4.9, the
simulated channel consisted of the TX and RX chips and a symmetrical structure with
package models connected to daughter PCB stripline traces with PTH vias subsequently
leading to high-speed connectors and finally connecting to the motherboard stripline trace
via short stubbed PTHs. A 1 million bit long, Pseudo-Random-Bit-Sequence (PRBS) is used
for the simulation.
Transmitter Chip (TX) Receiver Chip (RX)
18" PCB Trace-
Figure 4.8: High-speed channel used for time-domain system simulation
Table 4.2 depicts the average height of the eye at the receiving end of the channel.
The width is not displayed in the table since the coupling between the signals was not taken
into account in the time-domain simulations. As table 4.2 shows, at both 6.2Gbps and
11Gbps transfer rates, a significant eye height improvement is seen at the receiving eye. In
the case of the 6.2Gbps transfer rate, the average eye height for the base and optimized
cases are 158mV and 164mV respectively, therefore obtaining a 3.8% improvement over
the base case with the discontinuity cancellation. For the 11Gbps rate while the worst case
eye was closed in both cases due to the lossy PCB trace, the base and optimized cases
produce a 161mV and 168mV average eye height respectively, in other words generating a
4.3% improvement over the base case due to the discontinuity optimization.
Table 4.2: Time-Domain simulation of discontinuity cancellation for a typical chip-to-chip

channel
Package Ave. Eye Height (6.2 Gbps) Ave. Eye Height (11.0 Gbps)
Base Case 158 mV 161 mV
Discontinuity Optimized Case 164 mV 168 mV
The time-domain results of the system analysis highlight the significance of the
performance improvement due to the discontinuity cancellation. This technique is
especially valuable given that the cost of implementing it in the package is minimal and
simply requires knowledge of the channel discontinuities at the frequencies of interest.
Chapter 5 : PCB Plated-Through-Hole (PTH) Via Design
A critical region in a high-speed chip-to-chip channel is the escape region from the
package to the PCB traces. The high density of signals transmitted through the Ball Grid
Array (BGA) structure of the package require more than a single PCB layer for signal routing
and thus the use of vias and more specifically PTH vias. PTH vias are placed under the pad
array of the package and transport signals into interior layers of the multi layer PC board. As
the frequency content of signals increases to produce sharp rise and fall times for the digital
pulse, the PTH via portion of the PCB becomes visible to the signal and signal loss and
coupling in the vias becomes of major concern.
The fundamental frequency of the spectrum in Non Return to Zero (NRZ) signaling
occurs at one half the data rate. The fundamental frequency of a 6.25Gbps signal, for
example, would occur at 3.125GHz with some contributions from the third (9.38 GHz) and
fifth (15.25 GHz) harmonics. The PTH via can behave as a parasitic element or a
transmission line discontinuity based on the data rate of operation, as we saw in chapter 4.
It acts as a notch filter centered around a frequency primarily determined by the unused
portion of the via, henceforth referred to as the stub. Due to the stub resonance, signals
traveling through the via experience major reflections at frequencies near the notch filter
frequency [34, 35].
High-speed differential signaling greatly improves high-speed signaling especially

shielding it from excessive noise and coupling from nearby signals. This is because the
receiver only sees the relative voltage or voltage difference between the two transmission
lines in the differential pair. The inherently superior noise immunity and high gain of
differential circuits makes them less sensitive to the attenuation of the signal in the
transmission medium, thus allowing for signal levels to be reduced with consequent
improvements in switching speed, power dissipation and noise [36]. We thus analyze the
performance and coupling of high-speed differential signaling in the PCB via escape region
through 3D electromagnetic simulations performed by Ansoft's HFSS (High Frequency
51
Structure Simulator) [37]. The structure is a 28 layer PCB with 3 differential signal pairs
considered in different coupling configurations.
5.1 PCB PTH-Via Structure Setup
The PCB modeled is a 28 layer PCB with copper planes of 0.6mil thickness each (Figure
5.1). The dielectric material used is FR4 with Er = 4.2, which is widely used in
Incoming trace PTH-via Signal Layer Number

A!
Ut ~
L12
i I
~-~xii-~-~~ ~~~~--~- i--~Y3
L11
i L10
1i I L.9
d_
L.8
--i ;--i L.7
j i.. [.6
L
-i ---
rt--------- L5
s,
; L
L4
-3
cr~
L2
~-------i
b, 3 L1
--
0-11- r~
Figure 5.1: Modeled 16 layer PCB structure on the right displaying the via traces in red;
Dimensions and structure of a differential via pair on the left.
the PCB industry, and the thickness of each of the dielectric layers is 10.2mils (with the
exception of the core layer which is 4.8 mils thick). The PTH via diameter is 8mils with a via-
to-via pitch of 40mils (about 1mm), corresponding to the pitch of the BGA solder-balls on
the package. The via-pad diameter is 21mils and each GND/VDD has a via clearance with
diameter 30mils. In addition, each via is directly connected to a BGA pad of diameter 21
mils, required for structural stability.
In a practical application, the signal is routed from the BGA pad through the via and
onto a different PCB plane. We model the signal being routed to different layers in the PCB
which results in different via and via-stub lengths. The traces are used for measurement
purposes and their effect is removed after simulation by de-embedding them in a post-
process. Table 5.1 displays the different layers in the PCB that a signal can travel through
along with the corresponding via and stub lengths. In addition to the models with via stubs,
structures with the via stub back-drilled are also constructed and simulated to compare
with the stubbed models. In the back-drilled models, the signal escape layers in the PCB are
similar to those in Table 5.1 with the stub length corresponding to the back-drill length.
Table 5.1: Stubbed PCB Via Configuration

Escape Level Stub (mils) Via (mils) Stub A/4 Resonance Freq (GHz)
12 140.7 16.8 10.0
11 129.8 27.7 10.8
10 119.0 38.5 11.8
9 108.2 49.3 13.0
8 97.4 60.1 14.5
7 86.6 70.9 16.3
6 70.3 87.2 20.0
5 59.5 98.0 23.7
4 48.7 108.8 28.9
3 37.9 119.6 37.2
2 27.1 130.4 52.0
1 16.2 141.3 86.7
Three different via configuration models are constructed to study the effect of
differential coupling among the vias in the PCB escape region. As seen in Figure 5.2 and
Figure 5.3, these models are boxed, staggered and diagonal configurations.
Boxed Vias Staggered Vias Diagonal Vias
O0 0000 000000
BGAAssignment: 00 0000 000000
00 0000
Maximum compactness: Medium compactness: some Minimum compactness:

high coupling between vias coupling between vias vias farthest apart
Figure 5.2: Via configurations with different signal densities.
The different power planes as well as the via clearance on every GND/VDD plane is depicted
in the 3D models in Figure 5.3. One will also note that except for one via pair in the
staggered configuration, all incoming and escape traces are routed on the same side of the
model. This allows for consistency in the return path current of the different models.
q
Figure 5.3: Top and 3D view of the Box, Diagonal, and Staggered via configurations
5.2 Signal Transmission
Impedance discontinuities are the main causes of signals not being transmitted
through the PCB escape region. The BGA pad, via pad, and the via stub are sources for
capacitive discontinuity while the via itself contributes both inductive and capacitive
impedances to the signal path. Capacitive impedance arises from the proximity of the signal
paths to power and ground planes. In our design, the via pad capacitance is negligible
compared to the BGA pad due to the via clearances that remove the GND/VDD planes from
underneath the via pad. The capacitive coupling between the BGA pad and ground,
however, causes the BGA pad to be a highly capacitive region of the signal path and thus
creating an impedance discontinuity that is lower than the matched 100 0 characteristic
impedance.
The via parasitic is also significant especially given that the length of the via is much
longer than the BGA pad. Every via has parasitic capacitance to ground. Moreover, PTH vias
are electrically long structures relative to the interested frequencies of operation and thus
behave like transmission lines [38]. The minimum transit time, tv, through the stub structure
can be approximated by:
t- (5.1)
C
Thus, in the given PCB setup, equation 5.1 evaluates to 0.027ns.
The potentially more impacting discontinuity is the via stub. The via stub acts as an un-
terminated transmission line, generating a quarter-wave resonance (a zero) in impedance at
a frequency corresponding to four times tstub. Any signal power near this resonance
frequency will be subsequently attenuated by the via stub [39].
The results of the PCB via simulations with the stub back-drilled display two distinct
behaviors for relatively shorter and longer vias. The cut-off for these two groups is
approximately at half the PCB length, thus we group the vias escaping in layer 1-6 as long
vias and layers 7-12 as short vias. As shown in Figure 5.4, in the long via region, increased
via lengths contribute to less insertion loss and thus degraded performance at higher
decrease the
frequencies. For the short vias, however, we find that shorter vias tend to
insertion loss, as seen in Figure 5.5. It is also noted that in all the PCB models of vias with
the stubs back-drilled, via length is almost irrelevant for insertion loss at frequencies bellow
8 GHz.
0 2 4 6 8 10 12 14 16 18 20
freq, GHz
Figure 5.4: Insertion loss for differential vias with back-drilled stubs (longer via lengths)
[d
freq, GHz
Figure 5.5: Insertion loss for differential vias with back-drilled stubs (shorter via lengths)
In order to characterize the signal net structure, we perform Time Domain

Reflectometry (TDR) simulations. TDR refers to measuring a system by injecting a signal and
observing the reflections back at the source end. Using this method, the impedance
discontinuities in the net can be observed. Figure 5.6 depicts the TDR simulation waveform
of the PCB via structure with different via lengths. The key observation is the discontinuity
difference between the long via in Layer 12 and the short via in layer 1. The longer via
displays a smaller capacitive discontinuity and a sudden increase in the impedance in the
center of the waveform. This is due to the relatively larger impedance of the longer via
which produces impedance cancellation of the capacitive impedance present in the
structure.
1
[Oh
Layer 1
Layer 2
Layer 4
Layer 5
Layer 7
Layer 10
Layer 11
Layer 12
740 760 780 800 820 840 860 880 900 920 940
time, psec
Figure 5.6: TDR impedance waveform of different length differential vias
These observations can explain why from the results of our simulations the insertion
loss tends to degrade both for smaller via lengths as well as for longer via lengths. From our
simulations we see that layer 7 via escape produces the best performace for our particular
PCB setup.
The PCB via structures with the stubs produces significantly larger losses due to the
resonance caused by the stub discontinuity. In Figure 5.7, we observe that the resonance
frequency of the vias with different stub lengths occurs at a lower frequency than the ones
predicted by the quarter-wave length theoretical approximation seen in Table 5.1. For
example, the simulated resonance for the 16.2mil stub is 8GHz with the theoretical
approximation being 10GHz; and the simulated resonance for the 59.5mil stub is 18.5GHz
with the theoretical approximation being 23.7GHz. This difference is due to the fact that
theoretical values were calculated assuming the stub inside a dielectric material with
relative permittivity of Er = 4.2. The PCB, however, consists of layers of FR4 material (Er =
4.2) separated by layers of conducting planes which interfere with the electromagnetic
waves passing through the FR4 material, thus creating a different effective dielectric
constant. In addition, the coupling of the vias with conducting planes was assumed to be
negligible while in reality this is not the case and the coupling will reduce the resonance
frequency [23].
[c
0 2 4 6 8 10 12 14 16 18 20
freq, GHz
Figure 5.7: Insertion loss for three differential via pairs with stubs
5.2.1 Time-Domain System Analysis
We now look at the effect of the stub and via lengths in light of the system as a whole
in order to fully quantify the impact on overall channel performance. As in Chapter 4, we
incorporate the PTH-via structure in a typical channel scenario for system level analysis and
simulate the time-domain eye diagram to quantify the PCB-via design variables considered
in this chapter.
Simulations without any channel equalization produced sever eye closure at the
receiver end due to large losses caused by the lossy PTH, thus making any comparison
impossible. Therefore, we employed Feed Forward Equalization (FFE) to obtain an eye
opening at the receiver in order to compare the effect, after equalization, of the PTH stub
and via length. Once again, we use IBM High-Speed Signal Clock Data Recovery (HSSCDR) to
perform the eye diagram simulations in this section. A 6.2Gbps and an 11Gbps high-speed
transmitter (TX) and receiver (RX) circuit, based on an in production IBM circuit design, is
used to perform the simulations. Two channels were considered in our time-domain
simulations: one including the PCB PTH-via variation only at the package interface and a
second worst-case channel which includes the varying PTH-via at the connector/PCB
interface as well. The channels are displayed in Figure 5.8. We used a 1 million bit long,
Pseudo-Random-Bit-Sequence (PRBS) for the simulation.
Transmitter ChiP (TX) Receiver Chip (RX)
V- Package-PCB Interface PTH-via
5" Daughter Board Trace
Connector-PCB Interface PTH-via
' 18" PCB Trace-
Figure 5.8: High-speed channel used for time-domain simulations
Tables 5.2 and 5.3 depict the average height of the eye at the receiving end of the
channel for the 6.2Gbps and 11Gbps transfer rates with FFE. The eye width is not of interest
here since the coupling between the signals was not taken into account in the time-domain
simulations.
It is observed that longer stub lengths translate into larger average eye height closure
with more severe eye closure occurring at the 11Gpbs transfer rate compared to 6.2Gbps. A
141 mil long stub closes the average eye height by an additional 17% in the case of the
channel with the PTH-via inserted at the package while keeping the connector PTH-vias as
low loss vias. Moreover, the worst-case channel amplifies the loss caused by the stub since
more discontinuities are introduced into the channel. The effect of the back-drilled vias is
overshadowed by the equalization and thus the average eye height is similar for the two via
lengths considered. It is also noted that the worst-case channel does not affect the back-
drilled PTH performance since no stub discontinuity exist in these cases.
Table 5.2: 11Gbps time-domain eye simulation with FFE
PCB Layer Via Length (mils) Stub Length (mils) Ave. Eye Height WC-Ave. Eye Height
1 141.3 16.2 63.3% 61.3%
5 98.0 59.5 61.3% 53.3%
12 16.8 140.7 46.3% 31.8%
2 - Back-drilled 130.4 None 63.3% 63.0%
Table 5.3: 6.2Gbps time-domain eye simulation with FFE
PCB Layer Via Length (mils) Stub Length (mils) Ave. Eye Height WC-Ave. Eye Height
1 141.3 16.2 62.3% 61.0%
5 98.0 59.5 61.3% 58.7%
12 16.8 140.7 59.0% 49.7%
As shown in Figure 5.9, the channel insertion loss of the layer 12 via with the longest
stub incurs large losses compared to the shorter subs or no stubs. In addition, the worst
case channel insertion loss can also be seen to be more severe at the same frequencies
compared to the regular channel. Eye diagrams for the layer 12 and layer 1 PTH-vias are
depicted in Figure 5.10.
In
-10
-20 .. ..
--- ,---------
, , ,i , - ---
-30 ..... ..... 1............. 1 ............

... -- - ' .-- ......
-40 ...... •...... .. ....... : . ...... - -----
-50
-60 :....---
.
Layer 1
-70 Layer 5
-80 Layer 2 - Back-drilled
Layer 5 -. Back-drilled
-0 1 2 3 4 5 6 7 8 9 10 11
[Hz] x1 [Hz] x 10
Figure 5.9: Insertion loss plots for channels with different PTH-vias. Left plot: worst case
channel for 11Gbps simulation. Right plot: channel with package PTH-via and 6.2Gbps
transfer rate simulation.
.a . ... . h I eftiva.eyt iURwLitch pff ... -.. s" D. 1.111
It I lI
1.1OA 4 a 42 0 82
Slfun
R 0 '080
"'I
'1
b)
6.8
-06 4 .0.2
' ' ' iI' '
0 02 04
..
606
04
..
L )I
41 48 -0.6 -0.4 -0.2 0
(U
0.2 0.4 0 08 1
Figure 5.10: Eye diagram for 11Gbps simulation of a) Layer 12 PTH-via with worst case
channel, b) Layer 12 PTH-via, and c) Layer 1 PTH-via.
5.3 Signal Coupling
Chapter 3 described how signal coupling is caused by capacitive and inductive coupling
between two conducting signal paths. These parasitic couplings are governed by the PCB
structure and the distance and configuration of the signal nets. Noise coupling is of major
concern in the design of high wiring density packages with high speed SerDes (HSS) signals.
Adequate isolation between high speed data links is a good measure of the link
performance along with the increase in data rates and wiring density. IBM ASIC isolation
requirements for include a-30dB or lower coupling between transmit differential signals and
-35dB or lower coupling between receiving signals up to the fundamental frequency of the
transmission data [40].
The simulated differential via coupling for the three different via configurations
studied suggests that the coupling in the box configuration is higher than the diagonal via
configuration which in turn is higher than the coupling in the staggered case, as depicted in
Figures 5.11-12. This result is counter intuitive given that the distance between the via pairs
is smaller in the staggered case as compared to the diagonal configuration. As we saw
earlier in this section, the staggered PCB configuration contained a via pair with traces
routed on different sides on the PCB model box. This inconsistency between the models
might explain why the staggered performs better than the diagonal configuration especially
that the staggered case optimizes for return current, thus minimizing the coupling with
other nets. However, the effects of the traces into and out of the PTH via were de-
embedded and assuming perfect de-embedding, this should not be the case. Moreover, the
single-ended coupling results don't confirm this de-embedding error.
[(
0 2 4 6 8 10 12 14 16 18 20
freq, GHz
Figure 5.11: Near-End crosstalk (NEXT) for the Box, Staggered, and Diagonal via
configurations
64
0 2 4 6 8 10 12 14 16 18 20
freq, GHz
Figure 5.12: Far-End Crosstalk (FEXT) for the Box, Staggered, and Diagonal via configurations
At frequencies approaching the stub resonance frequency, NEXT differential coupling
in vias with subs is significantly larger than the coupling between vias of the same length yet
with the stubs back-drilled, as depicted in Figure 5.13. In addition, vias with longer stubs
produce larger differential coupling at lower frequencies.
The FEXT differential coupling in stubbed vias, however, is not affected as much as the
NEXT near the resonance frequency of the stub since the stub is a source of reflection and
prevents the signal from reaching the far end on the aggressor net and thus cannot affect
the victim net as significantly. This is shown in Figure 5.14 for the L5 via stub.
[dE
0 2 4 6 8 10 12 14 16 18 20
freq, GHz
Figure 5.13: Differential NEXT for the diagonal configuration for layer 2 back-drilled (Blue),
layer 5 back-drilled (Green), and layer 5 with stub (Red)
[dl
0 2 4 6 8 10 12 14 16 18
freq, GHz
Figure 5.14: Differential FEXT for the diagonal configuration for layer 2 back-drilled (Blue),
layer 5 back-drilled (Green), and layer 5 with stub (Red)
Finally, we see that the amount of NEXT differential coupling increases along with the
via length as area of the vias where electromagnetic waves can interact is increased. This is
shown in Figure 5.15.
0 2 4 6 8 10 12 14 16 18 20
freq, GHz
Figure 5.15: Differential NEXT coupling for different via lengths with back-drilled stubs
Chapter 6 : Conclusion
The package, the PCB, connectors, and cables are all components of a high-speed
channel. Optimization of each of these components is only valuable when significant
improvement is seen in the high-speed channel as a whole. Traditionally in the
microelectronics industry, each component has been designed by a separate team
responsible for optimizing their specific component for the needs of the market and given
the tools and technologies available at the time. As the improvement margins within each
component have reached their limits, co-designing especially at different component
interfaces is becoming more commonplace. A package-PCB co-design, for example, might
result in a less desirable or even lower performance design in the PCB that will facilitate a
solution in the package that allows for overall improved performance in the high-speed
channel system. An additional factor other than improved performance is reduced cost
which is sometimes as significant as performance, in the industry.
In this thesis, we have analyzed critical aspects of high-speed channel design with a
focus on the package and PCB region. A technique for reducing inherent discontinuities at
the package level was studied and performance improvements were obtained at relatively
low cost. In addition, the critical region of package and PCB interface was studied and
system level simulations were performed, in time and frequency domain, in order to allow
for better design rules for package and PCB design.
We have seen that better transmission performance in the package can be achieved
by compensating for inherent discontinuities by inserting a counter-discontinuity of short
electrical length. As the lengths of both the counter-discontinuity and existing
discontinuities is increased, distributed network behavior of the transmission lines
becomes more prominent resulting in reduced bandwidth. Fortunately, the package
dimensions of the laminate core vias and solder ball structures are well within a tenth of a
wavelength of application speeds of interest and hence these techniques for inserting
counter-discontinuities in complementary phase are effective for improving transmission
while maintaining the necessary bandwidth. System level time domain analysis also
confirms the transmission performance improvement gained by this technique. Although
only capacitive discontinuity examples of vias and solder balls were presented, this
technique can be applied to inductive discontinuities such as short bond-wires as well.
Maximum channel performance can only be obtained when package design is

performed in conjunction with other nearby structures in the channel. As such, the
performance of PTH-vias in the package and PCB region was studied for signal transmission
performance and coupling configurations. We have seen that long stubs reduce
performance significantly especially that PTH-vias occur very often and repeatedly in a
typical channel. As back-drilling is a costly operation, utilizing shorter stubbed, albeit long,
PTH vias can help reduce the performance loss due to the stubs significantly. Signal coupling
is also significantly affected by the stub length, and shorter stubs help reduce coupling as
well. NEXT coupling is far more impacted with longer stubs especially at higher frequencies
near the resonance frequency, while FEXT is not impacted as much due to the high loss in
the channel due to the stub.
It has been shown that although vias are capacitive in general, via length increases the
relative inductance of the via and with our PCB structure dimensions even became
inductively dominated at longer lengths, thus providing an optimized length at which
maximum channel transmission occurs due to minimal discontinuity. This, however, is the
case for back-drilled vias with no stub being present. In the stubbed PTH case, the stub
length overshadows any improvement due to optimized via discontinuity.
Coupling between the PTH and the ground layers of the PCB as well as with other
signal nets effects the resonance frequency due to the stub by lowering the resonant
frequency compared to the calculated theoretical quarter-wavelength frequency. In
addition, signal coupling was shown to be directly related to via length, and decreased with
shorter vias. With regards to different BGA configurations, the Box via created the largest
differential coupling compared to Staggered and Diagonal configurations.
As we have seen, channel design at the package level and the larger system level is
largely application dependant. The problem is not only to maximize performance, but to do
so at a reduced cost while meeting the demand of the particular application. The tradeoff
between cost and performance requires exact knowledge of the channel limits and design
implications. For example, back-drilling of stubs is the clear solution to optimizing
performance, yet as we have shown, the impact of short stubs is not significant enough to
warrant the cost of back-drilling for all applications.
Chapter 7 : Future Work
High-speed channel design is a challenging field especially when cost is added to the
equation. Signaling data rates keep increasing and new problems start occurring as these
rates increase. Therefore, further work has to be done to determine new limiting factors at
higher rates. One of these factors is the increased importance of coupling analysis in signal
integrity studies. At higher rates, not only immediate neighboring signals but a
neighborhood of signals needs to be analyzed for coupling effects. Time-domain simulations
need to take coupling into account as well since the eye width is significantly impacted by
this jitter due to coupling.
Channel equalization impacts channel performance in non-linear ways. As both Feed

Forward Equalization (FFE) and Decision Feedback Equalization (DFE) become commonplace
at higher signaling rates, channel design must be studied together with equalization effect
on the design similar to the co-design process of channel components. The challenge is to
optimize for cost by fully utilizing channel equalization.
Finally, a section beyond the scope of this thesis yet crucial to signal integrity is power
integrity of the high-speed circuitry. Current in power rails for analog circuits are switching
at faster rates creating large leakage current and thus large voltage noise due to power net
parasitics. As signaling rates increase, power delivery to the HSS cores for serial
communication must be closely studied to ensure reliable power delivery and thus circuit
performance.
Bibliography
[1] R. R.Schaller, "Moore's Law: Past, Present, and Future," IEEE Spectrum, vol. 34, pp. 52-
59, 1997.
[2] C.Metz, "IP Routers: New Tool for Gigabit Networking," IEEE Internet Computing, vol. 2,
no. 6, Nov.-Dec. 1998, pp. 14-18.
[3] K-Y.K. Chang, S-T. Chuang, N. McKeown and M. Horowitz, "A 50 Gb/s 32x32 CMOS
Crossbar Chip Using Asymmetric Serial Links," IEEE Symposium on VLSI Circuits, June 1999,
pp.19-22.
[4] T.P. Thomas, I.A. Young, "Four-way Processor 800 MT/s Front Side Bus with Ground
Referenced Voltage Source I/O," IEEE Symposium on VLSI Circuits, June 2002, pp. 70 -71.
[5] A. Deutsch, "Electrical characteristics of interconnections for high-performance

systems," Proceedingsof the IEEE, Volume 86, Issue 2, Feb. 1998 Page(s):315 - 357.
[6] T. Beukema et al., "A 6.4-Gb/s CMOS SerDes Core with Feed-Forward and Decision-
Feedback Equalization," IEEE Journalof Solid-State Circuits, Volume 40, Issue 12, Dec. 2005
Page(s):2633 - 2645.
[7] L. Green, "Understanding the Importance of Signal Integrity," IEEE Circuits and Devices
Magazine, Volume 15, Issue 6, Nov. 1999 Page(s):7 - 10.
[8] IBM SerDes. [Online] [Cited: March 10, 2008.]

(http://www-306.ibm.com/chips/news/2004/0210 asics.html
[9] V. Stojanovid, M. Horowitz, "Modeling and Analysis of High-Speed Links," Custom
Integrated Circuits Conference, Sept. 2003 Page(s):589 - 594.
[10] W. J. Dally, J. Poulton, "Transmitter Equalization for 4-Gbps Signaling," IEEE Micro,
Volume 17, Issue 1, Jan.-Feb. 1997 Page(s):48 - 56.
[11] A. Fiedler, R. Mactaggart, J. Welch, S. Krishnan, "A 1.0625 Gbps Transceiver with 2x-
Oversampling and Transmit Signal Pre-Emphasis," Solid-State Circuits Conference, Feb. 1997
Page(s):238 - 239, 464.
[12] H. Partovi et al, "A 62.5 Gb/s Multi-Standard SerDes IC," IEEE Custom Integrated
CircuitsConference, Sept. 2003 Page(s):585 - 588.
[13] R. Farjad-Rad, C.-K.K. Yang, M.A. Horowitz, "A 0.3-pm CMOS 8-Gb/s 4-PAM Serial Link
Transceiver," IEEE Journalof Solid-State Circuits, Volume 35, Issue 5, May 2000 Page(s):757
- 764.
[14] C.-K.K. Yang, V. Stojanovid, S. Modjtahedi, M.A. Horowitz, W.F. Ellersick, "A Serial-Link
Transceiver Based on 8GSample/s A/D and D/A Converters in 0.25pm CMOS," IEEE Journal
of Solid-State Circuits, vol. 36, no. 11, November 2001, pp. 1684-1692.
[15] W.T. Beyene, N. Cheng, C.Yuan, "Design and Analysis of Multi-Gigahertz Parallel Bus
Interfaces of Low-Cost and Band-Limited Channels," Electrical Performance of Electronic
Packaging, Oct. 2003 Page(s):213 - 216.
[16] J.L. Zebre et al, "Equalization and Clock Recovery for a 2.5-10-Gb/s 2-PAM/4-PAM
Backplane Transceiver Cell," IEEE Journal of Solid-State Circuits, Volume 38, Issue 12, Dec
2003 Page(s):2121 - 2130.
[17] J.F. Bulzacchelli et al., "A 10-Gb/s 5-Tap DFE/4-Tap FFE Transceiver in 90-nm CMOS
Technology," IEEE Journal of Solid-State Circuits, Volume 41, Issue 12, Dec. 2006
Page(s):2885 - 2900.
[18] Ming-Dou Ker; Yuan-Wen Hsiao, "On-Chip ESD Protection Strategies for RF Circuits in
CMOS Technology," Solid-State and Integrated Circuit Technology, 2006 Page(s):1680 -
1683.
[19] Hall, Stephen H., High-Speed Digital System Design: A Handbook of Interconnect Theory
and Design Practices. s.l.: Wiley-IEEE Press, 2000.
[20] D. Zwitter et al., "Competitiveness and Technical Challenges of Low Cost Wirebond
Packaging for High Speed SerDes Applications in ASICs," Electronic Components and
Technology Conference, May 2007 Page(s):777 - 784.
[21] Nanju Na, J. Audet, D. Zwitter, "Discontinuity Impacts and Design Considerations of
High Speed Differential Signals in FC-PBGA Packages with High Wiring Density," Electrical
Performance of Electronic Packaging, Oct. 2005 Page(s):107 - 110.
[22] H. Johnson, M. Graham., High-Speed Digital Design: A Handbook of Black Magic. s.I.:
Prentice Hall, 1993.
[23] M.S. Sharawi, "Practical Issues in High Speed PCB Design," IEEE Potentials,
Volume 23, Issue 2, Apr-May 2004 Page(s):24 - 27.
[24] A,Johnk CT., Engineering Electromagnetic Fields and Waves. New York :Wiley, 1988.
[25] J. Zerbe, C. Werner, R. Kollipara, V. Stojanovi', "A Flexible Serial Link for 5-10Gb/s in
Realistic Backplane Environments," DesignCon 2004.
77
[26] S.Sercu, J. De Geest, "BER Link Simulations," DesignCon 2003.
[27] S. Gong et al., "Packaging Impact on Switching Noise in High-Speed Digital Systems," IEE
Circuits, Devices and Systems, Volume 145, Issue 6, Dec. 1998 Page(s):446 - 452.
[28] P. Muthana et al., "I/O Decoupling in High Speed Packages Using Embedded Planar
Capacitors," Electronic Components and Technology Conference, May-June 2007
Page(s):299 - 304.
[29] Nanju Na; T. Budell, C. Chiu, E. Tremble, I. Wemple, "The Effects of On-Chip and
Package Decoupling Capacitors and an Efficient ASIC Decoupling Methodology," Electronic
Components and Technology Conference, Volume 1, June 2004 Page(s):556 - 567.
[30] Nanju Na; J. Audet, D. Zwitter, "Discontinuity Impacts and Design Considerations of
High Speed Differential Signals in FC-PBGA Packages with High Wiring Density," Electrical
Performance of Electronic Packaging, Oct. 2005 Page(s):107 - 110.
[31] E. Pillai, J. Weiss, "Novel T-Coil Structure and Implementation in a 6.4-Gb/s CMOS
Receiver to Meet Return Loss Specifications," Electronic Components and Technology
Conference, May-June 2007 Page(s):147 - 153.
[32] Voldman, Steven., ESD: RF Technology and Circuits. England : John Willey and Sons Inc.,
2006.
[33] Pozar, David M., Microwave Engineering. s.l. : John Willey and Sons Inc., 1998.
[34] T. Kushta et al., "Resonance Stub Effect in a Transition from a Through Via Hole to a
Stripline in Multilayer PCBs," IEEE Microwave and Wireless Components Letters, Volume 13,
Issue 5, May 2003 Page(s):169 - 171.
[35] Kim Jingook et al., "Via and Reference Discontinuity Impact on High-Speed Signal
Integrity," International Symposium on Electromagnetic Compatibility, Volume 2, Aug. 2004
Page(s):583 - 587 vol.2.
[36] W.T. Beyene, Shi Hao, J. Feng, C. Yuan, "Design and Analysis Methodologies of a 6.4
Gb/s Memory Interconnect System Using Conventional Packaging and Board Technologies,"
Electronic Components and Technology Conference, Volume 2, June 2004 Page(s):1406 -
1411 Vol.2.
[37] HFSS, Ansoft. [Online]

(http://www.ansoft.com/hfss).
[38] Shaowei Deng et al., "Effects of Open Stubs Associated with Plated Through-Hole Vias
in Backpanel Designs," International Symposium on Electromagnetic Compatibility, Volume
3, Aug. 2004 Page(s):1017 - 1022 vol.3.
[39] High-Speed Signal Propagation: Advanced Black Magic. Graham, H. Johnson and M.
s.l.: Prentice Hall, 2003.
[40] Nanju Na et al., "Design Optimization for Isolation in High Wiring Density Packages with
High Speed SerDes Links," Electronic Components and Technology Conference, May-June
2006 Page(s):187 - 193.

Mit PDF

Uploaded by

Copyright:

Available Formats

Mit 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

Mit PDF

Uploaded by

Copyright:

Available Formats

MASSACHUSETTS INSTITUTE

Package and PCB Solutions for High-Speed

Bachelor of Science in Electrical Engineering and Computer Science

Submitted to the Department of Electrical Engineering and Computer Science

Master of Engineering in Electrical Engineering and Computer Science

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

©2008 Massachusetts Institute of Technology. All rights reserved.

Submitted to the Department of Electrical Engineering and Computer Science

August 27, 2008

In Partial Fulfillment of the Requirements for the Degree of

Thesis Supervisor: Nanju Na

Thesis Supervisor: Vladimir Stojanovic

Chapter 1 : Introduction ..................................................................................... 17

This thesis presents an in depth analysis of different bottleneck areas in high-speed

Chapter 2 provides an overview of the structure of a typical high-speed channel and

As a step towards package performance enhancement, a concept of discontinuity

It is recognized that isolated package design cannot be the solution to overall

Figure 2.1: Diagram of a high-speed serial link

The aggregate termination impedance becomes frequency selective because of

2.2.1 Attachment of the Die to the Package

(a) Chip is mounted (b) Chip is mounted with

Thermally and mechanically, however, flip-chip packaging is dismal. The thermal

2.2.2 Routing of the Signals on the Package

2.2.3 Attachment of the Package to the Board

1 Mutual inductance - causes crosstalk

2 Series inductance - slows down signal propagation and creates electromagnetic

3 Parasitic capacitance - Slows down signal propagation

Figure 2.4: Cross-section of PCB trace configurations defined by parameters w, t, h, and b:

In order to improve the performance of high-speed links, we must have a good

3.1 Transmission Line Losses

In modern interconnects smaller dimensions and high-frequency content cause the

3.1.1 Conductor DC Losses

3.1.2 Dielectric DC Losses

3.1.3 Frequency-Dependant Conductor Losses

3.1.3.1 Frequency-Dependant Resistance and Inductance

The total inductance of a conductor is caused by magnetic flux produced by the

3.1.3.2 Effect of Conductor Surface Roughness

3.1.4 Frequency-Dependant Dielectric Losses

Frequency-dependant dielectric loss increases linearly with signal frequency:

Table 3.1: Loss Tangent for Typical PCB insulator.

Material FR4 Polyamide GETEK Rogers4350 Teflon

Figure 3.3: Transmission line system with multiple line impedance.

3.3 Crosstalk (FEXT, NEXT)

3.3.1 Mutual Inductance and Mutual Capacitance

VsLm = dlaggressor (3.6)

3.3.2 Types of Crosstalk

3.4 Power Integrity (digital, analog)

On-Chip Ground Ideal chip capacitance

Figure 3.5: Power supply circuit including package parasitic model

In the remaining portion of this thesis we concentrate on challenges in high-speed

This chapter discusses a package design technique to enhance high-speed signal

Discontinuities in high speed package interconnects are having more significant

u;4 - Chinlp Pads

4.1 The Counter-Discontinuity Cancellation Technique Description

In this chapter we look at an alternative approach for obtaining further improvement

transmission. The technique uses transmission line behavior in short

4.2 Theory: Distributed Network Behavior of Short Discontinuities

Total transmission at the line end of Zj: T=TO+T1+T2+...

) at t=O, (1+T1)(1-F1 )=(1-F' 2)