22 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO.

1, JANUARY 2001

CMOS Op-Amp Sizing Using a Geometric

Programming Formulation
Pradip Mandal and V. Visvanathan, Member, IEEE

Abstract—The problem of CMOS op-amp circuit sizing is The main reason for this is that the number of performance
addressed here. Given a circuit and its performance specifications, functions in an analog circuit is much larger than that in a dig-
the goal is to automatically determine the device sizes in order ital circuit. Further, analog performances are very sensitive to
to meet the given performance specifications while minimizing a
cost function, such as a weighted sum of the active area and power
the design variables and variation in the performance across the
dissipation. The approach is based on the observation that the first design space is quite high. In other words, the analog design
order behavior of a MOS transistor in the saturation region is such problem is a complex tradeoff problem that is knowledge inten-
that the cost and the constraint functions for this optimization sive. However, the research community has been aggressively
problem can be modeled as posynomial in the design variables. The working for computer-aided analog design. A good survey of
problem is then solved efficiently as a convex optimization problem. recent analog synthesis techniques is available in [1].
Second order effects are then handled by formulating the problem
as one of solving a sequence of convex programs. Numerical Existing approaches of automatic circuit sizing are broadly
experiments show that the solutions to the sequence of convex pro- classified into three main categories, namely knowledge-based
grams converge to the same design point for widely varying initial optimization, simulation-based optimization, and analytical
guesses. This strongly suggests that the approach is capable of equations-based optimization. In this context, we note that,
determining the globally optimal solution to the problem. Accuracy unlike in the digital domain, the standard cell based approach
of performance prediction in the sizing program (implemented
in MATLAB) is maintained by using a newly proposed MOS
[2] is quite restrictive in the analog domain.
transistor model and verified against detailed SPICE simulation. Since analog design requires detailed circuit knowledge, a
major approach of implementing an analog synthesis tool has
Index Terms—Cell-generation, device-models, optimization,
been the knowledge-based approach. Some of the existing tools
transistor-sizing, VLSI.
which follow this approach are BLADES [3], OASYS [4], and
IDAC [5] and [6]. However, the application of this approach has
I. INTRODUCTION been limited due to requirement of having to codify extensive
circuit knowledge and design heuristics.
T HE CURRENT trend in microelectronics is to integrate
a complete system that previously occupied one or more
boards on one or a few chips. Although most of the function-
On the other hand, DELIGHT.SPICE [7], ASTRX/OBLX
[8], FRIDGE [9], MAELSTROM [10], and ANACONDA [11]
use the simulation-based optimization approach. This approach
ality in an integrated system is implemented in digital circuitry,
does not require much circuit knowledge. Hence, the main
analog circuits are needed to interface between the core digital
advantage of this approach is that a wide range of circuits can
system and the real world. Therefore, to realize an integrated
be synthesized. However, the basic limitation comes from the
system on a single chip, the digital and analog circuits are com-
requirement of costly circuit simulation in each iteration of the
bined together. This integration of analog and digital circuits re-
optimization algorithm.
sults in so called mixed-signal integrated circuits which have a
To reduce the CPU time of optimization-based techniques, the
large market of applications in the telecom, consumer products, third approach is analytical equations-based optimization, where
computing, and automotive sectors. the circuit performances are evaluated using analytical equa-
Increase of design complexity and, at the same time, demand tions. OPASYN [12] uses simple analytical equations of op-amp
of design cycle time reduction due to highly competitive market performance. OPTIMAN [13] uses a symbolic simulator, ISAAC
can be managed only by the use of computer aided design. [14], to get the analytical models of the ac performances of a
Though in an integrated system, the analog circuitry occupies circuit. However, in [13], analytical models for dc and transient
a small physical area compared to the digital counterpart and performances have to be provided by an expert designer. In
becomes the bottleneck in design time reduction. circuit sizing, the use of a single weighted cost function [12] is
inadequate since sizing is a constrained optimization problem
Manuscript received July 20, 1998; revised July 18, 1999 and April 7, 2000.
with complex tradeoffs among the constraints. With simulated
An earlier version of this paper was presented at the 12th International Confer- annealing [13], the drawback is that it is computationally inten-
ence on VLSI Design, Goa, India, January 1999. This paper was recommended sive (even with the use of analytical equations of performances)
by Associate Editor K. Mayaram.
P. Mandal was with the Indian Institute of Science, Bangalore, India. He is
and cannot be realistically used in an interactive setting.
now with Philips Semiconductors, Bangalore, India (e-mail: pradip.mandal@ It therefore appears that an analytical equation based con-
philips.com). strained optimization method is the most promising approach
V. Visvanathan was with the Indian Institute of Science, Bangalore, India. for automatic circuit sizing. However, the existing technique
He is now with Cadence Design Systems, Allentown, PA 18195 USA (e-mail:
vish@cadence.com). that uses this approach [15] suffers from the drawback that it
Publisher Item Identifier S 0278-0070(01)00349-9. needs expert designer knowledge to sequentially introduce the
0278–0070/01$10.00 © 2001 IEEE
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 23

constraints. If this is not done, the method may fail to provide

even a feasible design point. Further, any optimal point that is
provided is only a local optimum design point. Thus, the objec-
tive of this work is to propose an analytical equation-based con-
strained optimization method that is fast, robust, and provides
the globally optimal design point.
The main purpose of developing an automatic sizing tool is
that the design space of a circuit will be quickly explored to
find a design point at which the circuit satisfies the required
specification and, at the same time, some cost (e.g., area, power)
is minimized. This places a number of stringent requirements on
the tool, which are enumerated below. Fig. 1. MOS transistor characteristic.
• In order to support an iterative design methodology, the
tool should be fast. characteristic of a transistor into its linear region of operation.
• The final solution point should be independent of the ini- The use of the extrapolated transistor characteristic results in a
tial guess. design formulation that is simple and robust.
• The optimization technique should be extremely robust. The second idea on which our sizing technique is based is that
In particular, the designer should not be burdened with the the first-order and higher order behavior of the MOS transistor
task of tuning the optimizer. have been separated. With the first-order model, it is shown that
In this paper, a circuit sizing method for CMOS op-amps is the constraint functions and the objective function for the design
proposed, which meets these three requirements as verified via optimization problem that has been addressed are posynomials
a prototype implementation applied to a number of two-stage in the design variables, namely transistor sizes and biases. In
op-amps. The effectiveness of the method is due to the use other words, with the first-order model, the op-amp synthesis
of convex optimization techniques via a geometric program- problem is formulated as a geometric programming problem.
ming formulation. It is similar to the recent work described in With a logarithmic transformation, a geometric programming
[16]–[19], which appear to have been done independently and problem becomes a convex optimization problem.
in parallel to this work [20]. However, compared to [16], our Finally, the higher order effects are handled by iteratively
technique addresses second-order effects and is far less restric- updating the first-order model parameters using the higher
tive in terms of the range of MOS models it supports. This is order transistor models, and by solving a sequence of convex
because, unlike [16]—where the problem is solved as a single programs. The technique of iteratively updating the first-order
geometric program—we show that it is necessary to formulate model parameters is elaborated in Section IV. The parameter
and solve the problem as a sequence of (convex) geometric pro- update is based on iteratively refining the dc operating point.
grams. In other words, the formulation of [16] is just the first Section III-A describes a systematic approach for finding the
step of our approach. This difference is crucial in the context of dc operating point.
submicrometer technology, since the models necessary to sup-
port our methodology need to support the geometric program- A. Extrapolated Transistor Characteristic
ming paradigm only in their first order effects, while the models
Usually transistors in an analog circuit are biased in the
of second order behavior can be arbitrary. The sequential convex
programming formulation is made possible by judiciously com- saturation region where the drain conductance is low, which
bining various techniques, most notably relaxed dc formulation helps to get high ac performance. Therefore, for analog design
[15] and casual dc analysis [22] with the powerful geometric analysis, one can use only the saturation region characteristic.
programming paradigm [23]–[25]. Systematic use of the saturation region characteristic is pictori-
The organization of the rest of this paper is as follows. In ally illustrated in Fig. 1. The continuous curves are the actual
Section II, the basic approach of op-amp sizing is described. In – characteristic curves of a transistor with different .
Section III, with the first-order Shichman–Hodges (S–H) model, The dotted line divides the whole region of operation into two
the op-amp sizing problem is formulated as a convex program- parts, linear (left) and saturation (right) regions. The dotted line
ming problem. Section IV describes how the higher order effects cuts a characteristic curve at a point where is equal to the
can be captured through an iterative approach. In Section V, the drain saturation voltage . By extrapolating the saturation
sizing technique is applied on a large number of op-amp circuits. region characteristic curves into the linear region, as shown by
Experimental results are provided in Section VI. By the use of the dashed lines in Fig. 1, a set of artificial characteristic curves
a new accurate MOS model, the accuracy of performance pre- are obtained. These characteristic curves are simple and smooth
diction has been increased in Section VII. Section VIII provides over the entire region of operation. For analog design analysis,
a discussion, while Section IX summarizes the work that is de- rather than using the actual characteristic curves, these artificial
scribed in this paper. characteristic curves can be used. However, to ensure that at
the final design point the transistors are actually in saturation,
it is necessary to satisfy the constraint for all
II. THE BASIC APPROACH
transistors. In the following section, we propose a systematic
The proposed op-amp sizing technique is based on three basic approach of circuit analysis for finding a set of design space
ideas. The first one is that of extrapolating the saturation region constraints which helps to satisfy the constraint on .
24 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

If the S–H model is used for the extrapolated device char-

acteristic curves, over the entire region of operation transcon-
ductance , drain conductance , and effective gate-to-source
voltage ( – ) are, respectively,

proportionality constant

and Fig. 2. Simple op-amp, a running example.

(1)
III. SIZING FORMULATION WITH A RUNNING EXAMPLE
Note that, assuming constant , the device model pa- Various steps of op-amp sizing are described here by using a
rameters , , and – are product of power (PoP) simple running example, shown in Fig. 2. Consider the design
functions of , , and . A PoP function is the product optimization problem as
of a positive coefficient and various variables that are raised to
some power (any real number). This is the key information on minimize a weighted sum of gate area and power while
which our sizing formulation is based. low frequency gain,
The extrapolated device characteristic curves and the design
unity gain frequency,
space constraints are used in our proposed sizing technique.
slew rate,
B. Geometric Programming Problem common mode range, CMR CMR
A geometric programming problem is of the following form:

minimize
subject to (2)
In this design problem, and of the transistors and the
and
bias current are the design variables. In the op-amp,
The objective function and each one of the constraint and are matched pairs. The various intermediate steps
function are posynomials of the design variables ’s. In of the design formulation follow.
equation form, a posynomial function is
A. DC Analysis
(3) In this step, the circuit is analyzed to determine analytical
equations that give its dc operating point. These equations are
solved to get the dc operating point, which is then used for pre-
where ’s are positive coefficients and ’s are arbitrary real
dicting its performance.
numbers. Note that in the above equation there are a number of
It is observed that in a CMOS analog circuit, various node
PoP terms. The coefficient is referred to as the constant of
voltages can be defined by the gate-to-source voltages of various
the corresponding PoP term. For any , the corresponding ’s
transistors. Further, the gate-to-source voltages of the transistors
are referred to as the powers of the th PoP term. With a loga-
can be determined by its size and its dc current. Finally, the dc
rithmic transformation on the design variables, the posynomial
current through all the transistors can be essentially determined
function is
by only a few transistors, which we refer to as current source
(4) transistors.
An exception to the approach described above is the output
node, whose voltage usually cannot be determined from the
where . It can be easily shown that is a convex ’s of various transistors. However, in actual applications,
function [23] of the transformed variables ’s. Therefore, op-amps are used in a closed loop configuration, where the
with the logarithmic transformation on the design variables, output node voltage is stable at a predetermined value (usually
the geometric programming problem becomes a convex zero) [22].
programming problem. In the following, it is shown that the In the example op-amp, the bias stage current is and the
op-amp sizing problem is a geometric programming problem. drain current of the current source transistor is
As a result, the op-amp design problem becomes a convex
programming problem. The well known property of a convex
programming problem is that any of its local minima is also a
global minimum. (5)
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 25

V V
TABLE I
AND OF VARIOUS TRANSISTORS

The current determines the current through the other

transistors. Since the transistor pairs and are
matched, their drain currents are

(6)

From (1), the gate-to-source voltage of all transistors can be

written as

(7)

Now, using a casual dc analysis which is described in [22],

various node voltages are expressed in terms of ’s and var-
ious biases. Fig. 3. Iterative approach of finding dc operating point.
The analysis starts from the positive supply and moves to-
ward the negative supply rail. Consider the transistor . The
currents are then used to find the current through the remaining
drain-to-source and gate-to-source voltages of the transistor
transistors in the circuit.
are1
In the next step, ’s are determined from ’s, ’s,
’s, and ’s of the transistors. In the subsequent
step, from the ’s, various node voltages are determined by
Therefore, using (5)–(10). In the final step, ’s and ’s are evaluated
using their equations, which are given in Table I.
(8) The values of the ’s and ’s, which are obtained at
the end of each iteration, are then used in the next iteration to
Note that the node voltages and are the same, since
get a more accurate estimate of the dc operating point. The ter-
and are assumed to be at the same dc bias. Now considering
minating condition is that all node voltages in two consecutive
the transistors and , respectively,
iterations are very close. It is found that this formulation re-
(9) sults in a highly contractive fixed-point scheme that converges
very fast. This is because, across the design space, the value
and
of ’s are close to one and the back bias depen-
(10) dency function of the threshold voltage is strictly monotonic
with small slope.
These node voltages are now used to express the ’s and
’s of the various transistors as given in the Table I. B. Design Space Constraints
The above analysis shows that (5)–(10) and Table I represent
a complete set of nonlinear equations whose solution provides Here we find the design space constraints by which all the
accurate node voltages. An effective way of solving this set of transistors are kept in saturation, i.e., away from subthreshold
nonlinear equations is through a fixed-point scheme. The overall and linear regions. To keep all the transistors away from the
method of finding node voltages is shown in Fig. 3. subthreshold region with a margin of , the constraints to
In the first step, channel length modulation factor, be satisfied are
, and threshold voltage, , of various tran-
(11)
sistors are updated based on the and values in the
previous iteration. Note that in the first iteration, ’s
On the other hand, to keep a transistor away from the linear
are assumed to be one and all the ’s are taken to be .
region, we require
Next, the drain current of all current source transistors in the
circuit (e.g., in the example circuit) are found. These for n-type
(12)
V V or for p-type
M
1 and represent, respectively, the magnitudes of drain-to-source
and gate-to-source voltages of the p-transistor . where
26 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

Note that in the S–H MOS model, is equal to . How- The negative common mode range of the op-amp is CMR
ever, this new notation is introduced for ease of extending the . Note that CMR has negative
design formulation for any other MOS model, such as the one value. Therefore, to get the specified CMR
described in [20].
In the example circuit, the gate and drain voltages of the tran- CMR
sistors and are the same. Therefore, these two transistors or
are always in saturation. However, to keep (and ) in satu-
(19)
ration, we require . In quiescent condition, CMR
. Further, from (8), . Therefore, the
On the other hand, the positive common mode range is
design inequality is
CMR , which should be greater than
CMR . In other words
or
(20)
(13) CMR
The objective function is a weighted sum of total effective
To keep transistor away from linear region, we require gate area (in micrometers micrometers) and total quiescent
. Using the expressions of and in (9) and (10), current of the op-amp (in microamperes), which is given by
we get
(21)

In this equation, and are two specified weights.

(14) To summarize the formulation which is given in this sec-
tion, an op-amp design problem is a constrained optimization
Note that, assuming constant ’s and ’s, (11), (13), problem of the following form:
and (14) are posynomial functions of ’s, which are con-
minimize
strained to be less than or equal to one. As will be made clear (22)
subject to
in the following sections, this is an important step toward our
formulation. The objective function is given in (21) while various constraint
Finally, to keep the transistor sizes within the specified limit, functions are provided in (11) and (13)–(20). It is observed that,
the following inequalities should be satisfied: assuming constant , and ( , the constraint
functions ’s are posynomials of , , , and of var-
ious transistors. Further, the objective function is a posyn-
omial of , , and of the transistors. With the same as-
sumption, it is also found that of the transistors are PoP func-
tions of the design variables ( , , and the bias current ).
(15)
Therefore, with the extrapolated S–H model where the device
parameters , , and are also PoP’s in the design vari-
ables, the objective and the constraint functions are posynomials
of the design variables. With logarithmic transformation, the op-
timization problem becomes a convex programming problem.

C. Performance Constraints and Objective Function IV. FORMULATION OF OP-AMP DESIGN AS A SEQUENCE OF
The low frequency gain of the op-amp is . CONVEX PROGRAMMING PROBLEMS
Therefore, to meet the gain specification In the last section, assuming constant , , and
, the op-amp synthesis problem is formulated as a convex
(16)
programming problem. Here, to account for the effect of vari-
ations in , , and , the overall method is
The unity gain frequency of the op-amp is . So to get shown in Fig. 4. The various steps of the method are as follows.
the specified
Step 0) Accept the designer specifications and the circuit
to be sized. The user may provide an initial design
(17)
point, which is optional. By default, the initial de-
sign point is taken as the minimum feature size for
Slew rate of the op-amp is . So the constraint to get
all the transistors. Using two iterations of the dc
the specified slew rate is
analysis technique described in Section III-A, an
approximate dc operating point at the initial design
(18) point is found.
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 27

pass, to get the updated values of ’s, ’s,

and ’s, the values of ’s and ’s are taken
from the operating point at the solution point in the
previous iteration.
Step 5) In this step, the convergence of the sequence of so-
lution design points and node voltages are checked.
If the design points and the node voltages of the last
two iterations are very close, then the op-amp netlist
with sized transistors is provided. Otherwise, we go
back to Step 1.
Note that within an iteration, the values of ’s and ’s,
which are used in Step 1 to determine the convex programming
problem, are not mathematically consistent with the solution
point. However, as the iterations proceed and when the sequence
of solution points converges and the node voltages converge, the
inconsistency becomes negligible. Such an approach is in spirit
similar to the relaxed dc formulation, which is described in [15].

V. TWO-STAGE CMOS OP-AMP SIZING

The op-amp sizing technique, which is described in the
last two sections, is applied here for designing a number of
two-stage op-amps. For ease of referencing, in Section V-A,
the op-amp topologies are characterized by five binary vari-
ables. While details of the design formulation of the two-stage
op-amps are available in [20], a summary of dc analysis of
the op-amps, design space constraints, and the performance
constraints are provided, respectively, in Sections V-B–V-D.
The design formulation is similar to that of the running example
op-amp, which is given in Section III.
Fig. 4. Op-amp sizing method through sequential convex programming.
A. Characterization of Two-Stage Op-Amps
Step 1) In this step, the constants and powers of the PoP rep- In the class of two-stage op-amps, there is a basic structural
resentation of the various ’s are determined. The similarity, namely the hierarchical structure of different config-
values of ’s, which are used to determine the urations is the same. It is only the subcircuits, which are the
constant, come from the previous iteration. For the leaf cells of the hierarchy, that are different across the various
example circuit, refer to (5) and (6) for the relevant topologies [29].
PoP functions. Similarly, the constants and powers A two-stage op-amp consists of an input stage, a second stage,
of the PoP representation of the various ’s, ’s, and a compensating circuit. The input stage has three parts:
and ’s are also determined. Using the values of current source, differential pair, and current mirror. The second
’s, which come from the previous iteration, the stage has two parts: transconductance amplifier and active load.
threshold voltage of the transistors are also evalu- Each one of the four subcircuits, namely differential pair, current
ated here. mirror, transconductance amplifier, and active load, can be either
Step 2) From the constants and powers of ’s, ’s, simple or cascoded. For a cascoded current mirror, a level shifter
’s, ’s, and the values of ’s, the objec- is required between the input stage and the second stage. The
tive function and the constraint functions in the compensating circuit consists of a capacitor and a resistor. Fur-
transformed design space are derived. Both the ther, the transistors in the differential pair can either be n-type or
objective function and the constraint functions are p-type. The choice of polarity of the transistors in the differential
convex functions. pair also determines the polarity of the transistors in the other
Step 3) The convex programming problem is then solved subcircuits. All the possible op-amp topologies are characterized
to find the global optimal solution of the current by five binary variables, which are defined in Table II. With
iteration. n-type differential pair transistors, the supercircuit [29] of the
Step 4) In this step, the dc operating point is updated for the considered set of op-amps is given in Fig. 5. In the same figure,
new solution design point. The method of finding the relation among the binary variables and the subcircuits is
the dc operating point is the same as the iterative also indicated by naming the enclosed subcircuits by the corre-
method described in Section III-A. The key differ- sponding binary variables. So, from the supercircuit, depending
ence, however, is that only one pass of the flow on the values of the binary variables, other topologies can be
graph shown in Fig. 3 is executed. In the single constructed by selectively including the enclosed subcircuits.
28 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

Fig. 5. Supercircuit of the two-stage op-amps.

TABLE II However, a simple way to get is that the sizes of tran-

DEFINITION OF THE TOPOLOGY CHARACTERIZING BINARY VARIABLES
sistors and are equal, the length of the two transistors
and are the same, and their widths satisfy the following
constraint:

(25)

Now with , the current through the transistors

and are

(26)
B. DC Analysis
DC analysis includes finding drain currents through various The gate-to-source voltages of all the transistors are determined
transistors, determining ’s of the transistors, expressing by
various node voltages in terms of ’s, and, finally, writing
’s and ’s in terms of the node voltages. The current
through the input stage and the second stage are, respectively, where

(27)

(23)
Through a casual dc analysis, which is discussed in Sec-
and tion III, various node voltages are expressed in terms of ’s
and bias voltages

The current through the other transistors are

for (28)
(24)
for
The sizes of transistors in the level shifter are so chosen that
the node voltages and are equal. With , the
balance equation (39), which helps to achieve small systematic Node 16 is the output node of the op-amp. As described in Sec-
offset, remains simple even with fully cascode first stage. There tion III, this node does not have a well defined equation. How-
are a number of possible ways by which this can be achieved. ever, in actual application, op-amps are used in close loop con-
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 29

figuration, and in this configuration the output node voltage is, and for cascode differential pair, the required inequality is
usually, stabilized to zero. It may be noted that
.
The back bias of the transistors in the op-amp in terms of i.e.
various node voltages are as follows: (33)
for In cascode differential pair, the transistors (and ) are
kept in saturation by the inequality
(29)
i.e.

Finally, drain-to-source voltages of the transistors are expressed

in terms of the node voltages as follows (34)
The transistor is kept in saturation by the inequality

i.e.

(30) (35)

In the second stage, the transistor is kept in saturation

by the inequality

i.e.
Note that the expressions of some ’s include the binary
variables ’s. This is because these ’s are topology de-
pendent. Therefore, the binary variables are used to define the
(36)
’s across the various op-amp topologies.
and the transistor is kept in saturation by the inequality
C. Design Space Constraints
The design space constraints include the technology limits
of the transistor sizes and the constraints on design variables,
which help to bias the transistors in saturation. The constraints (37)
corresponding to the technology limits are straight forward, and Equation (37) is derived based on the assumption that the
hence they are not explicitly given here. Recall from Section III . The design of the level shifter to get this biasing condition
that the inequalities that keep the transistors away from the is already discussed in Section V-B. Note, finally, that the re-
linear region are given by maining transistors in the supercircuit are automatically biased
for n-type in the saturation region.
(31) All the transistors in the op-amps are kept away from the sub-
or for p-type
threshold region by the inequalities,
where
(38)
Using the appropriate values of the node voltages from (28) in
the above inequality, we get the required constraints.
Along with the various inequality design space constraints,
Consider the input stage of the supercircuit. To keep the tran-
there are two equality constraints also. One of them is given in
sistor (and ) in saturation, the required inequality for
(25), which is related to designing the level shifter to get
simple differential pair is
. The other equality constraint is the balance equation, which
is given by

i.e.

(39)
(32)
30 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

The balance equation is necessary to keep the systematic offset On the other hand, when the active load is cascoded ( ),
small [27]. the negative output swing constraint is
The concluding remarks about the design space constraints
are that they are posynomial in ’s if the ’s and ’s (46)
are assumed to be constant. With a further simplifying assump-
Note that the magnitude of the bias voltage , rather than its
tion, i.e., constant , the ’s are PoP functions
actual value, is taken as the design variable.
of the design variables for the S–H model. Therefore, the de-
To meet the slew rate specification, the following two inequal-
sign space constraints are posynomials of the design variables.
ities should be satisfied
Finally, the equality constraints are PoP functions of design vari-
ables. Therefore, each one of the equality constraints can be
translated into two PoP terms less than or equal to one. (47)

D. Performance Constraints and Objective Function

This section provides analytical expressions of various per- 2) AC Performances: The low frequency gain of an op-amp
formance functions of the op-amps. Performance constraints are is
then derived by constraining the performance expressions by
their respective specified value. The analytical expression of the
objective function is also given here.
1) DC and Transient Performances: For simple input dif- where output conductances of the two stages are
ferential pair ( ), the positive common mode range is
CMR . Therefore, to meet
its specification, the constraint is (48)

CMR
To satisfy the gain specification, we require
(40)

On the other hand, for cascoded input differential pair (i.e.,

), CMR and the corresponding i.e.
performance constraint is (49)

CMR (41) where

The negative common mode range is given by CMR

, and to meet its specification

CMR
(42)

With simple output transconductance amplifier ( ), the

positive output swing is and
the required constraint to meet specification is
Transfer function of a two-stage compensated op-amp can be
well approximated with its low frequency gain, one zero and
(43)
three poles. With appropriate choice of the compensating re-
sistor ( ), the zero of the transfer function of the
On the other hand, when the transconductance amplifier is cas- op-amp can be placed at very high frequency [27]. The three
coded (i.e., ), the positive output swing is poles of the transfer function are well approximated by
and to meet its specification we require

(44)
(50)
With a simple second-stage active load, to meet the negative
output swing specification, we require and

In order to have a transfer function that is well approximated

(45)
by a single pole system in the frequency range of interest, the
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 31

second pole should be beyond the gain bandwidth product where

( ), i.e., the following should be satisfied CMRR
(51)
CMRR
For a true single pole behavior, unity gain frequency (UGF) is
equal to . However, with the approximate single pole
CMRR
behavior that we get by (51), well approximates the
UGF. Hence, the specified UGF can be achieved by
CMRR
UGF
(52)
Note that the above equation for CMRR—which, as our
The phase margin is results in Section VII show, matches SPICE simulation
well—does not include the effect of mismatch. Therefore, for
PM UGF UGF a more realistic estimate of CMRR, the effect of statistical
mismatch variations should be routinely incorporated in SPICE
To meet phase margin specification simulation and in analytical techniques such as ours.
3) Objective Function: In the optimization formulation, a
UGF UGF PM
weighted sum of the total gate area of the transistors and the
or power dissipation is taken as the objective function to be mini-
UGF UGF mized.
PM The total power dissipation is
UGF
PD
or

and the total effective gate area is

UGF UGF UGF
EFF AREA
PM
or

PM

(53)

The low-frequency CMRR of an op-amp is

Note that all the performance constraint functions are posyno-
mials in the device parameters ’s, ’s, ’s, and ’s.
CMRR These device parameters are PoP function of design variables.
CMG
where Therefore, the performance constraint functions are posynomials
in the design variables. Further, the power dissipation and the
CMG total effective gate area are posynomials of the design variables.
and As with the design space constraints, the performance constraints
are expressed in terms of device parameters, which are in turn
functions of the actual design variables.
i.e.
VI. SIMULATION RESULTS
CMRR The CMOS op-amp sizing technique that has been described
so far has been implemented in MATLAB [30] for the class
(54) of two-stage CMOS op-amps. In the implementation, each
convex programming problem is solved by using the sequential
To meet its specification, we require quadratic programming method, which is available in the op-
timization toolbox [30]. This was done in the interest of quick
CMRR prototyping in order to provide a proof of concept. As our exper-
CMRR imental results show, this implementation is very fast in spite of
i.e. the fact that we did not use any of the far more efficient special
(55) purpose techniques that exist for convex optimization [25].
32 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

TABLE III TABLE V

SPECIFICATIONS, S–H MODEL BASED PREDICTED PERFORMANCES AND SPICE OPTIMAL DESIGN POINT OF THE FIRST EXAMPLE
SIMULATED PERFORMANCES AT THE FINAL DESIGN POINT OF THE OP-AMP
IN THE FIRST EXAMPLE

TABLE VI
COMPARISON OF THE FOUR FINAL DESIGN POINTS OBTAINED BY STARTING
TABLE IV FROM FOUR INITIAL GUESSES IN THE FIRST EXAMPLE
OPTIMIZATION STATISTICS FOR THE FIRST EXAMPLE BY STARTING FROM
SMALL SIZES OF ALL THE TRANSISTORS

and the maximum difference in the node voltages was less than
1 mV. With the other three initial guesses, the convergence be-
havior was essentially the same. The sequence of convex pro-
grams converged in less than 50 s of CPU time with at most five
iterations.
A number of two-stage op-amps were sized for a 1.6- m tech- The four solution design points, which were obtained by
nology. The experimental results are given in this section. These starting from the four initial design points 1)–4), are denoted
results include CPU time required to find the final design point, by , , , and , respectively. The solution
the convergence behavior of the method, and a comparison of point is given in Table V, while Table VI reports the
predicted performances with SPICE simulations based on the differences among the four solution design points. Note that
level two MOS model. all four solution design points are essentially the same. It
Example 1: The topology for the first example was the should also be noted in this context that in this and the other
simplest op-amp (all subcircuits were simple). The set of per- examples in this paper, we have reported the details of the
formance specification of this example is given in the second iterative process with precision in the millivolt range for node
column of Table III. voltages and in the nanometer range for device sizes. This has
The circuit was sized by starting from four different initial been done in order to demonstrate that our technique converges
design points. The four initial design points are: 1) small size to a consistent dc operating point and that the final solution is
(minimum feature size m, m) of all the the same even with very different initial guesses. However, this
transistors; 2) small size of the input-stage transistors and large does not mean that the transistors have to be sized this precisely
size ( m, m) of second-stage transistors; to get the reported performance.
3) large size of the input stage transistors and small size of the An expert designer would pick a nonminimum channel length
second-stage transistors; and 4) large size of all the transistors. for the input stage transistors to get high gain. On the other hand,
The optimization statistics in case 1), i.e., starting with small for the second-stage transistors, in order to achieve high slew
size of all of the transistors, is shown in Table IV. In the table, rate, he/she would choose minimum channel length with large
the first column indicates the iteration number of the sequential channel width. The optimal design point given in Table V is
convex optimization algorithm. The second column gives the qualitatively similar to such a choice.
CPU time (IBM RS/6000, running AIX) required for solving The various performances at the optimal design point as
the convex programming problem using sequential quadratic predicted by the program using the S–H model and the corre-
programming. The last three columns of the table provide the sponding SPICE simulations (using the level two MOS model)
maximum difference in transistor lengths and widths, and node are given in the last two columns of Table III.
voltages at the two-solution design points in two consecutive it- Note that while the S–H model based predictions satisfy the
erations. The iteration process was stopped when the maximum specifications, many of the performances as actually measured
difference in the lengths and the widths was less than 0.02 m in SPICE (using level two model) do not. The inaccuracies are
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 33

TABLE VII VII. IMPROVING DESIGN ACCURACY USING ACCURATE

OPTIMAL DESIGN POINT IN THE SECOND EXAMPLE MOS MODEL
In the previous section it has been demonstrated that the pro-
posed sizing method is robust and it has the capability of picking
the optimal design point. However, in the implementation of
the technique, the S–H MOS model was used, which is not an
accurate model for small-channel length transistor. Therefore,
at the design points, which were obtained by the design op-
timization, the op-amps could not meet some required perfor-
mances. To improve the design accuracy, in the second version
of the prototype, the S–H model was replaced by a newly pro-
posed MOS model – (alpha–Analog). While the details of
the model are available in [20], in the following section the key
equations are provided. Section VII-B provides some modified
equations compared to those in Section V, which are required
for replacing the S–H model by the – model. Simulation ex-
amples are provided in Section VII-C.
TABLE VIII
SPECIFICATIONS, S–H MODEL BASED PREDICTED PERFORMANCES, AND
SPICE SIMULATED PERFORMANCES AT THE OPTIMAL DESIGN POINT OF A. The Model Description
THE SECOND EXAMPLE
The – MOS model is an extension of the simple -power
law MOS transistor model proposed by Sakurai and Newton
[31]. It may be noted that, since in the -power model the drain
conductance over the saturation region is taken as zero, it can
not be directly used for analog circuit analysis.
A summary description of the – model is given below.
As the model is intended for use in analog circuit design, the
model description is only for the saturation region of operation.
A transistor is in the saturation region when

and (56)

The threshold voltage and the drain saturation voltage are, re-
spectively,

due to the inadequacy of the S–H MOS model in the short

(57)
channel regime. However, for long channels the model is quite and
accurate. To demonstrate this, in the following example we re-
strict the channel length of the transistors to be more than 2.6 In the saturation region, drain current, transconductance, and
m and do the design optimization. drain-to-source conductance are, respectively,
Example 2: This example is the same as the first example
except that the lengths of the transistors were restricted to be
more than or equal to 2.6 m (though the technology limit is
1.6 m). The solution design point is given in Table VII. The
predicted performance of the op-amp at the final design point
is given in Table VIII. Note that at the final design point, the
predicted performances using S–H MOS model is close to those
of SPICE simulation. However, the total gate area and the power (58)
dissipation at the solution design point obtained in this example
are, respectively, 57% and 38% more than those at the solution
design point obtained in the first example. It is therefore clear In (57), is the zero back bias threshold voltage, while
that the S–H model is inadequate for synthesis tools targeted at and represent dependencies of threshold voltage on
modern short-channel processes. So, we need a MOS model that the back bias and the channel length , respectively. The
is accurate in the short-channel regime. To increase the design dependency functions are
accuracy in the prototype, the S–H model is replaced by a newly
proposed MOS model. The corresponding simulation results are
(59)
provided in the following section. and
34 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

In the same equation (57), the parameters and are given TABLE IX
as SPECIFICATIONS, NEW MODEL BASED, AND S–H MODEL BASED PREDICTED
PERFORMANCES AND SPICE SIMULATED PERFORMANCES AT THE FINAL
DESIGN POINT OF THE OP-AMP IN EXAMPLE 3
(60)

For a long-channel transistor drain current, is proportional

to , i.e., in (58) is two. However, in [31] it is shown that the
value of is dependent on the channel length and in the small
channel length regime, the value is quite close to one. We found
that the parameter in the equation for also depends on
the channel length. For the two parameters the following models
are taken:

(61)

The transconductance [in (58)] is obtained by taking the voltages of the transistors are different. The current through the
derivative of with respect to . input and the second stages are, respectively,
Now we consider the drain conductance . In the S–H
model, . Using this information, we have taken the
model of which is given in (58). In this model, and
are constant while the term represents the dependencies of
on and . The model of is as follows:
(64)
and

where (62)

and are second-order polynomials of

and the effective gate to source voltages of the transistors are
The term in (58) represents the dependency of on .
Over the saturation region of operation, the function can be well
approximated by an exponential function. The model we have (65)
taken for is

C. Simulation Results
Example 3: In this example, the topology is the same as that
in Example 1 (Section VI-A), i.e., simple op-amp. The specifi-
where and are second-order cation set of this example is shown in Table IX. Like the other
polynomials of and example, here the op-amp was designed by starting from the fol-
(63) lowing four initial design points:
The model parameter extraction procedure is given in [20]. 1) small size of all the transistors;
2) small size of the input-stage transistors and large size of
B. Application of the New Model second-stage transistors;
The new MOS model, which is described in the last section, 3) large size of the input-stage transistors and small size of
is used for op-amp design automation. In the new model, (58) the second-stage transistors;
provides the first-order models of , and as function 4) large size of all the transistors. Further, the op-amp was
of biases and transistor sizes. These first-order models are PoP designed by starting from 25 random initial guesses that
functions. This property of the new model helps to use it in are uniformly distributed in the space of design variables.
design optimization through sequential convex programming. The optimization statistics for case 1) is given in Table X. In
In the sequential design optimization, (58), which provides first- the table, the maximum constraint violation reported for each
order models, is used within the main convex programming (CP) iteration is reported by the SQP in MATLAB at the start of the
optimization. For the iterative to update from one CP to the next solution process corresponding to that iteration. Note that all the
CP in the optimization, the higher order effects are captured by specs have been normalized to have a value of one. Thus, a con-
using (57) and (59)–(63). With the new model, all the design straint violation of two is a 200% violation, while a constraint
equations provided in Section V remain the same. However, the violation of 0.002 is a 0.2% violation. The final solution design
expressions of the drain currents and effective gate-to-source point is given in the Table XI.
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 35

TABLE X TABLE XII

OPTIMIZATION STATISTICS FOR EXAMPLE 3 BY STARTING FROM SMALL SIZE SPECIFICATIONS, NEW MODEL BASED, AND S–H MODEL BASED PREDICTED
OF ALL THE TRANSISTORS IN THE OP-AMP PERFORMANCES AND SPICE SIMULATED PERFORMANCES AT THE FINAL
DESIGN POINT OF THE OP-AMP IN EXAMPLE 4

TABLE XIII
OPTIMIZATION STATISTICS IN EXAMPLE 4 BY STARTING FROM SMALL SIZES
OF ALL THE TRANSISTORS IN THE OP-AMP
TABLE XI
THE OPTIMAL DESIGN POINT IN EXAMPLE 3

For each of the other 28 cases, the required CPU time is

less than 60 s and the required number of iterations are 8 to
11. Note that in this example, the required CPU time is more
than that in Example 1. All the solution points, which are ob-
tained by starting from the four deterministic and 25 random point with all transistors at minimum sizes are provided in
initial guesses, were compared. It is found that in these solu- Table XIII. Note that because of the circuit complexity, the
tion points, the maximum deviation in ’s and ’s are, re- required iteration number is high. However, it converges
spectively, 0.0077 m and 0.0032 m. The performance of the steadily to the final solution point and the required CPU time
circuit at the final design point was predicted using the new is less than three minutes. The final solution design point is
model, SPICE simulation, and the S–H model, which are given given in the Table XIV. The design optimizations, by starting
in Table IX. Note that the predictions based on the new model with the other initial design points, also converge to the final
and SPICE simulation are quite close to the specifications, while design point within five minutes CPU time and with less than
the S–H model based prediction overestimates some of the per- 30 iterations. All the solution points, which are obtained by
formance metrics. starting from the four deterministic and 25 random initial
Example 4: In this example, the topology is the most guesses, are very close to each other. It is found that in these
complex op-amp among the considered class of op-amps. solution points, the maximum deviation in ’s and ’s are,
Both the first and second stages of the op-amp are cascoded. respectively, 0.0135 m and 0.0032 m. The performance of
The design specifications are shown in the second column the circuit at the final design point was predicted using the
of Table XII. Like the other examples, here the op-amp was new model, SPICE simulation, and the S–H model, which are
designed by starting from the four widely varying deterministic given in the Table XII. In the same table, the second column
initial design points and 25 uniformly distributed random provides the specifications. Like the previous example, here the
initial design points. The optimization statistics for the starting predictions based on the new model and SPICE simulation are
36 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

TABLE XIV we would only use a standard model in the “device model” box
THE OPTIMAL DESIGN POINT IN EXAMPLE 4 and the only technology in the “find constants and powers …”
box would be linear-response surface modeling. Specifically,
one would first model the of the current source transistors
and then model the , , and of the other transistors
using the standard device model for function evaluation. While
doing this modeling, the quantities and would be kept
at the values determined in the previous iteration of the SCP.
Note in this context that for a PoP model, if is a PoP func-
tion of , , and , then so is of , , and , i.e.,
and are completely interchangeable. Further, as the it-
erants of the SCP proceed to convergence—as indicated by the
proximity of the solutions and the reduction of the maximum
constraint violation (see Table X)—the region over which the
PoP response surface is built can be shrunk, resulting in a better
fit with the standard model. Note that this approach is reminis-
cent of the iterative simulate-approximate-optimize strategy of
CENTER [35].
There is one cautionary note, however. As the independent
variables are varied in order to build the response surface, it is
important that the device model stays in the domain (e.g., satura-
tion) in which a PoP model is a good approximation. Thus, tradi-
tional fractional-factorial design [33], [34], which places sample
points at the corners of the experiment-design box, should be
avoided. Instead, Latin-hypercube sampling [36] should be used
since it spreads the samples more uniformly in the box [35].
Further, those samples that do not fall in the required regime of
device operation, i.e., saturation, should be discarded.
It should be noted that the approach outlined above, though
promising, has not yet been implemented. It remains to be
quite close to the specifications, while the S–H model based seen whether this approach of using only standard models will
prediction overestimates some of the performance metrics. converge to the same final solution from widely varying initial
guesses, as has been demonstrated in this paper for the –
model.
VIII. DISCUSSION The next issue that we will address is that of restricting the
The results presented so far have left a few questions unan- devices to operate in the saturation region. We do this because
swered, while also opening up new avenues of research. We ad- in standard CMOS op-amp design, the mosfets that are used as
dress these points in this section. loads or amplification devices are biased in the saturation re-
The key concept in our approach is that the CMOS op-amp gion for, among other reasons, the low that is achievable in
sizing problem can be formulated as a sequence of (convex) this region [28]. There are of course specific exceptions to this
geometric programs. This is achieved by modeling , , rule, e.g., when a parallel connection of an NMOS and PMOS
and as a PoP function of the transistor sizes and the bias device is used to build a resistor, or the common mode feedback
current at a “relaxed” estimate of the dc operating point [see transistor in a fully differential two-stage op-amp. At this time,
(1) and (58)]. Because of this iterative formulation, as the it- we handle these situations as special cases that we can fit into
erants proceed and approach convergence, the coefficient and our paradigm. However, the question remains as to whether the
powers of the first-order PoP model are made accurate via the SCP approach can be applied to MOS circuits without any con-
use of second order model functions, such as those described cern about the region of operation of the individual devices, or
in (59)–(63) for the – model that we have introduced. While even for bipolar circuits. The answer depends on how well the
such an approach maintains convexity in each optimization step, iterative model-optimize approach that we have outlined in the
the need for a special model (e.g., – ) that will have to accu- context of using a standard model can be extended to these sit-
rately mimic a standard model (e.g., BSIM3) could be viewed as uations. Essentially, what we require is that the derived device
a drawback. Therefore, we now outline an approach to sequen- parameters, e.g., and , be modeled as PoP functions of the
tial convex programming (SCP) that does not use any special independent design variables and that these PoP models become
models. accurate approximations of the original device models as the it-
The key observation is that with a transformation, a PoP erants converge.
function becomes a linear function whose coefficients can be A requirement of our approach is that the performance
easily determined using linear least-squares curve fitting tech- constraints have to be manipulated into the form of an upper
niques [33]. Thus, in the inner loop of the flow graph of Fig. 4, bounded posynomial. While we have shown that this is indeed
MANDAL AND VISVANATHAN: CMOS OP-AMP SIZING USING A GEOMETRIC PROGRAMMING FORMULATION 37

possible for a large class of CMOS op-amps and a large number to the analog circuit sizing problem within the framework of
of specifications, this is nevertheless a restriction. Indeed, our sequential convex programming would be a worthwhile area of
investigations with complex folded-cascode op-amps show research.
that it is very simple to cast performance specifications as Finally, since we have formulated the performance metrics for
signomials (unlike posynomials, these are sums and differences the class of two-stage op-amps in the notation of [29], we trust
of PoP terms). We would then have a “signomial programming that it is obvious that their results can be enhanced by replacing
problem” (where the direction of the constraint inequality is their integer-nonlinear programming by the more powerful in-
not an issue). This can be solved as a sequence of geometric teger-convex programming.
programs [23] and would fit directly into our existing sequential
convex programming approach. IX. SUMMARY
The various issues discussed so far in this section have also
served to highlight the superiority of our approach over that An efficient technique for sizing CMOS op-amps has been
of GPCAD [16]. Since the approach in [16] is that of approx- proposed. In this method, the op-amp sizing problem is formu-
imating and solving the CMOS op-amp sizing problem as a lated as a sequential convex programming problem. Such a for-
single geometric program, i.e., one pass of the loop shown in mulation has two major advantages as enumerated below.
Fig. 4, their approach could be viewed as a first-order-approx- 1) Since the convex programming problem is very well un-
imate globally optimal solution. Further, none of the possible derstood, it is very straight forward to solve it in a robust
extensions discussed so far in this section are applicable to [16], and computationally efficient manner.
since these ideas are predicated on our sequential convex pro- 2) The sequence of solutions generated is a sequence of
gramming formulation. Thus, it would be fair to say that our global optimal of convex programming subproblems.
approach is more general than that of [16]. Intuition therefore suggests that the point to which this
Note also that the SCP provides a sequence of iterants that sequence converges is the globally optimal solution
are globally optimal solutions of convex subproblems. Further, of the original problem. This belief is supported by
our computational results show that this sequence converges to experimental results, where it is shown that the method
the same final solution for widely varying initial guesses. While converges to the same final design point for widely
these are very appealing features, it should be borne in mind varying initial guesses.
that this does not guarantee that the final solution is the global The method has been prototyped in MATLAB and applied to
optimum of the original sizing problem. a number of two stage CMOS op-amps. The experimental re-
Not withstanding the extensions described above, there may sults highlight the robustness and computational efficiency of
remain certain performance metrics like settling time, which the technique. Further, the optimal design point is qualitatively
cannot be modeled as a suitable analytic function. Nor is it likely similar to one that would be picked by an expert designer.
that symbolic analysis [14], which is excellent for generating ex- Although in the initial version of the implementation the S–H
pressions for ac performance metrics, would provide a solution. MOS model was used, in the short channel length regime the ac-
While it is possible to meet a given settling time specification by curacy of performance prediction was not very good. To address
suitably constraining slew rate, unity gain frequency, and phase this problem, in the second version of the prototype a new MOS
margin (which are modeled as posynomials), this nevertheless model, called the – model, has been used to replace the S–H
requires user intervention. Indeed, as has been observed in [8], model. A number of op-amps were sized using the new model.
even the one-time user effort required to derive analytic expres- The experimental results show the accuracy of performance pre-
sions of performance metrics (for a new op-amp) is a barrier to diction at the final design point. Further, the results demonstrate
the widespread use of techniques such as ours. This is a bar- that with the accurate model the sizing technique maintains its
rier worth breaching in order to bring the major advantages of robustness.
convex optimization into a truly automated circuit-sizing tool.
The best bet for achieving this is response surface modeling.
The idea is essentially the same as the one previously outlined ACKNOWLEDGMENT
for accommodating standard device models. The difference The authors would like to thank J. Fishburn for bringing the
is that instead of building PoP functions based on model work of Hershenson et al. to their attention and K. Singhal for
evaluations, we would need to build posynomial or signomial sharing her insights into response surface modeling. The author
models of the circuit performance metrics (e.g., settling time) would also like to thank the anonymous reviewers for their com-
in terms of the independent design variables using standard ments that have helped improve this paper.
circuit simulation for the “experimental measurements.” Unfor-
tunately, the rich history of response surface modeling [34] has
REFERENCES
been focused on modeling polynomials rather than the more
complex posynomials or signomials. In particular, determining [1] L. R. Carley, G. G. E. Gielen, R. A. Rutenbar, and W. M. C. Sansen,
“Synthesis tools for mixed-signal ICs: Progress on frontend and backend
the optimal number of PoP terms for the posynomial model and strategies,” in Proc. DAC, June 1996, pp. 298–303.
picking the correct powers for these terms without requiring [2] M. J. S. Smith, C. Portmann, C. Anagnostopoulos, P. S. Tschang, R.
prior analytical knowledge of the physical quantity being Rao, P. Valdenaire, and H. Ching, “Cell libraries and assembly tools
for analog/digital CMOS and BiCMOS application-specific integrated
modeled are difficult problems for which only limited results circuit design,” IEEE J. Solid-State Circuits, vol. 24, pp. 1419–1432,
are available [37]. Extending these results and applying them Oct. 1989.
38 IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL. 20, NO. 1, JANUARY 2001

[3] F. El-Turky and E. E. Perry, “BLADES: An artificial intelligence ap- [27] R. Gregorian and G. C. Temes, Analog MOS Integrated Circuits for
proach to analog circuit design,” IEEE Trans. Computer-Aided Design, Signal Processing. New York: Wiley, 1986.
vol. 8, pp. 680–692, June 1989. [28] P. E. Allen and D. R. Holberg, CMOS Analog Circuit Design. New
[4] R. Harjani, R. A. Rutenbar, and L. R. Carley, “OASYS: A framework York: Holt, Rinehart and Winston, 1987.
for analog circuit synthesis,” IEEE Trans. Computer-Aided Design, vol. [29] P. C. Maulik, L. R. Carley, and R. A. Rutenber, “Integer programming
8, pp. 1247–1265, Dec. 1989. based topology selection of cell-level analog circuits,” IEEE Trans.
[5] M. G. R. Degrauwe et al., “IDAC: An interactive design tool for Computer-Aided Design, vol. 14, pp. 401–412, Apr. 1995.
analog CMOS circuits,” IEEE J. Solid-State Circuits, vol. SC-22, pp. [30] A. Grace, “Optimization Toolbox User’s Guide,” The MathWorks Inc.,
1106–1115, Dec. 1987. Nov. 1992.
[6] , “Toward an analog system design environment,” IEEE J. Solid- [31] T. Sakurai and A. R. Newton, “Alpha-power law MOSFET model and
state Circuits, vol. 24, pp. 659–671, June 1989. its applications to CMOS inverter delay and other formulas,” IEEE J.
[7] W. Nye et al., “DELIGHT.SPICE: An optimization-based system for the Solid-State Circuits, vol. 25, pp. 584–594, Apr. 1990.
design of integrated circuits,” IEEE Trans. Computer-Aided Design, vol. [32] P. V. Wolf et al., An Introduction to the NELSIS IC Design
7, pp. 501–519, Apr. 1988. System. Delft, The Netherlands: Delft Univ. Technol., 1990.
[8] E. S. Ochotta, R. Rutenbar, and L. R. Carley, “Synthesis of high-perfor- [33] P. Mandal and V. Visvanathan, “Macromodeling of the a.c. character-
mance analog circuits in ASTRX/OBLX,” IEEE Trans. Computer-Aided istics of CMOS op-amps,” in Proc. Int. Conf. Computer-Aided Design,
Design, vol. 15, pp. 273–294, Mar. 1996. 1993, pp. 334–340.
[9] F. Medeiro et al., “A statistical optimization based approach for auto- [34] G. E. P. Box and N. R. Draper, Emperical Model Building and Response
mated sizing of analog cell,” in Proc. Int. Conf. Computer-Aided Design, Surfaces. New York: Wiley, 1987.
1994, pp. 594–597. [35] K. Singhal, C. C. McAndrew, S. Nassif, and V. Visvanathan, “The
[10] M. Krasnicki, R. Phelps, R. Rutenbar, and L. R. Carley, “Maelstrom: CENTER design optimization system,” AT&T Tech. J., vol. 68, pp.
Efficient simulation-based synthesis for custom analog cells,” in Proc. 77–92, May/June 1989.
36th DAC, 1999, pp. 945–950. [36] R. L. Iman, J. C. Helton, and J. E. Campbell, “An approach to sensitivity
[11] R. Phelps, M. Krasnicki, R. Rutenbar, L. R. Carley, and J. R. Hellums, analysis of computer models, Part I: Introduction, input variable selec-
“ANACONDA: Robust synthesis of analog circuits via stochastic pat- tion and priliminary variable assesment,” J. Qual. Technol., vol. 13, pp.
tern search,” in Proc. IEEE Custom Integrated Circuits Conf., 1999, pp. 174–183, 1981.
567–570. [37] N. R. Draper and H. Smith, Applied Reression Analysis. New York:
[12] H. Y. Koh, C. H. Sequin, and P. R. Gray, “OPASYN: A complier for Wiley, 1966.
CMOS operational amplifiers,” IEEE Trans. Computer-Aided Design,
vol. 9, pp. 113–125, Feb. 1990.
[13] G. G. E. Gielen, H. C. C. Walscharts, and W. M. C. Sansen, “Analog cir-
cuit design optimization based on symbolic simulation and simulated an-
nealing,” IEEE J. Solid-State Circuits, vol. 25, pp. 707–713, June 1990.
[14] , “ISAAC: A symbolic simulator for analog integrated circuits,”
Pradip Mandal received the B.E. degree in elec-
IEEE J. Solid-State Circuits, vol. 24, pp. 1587–1597, Dec. 1989.
tronics and telecommunication engineering from
[15] P. C. Maulik, L. R. Carley, and D. J. Allstot, “Sizing of cell level analog
the Bengal Engineering College, Shibpur, in 1989.
circuits using constrained optimization techniques,” IEEE J. Solid-State
He received the M.E. degree and the Ph.D. degree
Circuits, vol. SC-28, pp. 233–241, Mar. 1993.
in electrical communication engineering from the
[16] M. M. Hershenson, S. P. Boyd, and T. H. Lee, “GPCAD: A tool for
Indian Institute of Science, Bangalore, India, in 1991
CMOS op-amp synthesis,” in Proc. Int. Conf. Computer-Aided Design,
and 1999 respectively.
1998, pp. 296–303.
Dr. Mandal started his career in 1997 with
[17] , “Automated design of folded-cascode opamps with sensitivity
Motorola India Electronics, Bangalore, India, where
analysis,” in Proc. 1998 IEEE Int. Conf. Electronics, Circuits, Syst., vol.
he has worked on development of design automation
1, pp. 121–124.
tools for VLSI circuit and SAW filter. In 1998 he
[18] M. M. Hershenson, S. S. Mohan, S. P. Boyd, and T. H. Lee, “Opti-
joined Philips Semiconductors, Bangalore, where he is a Senior Technical
mization of inductor circuits via geometric programming,” in Proc. 36th
Specialist in the area of I/O circuit design. His research interests are analog
DAC, 1999, pp. 994–998.
circuit design and design automation for analog and mixed-signal integrated
[19] M. M. Hershenson, A. Hajimiri, S. S. Mohan, S. P. Boyd, and T. H.
Lee, “Design and optimization of LC oscillators,” in Proc. Int. Conf.
circuits.
Computer-Aided Design, 1999, pp. 65–69.
[20] P. Mandal, “Synthesis and biasing of CMOS op-amps,” Ph.D. disserta-
tion, ECE Dept., Indian Inst. Sci., Bangalore, India, 1997.
[21] P. Mandal and V. Visvanathan, “A new approach for CMOS op-amp
synthesis,” in Proc. 12th Int. Conf. VLSI Design, 1999, pp. 189–194.
[22] C. Taumazou and C. A. Markis, “Analog IC design automation: Part V. Visvanathan (S’77–M’82) received the B.Tech. degree in electrical engi-
I—Automated circuit generation: New concept and methods,” IEEE neering from the Indian Institute of Technology, Delhi, India, the M.S.E.E. de-
Trans. Computer-Aided Design, vol. 14, pp. 218–238, Feb. 1995. gree from the University of Notre Dame, IN, and the Ph.D. degree in electrical
[23] R. J. Duffin, E. L. Peterson, and C. Zener, Geometric Program- engineering and computer sciences from the University of California, Berkeley.
ming—Theory and Application. New York: Wiley, 1967. He was with AT&T Bell Laboratories, Murray Hill, NJ, and has been on the
[24] J. P. Fishburn and A. E. Dunlop, “TILOS: A posynomial programming faculty of the University of Maryland, College Park, and the Indian Institute of
approach to transistor sizing,” in Proc. Int. Conf. Computer-Aided De- Science, Bangalore. He is currently with Cadence Design Systems, Allentown,
sign, 1985, pp. 326–328. PA, where he is an Architect in the area of core algorithms for the simulation
[25] S. S. Sapatnekar, V. B. Rao, P. M. Vaidya, and S. M. Kang, “An exact of analog and mixed-signal integrated circuits. He has two pending U.S. patents
solution to the transistor sizing problem for CMOS circuits using and has authored or coauthored over 50 papers in archival journals and confer-
convex optimization,” IEEE Trans. Computer-Aided Design, vol. 12, ence proceedings in various areas of VLSI design and electronics design au-
pp. 1621–1634, Nov. 1993. tomation (EDA).
[26] H. Onodera, H. Kanbara, and K. Tamaru, “Operational-amplifier com- Dr. Visvanathan received the AT&T Bell Laboratories Distinguished Tech-
pilation with performance optimization,” IEEE J. Solid-State Circuits, nical Staff Award in 1989 and the Honorable Mention Award at the 6th. Int.
vol. 25, pp. 466–473, Apr. 1990. Conf. on VLSI Design, Bombay, India, in 1993.