Professional Documents
Culture Documents
Peter B. Aronhime, F. William Stephenson (Auth.) - Analog Signal Processing-Springer US (1994) PDF
Peter B. Aronhime, F. William Stephenson (Auth.) - Analog Signal Processing-Springer US (1994) PDF
by
Peter B. Aronhime
University of Louisville
F. W. Stephenson
Virginia Polytechnic Institute
& State University
A Special Issue of
ANALOG INTEGRATED CIRCUITS
AND SIGNAL PROCESSING
An International Journal
Vol. 6, No. 3 (1994)
Statistical Design
Global Design of Analog Cells Using Statistical Optimization Techniques ............................. .
. . . . . . . • . . . . . .F. Medeiro, R. Rodrfguez-Macfas, F. V. Fernandez, R. Dominguez-Castro, J. L. Huertas and
A. Rodrfguez-Vdzquez 3
DDA·Based Circuits
A Multiple Input Differential Amplifier Based on Charge Sharing on a Floating-Gate MOSFET .......... .
. . . . . . . . . . . . . . . . . . . • . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .K. Yang and A. G. Andreou 21
Design and Applications of a CMOS Analog Multiplier Cell Using the Differential Difference Amplifier ...
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . S.-C. Huang and M. Ismail 33
Optimization Techniques
On the Optimal Design of Switched-Capacitor Filter Circuits for Analog and Mixed-Signal Integrated Circuit
Realization ......................................................... N. C. Gustard and R. E. Massara 43
Filter Design
A 7.2 GHz Bipolar Operational Transconductance Amplifier for Fully Integrated OTA-C Filters .......... .
. . . . . . . . . . . . . . . . . . . . . . . . . • . . • • . . • . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M. Atarodi and J. Choma, Jr. 67
Current-Mode Synthesis Using Node Expansion Techniques ................... M. Desai and P. Aronhime 19
Macromodeling
Nonlinear Macromodeling with AWE ...................................... R. J. Trihy and R. A. Rohrer 89
ISBN 978-1-4419-5147-2 ISBN 978-1-4757-4503-0 (eBook)
DOI 10.1007/978-1-4757-4503-0
Editorial
This special issue was conceived in Detroit during the 36th Midwest Symposium on Circuits and Systems. We
were asked by Editor-in-Chief Mohammed Ismail to select a collection of analog papers from those presented at
the conference and seek to have the authors develop them further for publication in this special issue.
It seemed to us quite appropriate that we undertake this task as a technical challenge as well as a means of
recognizing the unique role which the Midwest Symposium has played for the past thirty-nine years. It is the oldest
conference devoted to circuits and systems and affords a convenient form for the presentation of papers by graduate
students as well as by faculty and engineers from the non-academic arena. Furthermore, despite its somewhat
restrictive title, the Symposium has been held in many locations outside its original Midwest home. In recent years,
conferences have been held in Puebla (Mexico), Calgary (Canada), Monterey (California), Louisville (Kentucky),
Lafayette (Louisiana) and Washington, D.C. Future venues include Rio de Janeiro (Brazil) and Davis (California).
Of course, all of these conferences were, or will be, located in the midwest section of the town or city I
We carefully reviewed many papers presented in Detroit and decided that we must restrict our pool to those of an
analog nature. Furthermore, the papers needed to address something broader in scope and on a topic lending itself
to a full length journal paper. Our final selection loosely falls within our chosen title of Analog Signal Processing,
but we ask the reader's indulgence on this issue. More important to us has been the need to select a solid group of
analog-related papers which represent the breadth of topics covered in a variety of sessions in Detroit.
Several papers address the questions of cell design for CMOS applications. Medeiro et al. describe statistical
optimization for the design of complex analog cells while Yang and Andreou develop the lossless property of an
MOS floating gate into circuits having potential use as analog signal processing building blocks. By contrast, Huang
and Ismail discuss the design of a differential difference amplifier with application as a basic block in a simple
four-quadrant multiplier. In a further paper on analog design, Gustard and Massara use a switched-capacitor case
study to illustrate the usefulness of numerical optimization techniques in the generation of silicon-level layout for
analog functional modules from high-level specifications.
Filter design is featured in three of the contributions, but in distinctly different ways. de Queiroz and CaiOba discuss
an optimal approach to the overdesign of gain-shaping filters, while Atarodi and Choma present a transconductaJtce
amplifier for use in active integrated filters operating at cut-off frequencies greater than 200 MHz. Finally, Desai
and Aronhime propose a method for the node expansion of passive prototypes from which may be derived a series
of oscillators and active filters.
The paper by Trihy and Rohrer is quite general and offers an approach for the simulation of nonlinear (linearized)
circuits. Using an asymptotic waveform evaluation, the paper offers strategies for macromodeling nonlinear circuits.
In conclusion, we wish to acknowledge the assistance of Dr. Jatindar Bedi, ECE Department, Wayne State
University, who served as Publications Chairman of the 36th Midwest Symposium. Dr. Bedi was extremely helpful
in providing us with pre-publication copies of several papers. In addition, we acknowledge the work of Dr. Michael
Polis (currently Dean of Engineering at Oakland University), General Chairman of the 36th Midwest Symposium
and other Operating Committee members, Dr. Harpeet Singh and Dr. Feng Lin(both at Wayne State University), Ben
Behera (Aero Service Corp.), Dr. Majid Ahmadi (University of Windsor), Dr. Nader Boustany (General Motors),
and Dr. Magdy Bayoumi (University of Southwestern Louisiana) for their efforts in organizing the Symposium.
Peter B. Aronhime
F. W. Stephenson
178 Aronhime and Stephenson
~ ;)~'J\11',
the University of Durham {UK) and a Ph.D. in Electrical En-
,~,. i
gineering, 1965, at the University of Newcastle upon Tyne
{UK). Dr. Stephenson has been Dean of the College of Engi-
neering at Virginia Tech since June 1994. From 1990-1994
he was Professor and Head of The Bradley Department of
Electrical Engineering at the same school. Previously, from
1986-1990, he was Associate Dean for Research and Gradu-
ate Studies in the College ofEnglneering. In addition, he was
t~ the founding Director of the Virginia Tech Hybrid Micro-
t. electronics Laboratory and co-founder of the Virginia Tech
Student Chapter of ISHM, the largest in the nation with 140
Peter Aronhlme received the B.E.E. degree from the Uni- members: He has held industrial appointments with Welwyn
versity of Louisville in 1962 and M.S. and Ph.D. degrees Electric and the Microelectronics Division ofElectrosil, both
from the Colorado State University in 1964 and 1971. He has in England. While with Electrosil, he worked on the applica-
held engineering positions with Bell Laboratories in Winston- tions of both monolithic and hybrid integrated circuits. His
Salem, NC, and Hughes Aircraft in Fullerton, CA, and he has main research interests are in the areas of hybrid microelec-
held academic positions with Tri-State University In Angola, tronics, RC active and switched-capacitor filter design. Prior
IN, and Jllinois Institute of Technology in Chicago. In 1976 to joining Virginia Tech In 1978, he taught at the Universities
he joined the University of Louisville, where he is Professor of Hull {UK) and Rochester, NY, where he was the RT French
of Electrical Engineering and Coordinator of the Computer Visiting Professor in 1976-77. He Is the co-author of Lin-
Science and Engineering Ph.D. Program. In 19g7 he was a ear Microelectronic Systems, Macmillan, 1973, and Active
visiting professor at Colorado State University. His research Filters for Communications and Instrumentation, McGraw
interests include network theory, computer-aided circuit anal- Hill {UK), 1979, and the editor of RC Active Filter Design
ysis/design/testing, modeling and instrumentation. Dr. Aron- Handbook, John Wiley & Sons, 1985. He has published re-
hirne is a member of Eta Kappa Nu, Sigma Xi, Phi Kappa search results in such journals as IEEE Transactions (CHMT
Phi, and Omicron Delta Kappa. and CAS), Tnt. Journal of Circuit Theory and Applications,
Microelectronics, and others. Dr. Stephenson is a member
of HKN, Tau Beta Pi, ODK, a Fellow of the Institution of
Electrical Engineers (lEE), and a Fellow of the Institute of
Electrical and Electronic Engineers (IEEE).
2
Analog Integrated Circuits and Signal Processing, 6, 179--195 (1994)
0 1994 Kluwer Academic Publishers, Boston.
angelllcnm.us.e~
Abstract. We present a methodology for automated sizing of analog cells using statistical optimization in a simula-
tion based approach. This methodology enables us to design complex analog cells from scratch within reasonable
CPU time. Three different specification types are covered: strong constraints on the electrical performance of the
cells, weak constraints on this performance, and design objectives. A mathematical cost function is proposed and a
bunch of heuristics is given to increase accuracy and reduce CPU time to minimize the cost function. A technique
is also presented to yield designs with reduced variability in the performance parameters, under random variations
of the transistor technological parameters. Several CMOS analog cells with complexity levels up to 48 transistors
are designed for illustration. Measurements from fabricated prototypes demonstrate the suitability of the proposed
methodology.
3
180 Medeiro, Rodrfguez-Macfas, Fernandez. Dom(nguez-Castro, Huertas and Rodrfguez-Vazquez
be manually provided. Hence, the methodology is only analog output buffer, which were sized using the pro-
partially open. Furthermore, the level of complexity for posed methodology, fabricated in different CMOS
AC automatic modeling is limited by the capabilities technologies, and whose performance was corrobo-
of symbolic analysis tools (currently, about 15 MOS rated from actual silicon prototypes. The proposed
transistors using high-frequency MOST models and technique is also extended to design for low variability
workstation standard configurations). Consequently, incorporating mismatching information in the design
this approach is not the most suitable for the auto- procedure. This is illustrated in the design of a CMOS
mated sizing of complex analog building blocks (for folded-cascode operational amplifier.
instance, fully-differential opamps), or for applications
where large signal specifications play a major role, for 2. Some Generalities on Optimization-Based
instance, oversampled modulators for high resolution Sizing
AID converters [ 13].
Whether closed or open, equation-based systems Analog sizing is a constructive procedure to map cell
have a common drawback in that sizing is carried out specifications into design parameter values. Design
using simplified analytical descriptions of the blocks. specifications are given a broad meaning here which
Hence, manual fine-tuning using an electrical sim- includes constraints on the electrical performance pa-
ulator and detailed MOS transistor models may be rameters of the cell as well as design objectives. Let us
necessary once rough automated sizing is completed. consider for illustration purposes the output buffer of
This drawback is overcome in the so-called simulation- figure 1, one of the examples covered in this paper. A
based systems [14], which also reduce sizing to a con- possible specification set for this circuit could include
strained optimization problem, and aim to solve it by constraints on its DC gain (Ao > target), input capaci-
following an iterative procedure built around an elec- tance (C;n <target), 3-dB frequency (f3 d 8 >target),
trical simulator. No design equations are required in and output voltage range (target < OS < target), in
addition to the design objective of minimum possible
these approaches; the design parameters are updated
power consumption. With regards to the design param-
at each iteration based on the results provided by sim-
eters, these include transistor dimensions and passive
ulations with detailed transistor models. Thus, they
component values.
are intrinsically open. A representative example of
In a generic circuit, the design parameters can
this methodology is DELIGHT. SPICE [14] where DE-
be viewed as components of a vector xT
LIGHT (a general algorithmic optimization tool) and
{x 1, x 2 , ... , x N} defining a multidimensional de-
SPICE are combined. Also, advanced electrical simu-
sign space. Thus, performance parameters and the
lators, like HSPICE [ 15], incorporate optimization rou-
tines. However, the optimization routines in both tools
search for a local solution, and consequently are typ-
ically used to redesign cells whose performance spec-
ifications are close to the design goals (for instance,
technology updating of a cell library), but are inappro-
priate to size analog cells from scratch. This is a real
challenge in analog design automation and requires the
development of other techniques.
This paper presents a simulation-based approach for
global sizing of arbitrary topology analog cells us-
ing statistical optimization. We demonstrate that by
combining proper cost function formulation and in-
novative optimization heuristics complex cells are de-
signed starting from arbitrary initial points, within rea-
sonable CPU times and with no designer interaction
required - a very appealing feature for ASIC appli-
cations. We present results obtained for two fully-
differential CMOS opamps, a comparator and an Fig. 1. A CMOS output buffer.
4
Global Design of Analog Cells using Statistical Optimization Techniques 181
features involved in design objectives are given as introduces additional complications to the analyti-
functions of x; referring again to the example cal solution process.
of figure I: A0 (x),C;n(x).fadB(x),OS(x), and Due to these difficulties, analog circuits are most
Power(x). Then the problem of sizing is formulated as conveniently sized by using an iterative, dynamic pro-
a constrained optimization problem; in particular, for cess. This concept is illustrated in figure 2: starting
the case of the buffer of figure I , from an initial design parameter estimate, x 0 , a dis-
crete sequence of movements (represented generically
minimize Power (x) as ~xn) is performed through the design parameter
A0 (x) > target space until an equilibrium solution point x• is found.
subjected to C1n(x) < target (I) A key component of this iterative loop is process
!adn(x) > target management: the calculation of the direction and mag-
target < OS(x) < target nitude of the movement ~Xn to be made at each itera-
tion. In manual design, ~Xn is chosen by the designer
Unfortunately, even for elementary analog cells like based on his/her knowledge of the circuit structure be-
that shown in figure I, the analytical solution to the ing sized - a difficult and time-consuming task even
sizing problem is not possible due, among other fac- for experienced analog designers. In automated de-
tors, to the following: sign, the selection of ~Xn must be performed by the
• Design equations, i.e., functional relationships computer based on the evaluation of some critical cir-
among performance parameters and design objec- cuit performance indicators. A convenient approach to
tives on one hand, and design parameters on the do this is to recast the problem formulation as a cost
other, are very difficult to obtain accurately. function 4i(x) which quantifies the degree of achieve-
• These relationships are typically highly nonlinear ment of the design goals and their relation to the design
and, consequently, unsolvable analytically. A fur- parameters. Thus, the parameter updating to be done
ther complication arises due to the large dimensions for the subsequent iteration ~Xn is selected at each
of the design and the specification spaces. iteration using functional analysis data of 4i(x). This
• The need to minimize some functions forces the cal- approach also provides simple and accurate criteria to
culation of first and second derivatives and hence finish the sizing process at points where the cost func-
tion is either maximized or minimized.
In the simplest case, ~Xn is calculated by using
pieces of information calculated only at Xn · However,
as demonstrated in this paper, the use of additional in-
formation from previous points, at time instances n- 1,
n - 2, etc., may produce more robust solutions of the
sizing problem, in the sense of yielding cells whose
specifications have lower variability when statistical
variations of the technological parameters are taken
into account. In this more general case, the updating
@! YF.S:
process is described as a high-order nonlinear discrete-
time system,
(a) =···o;;~·g;;p.;x.c;s·~ianagen;cn.················
manual managemcr·u
~~ au1oma1ed management
performance evaluations in figure 2 (equivalently, the
calculation of performance specification values and the
•non••••••••••••••••,.,.•••••• ••••••• values of the features involved in the design equations
as functions of x) are made using electrical simulation
(b ) PARAMETER UJ"'ATINC
and detailed transistor models to guarantee accuracy
Fig. 2. Iterative analog cell sizing: (a) general concept; (b) manual of the sizing process. Many different alternative im-
and automated design updating management. plementations of figure 2 are possible depending on:
5
182 Medeiro, Rodrfguez-Macfas, Fernandez. Domfnguez-Castro, Huertas and Rodrfguez-Vdzquez
a) formulation ofthecost function itself, b) the updating allowed. Hence, if any setting of the design parame-
procedure. Two major alternatives can be roughly iden- ters (equivalently, any point of the design parameter
tified, depending of the functional structure of S[ •) space) does not satisfy one strong restriction, it must
in (2): be rejected immediately.
• Deterministic, incremental techniques where .6x,. • Weak restrictions: These are the typical perfor-
calculation uses information about the derivatives mance specifications required of analog building
of the cost function. This is an important draw- blocks, i.e., A 0 > 80dB. Unlike strong restrictions,
back since analytical expressions for the cost func- weak restrictions allow some relaxation of the tar-
tion and its derivatives as functions of the design get parameters, making such circuit sizings which
parameters are not commonly available, so that the do not meet such specifications acceptable.
derivatives must be calculated by numerical interpo- • Design objectives: Stated as the minimi7.ation
lation. Another major drawback is that only .6x,. (maximization reduces to this case by either chang-
values which lower the cost function are considered. ing the sign or using the inverse of the function to
Hence, the optimization process is easily trapped in maximize) of some performance features,
local minima, rendering it very suitable only for fine
adjustment of the design. minimize Y>~~, (x) 1 ::; i ::; P (3)
• Statistical techniques, where .6xn is calculated at
random and hence requires no information about the for instance, minimize -GB of an opamp (equiva-
cost function derivatives. lently, maximize GB), where GB denotes the gain-
Parameter updating in deterministic techniques is bandwidth product; or minimize the occupied area
done only in the direction which lowers the cost func- of the circuit.
tion. This makes them very sensitive to the starting Mathematically, the fulfillment of these specifi-
point and hence inadequate for global circuit sizing. cations can be formulated as a multi-objective con-
This is overcome using statistical optimization tech- strained optimization problem,
niques where movements in the design space are done
!
heuristically, following statistical optimization princi- minimize Y>~~, (x}, 1 ::; i ::; P
ples [16]. The price to pay for an independent initial
point is a larger number of iterations and hence longer y.;(x) ;::: Y.; or Ya;(x) ::; Y..;,
CPU times. However, as shown here, proper formu-
lation of the cost function, the movement generator, 1 ::;j::; Q
and the cooling schedule, adapted to the nature of ana- subjected to
log synthesis, palliates the high computational cost and Ywk(x) 2:: Ywk or Ywk(x) ::; Ywk•
thus provides a convenient methodology for global de- 1 ::;k::; R
sign of analog cells.
(4)
6
Global Design of Analog Cells using Statistical Optimization Techniques 183
and Fwlc( •) are defined as vector, .1.x.,, is randomly generated at each iteration.
The value of the cost function is calculated at the new
- L w; log(IY>~o;/), parameter space point and compared to the previous
one. The new point is accepted if the cost function has
i
F.i(Y.J) K.J(Y•J• Ysi) (6) a lower value. Unlike deterministic techniques, it may
also be accepted if the cost function increases, accord-
Fwlc(Ywk) -Kwk(Ywk> Ywk) log ( f::) ing to a probability function,
(8)
where w; (called weight parameters for the design ob-
jectives) is a positive (alternatively negative) real num- depending on a control parameter, T. The random
ber if Y>~oi is positive (alternatively negative), and for character of movements and the statistical acceptance
K,j( •) and Kwk( •) we have of those which increase the cost function enable es-
caping from local minima and hence wide exploration
-oo, if strong of the design space. This probability of acceptance
{ restriction holds changes during the optimization process, being high at
oo, otherwise the beginning (for large T) and decreasing as the system
cools (decreasing T). This is the general concept lying
if weak behind simulated annealing optimization techniques-
restriction holds a process whose name is justified by its analogies to the
otherwise physical annealing in solids [ 16]. The tool proposed
(7) herein incorporates new heuristics relating to both pa-
rameter updating and the cooling schedule itself, as
where k~c (weight parameters assigned to weak restric- explained below.
tions) is a positive (alternatively negative) real num-
ber if the weak specification is of ~ (alternatively :::;) 4.1. Cooling Schedule
type. Weight parameters are used to give priority to the
associated design objectives and weak specifications. Cooling schedule refers to the strategy used to mod-
As shown in the cost function formulation, only rela- ify the temperature while the process evolves. Unlike
tive magnitude of the weight parameters of the same
type makes sense. In (7) weak specifications are as-
sumed positive. Sign criteria is reversed for negative
specifications.
Strong restrictions are checked first at each itera-
tion. If any of them are not met, the corresponding
movement must be rejected. Otherwise, weak restric-
tions are examined. Weak restrictions have priority
over design objectives. If some weak restriction is not
fulfilled, the cost function is built only with their con-
tribution. Hence, if no circuit sizing is able to cover all
weak specifications, the optimization process will pro-
vide results as close as possible. Once all of them are
met, the design objectives are evaluated and their in-
fluence in the cost function guides their maximization
or minimization.
7
184 Medeiro, Rodrfguez-Madas, Fernandez, Dom(nguez-Castro, Huertas and Rodrfguez- Vazquez
classical simulated annealing algorithms [ 16], where T of the initial and final temperature. Figure 4(b) illus-
in (8) decreases monotonically during the process, our trates this type of cooling schedule.
tool uses a composed temperature parameter,
4.1.2 Temperature Scale. As (9) shows, the temper-
T = a(x)T (n) 0 (9) ature scale parameter is a function of the position in
the design parameter space. More specifically, the
where n denotes the iteration count, T 0 ( n) (the nor- scale depends on which region of the parameter space
malized temperature) is a function of n, and a(x) (the
is reached after each movement. This is so done to
temperature scale) is a function of the position in the de- compensate the large differences that may eventually
sign parameter space. Our tool incorporates heuristics appear in the increments of the cost function in the dif-
to choose T0 and a for increased convergence speed, ferent regions. Thus, no temperature definition is used
namely: for those regions where strong restrictions do not hold,
• Non-monotonic and adaptive normalized tempera- due to the fact that any design entering this region is
ture. automatically rejected. On the other hand, in regions
• Use of a nonlinear scale, with different expressions where some weak specifications are violated, temper-
for different regions of the design parameter space. ature is given as
4.1.1 Normalized Temperature. Instead of a conven- T =To lkmaxl ==* a(x) = kmax (12)
tional slow monotonically decreasing temperature [ 18],
a sequence of fast coolings and re-heatings is used. In where kmax is the weight associated to the maximum
circuits with not very demanding specifications, this among the Fw( • )'sin (6), and T 0 is the normalized
enables to obtain feasible designs for low iteration temperature at the current iteration. Finally, if both
counts. Also, for those cases where demanding speci- strong and weak restrictions hold, temperature is given
fications are asked for, we have found that this strategy as
reduces iteration count by, on the average, a factor of
6. l\vo different evolutionary Jaws for the normalized
temperature are incorporated in the tool: exponential
decreasing, and linear decreasing. For illustration pur- where wi is the weight associated to the i-th design
poses figure 4(a) shows an exponential schedule with objective.
eight re-heatings. Initial and final temperatures, num-
ber of coolings, decreasing law and rate, etc., are com-
pletely controlled by the user. An alternative cooling 4.2. Parameter Updating
schedule makes To to change as a function of the per-
centage of accepted movements, Concerning the updating of design parameters three
kinds of heuristics have been adopted:
• Changes in the amplitude of the movement Axn as
T 0 (n) =T 0 (n- 1) + ,6 ( 1- p,~n)) (10) a function of the temperature. In particular, at high
T, large amplitude movements are allowed as they
where p is calculated as are likely to be accepted and favor wide exploration
of the design parameter space. On the contrary, at
number of accepted movements
p= ( 11) low T, acceptance probability decreases and, hence,
only small movements are performed (equivalent to
during the last M iterations, where M is an heuris- fine-tuning the design).
tic variable whose typical value is around 25; f3 in (10) • The possibility of defining logarithmic scales for
controls the rate of temperature change and has a typical independent variables. This has been done be-
value around 0.1; and p, ( n) is a prescribed acceptance cause many design parameters, i.e., transistor sizes,
ratio, which can be fixed or vary with some given law. bias currents, etc., may vary over several decades.
This schedule provides very good results for practi- For instance, a change of 2J.LA in a bias current
cal circuits, rendering the outcome of the optimization does not have the same significance if the previ-
process somewhat independent of the specified values ous bias current value is 5J.LA as if it is 100J.LA;
8
Global Design of Analog Cells using Statistical Optimization Techniques 185
1.0 ~-~--.-----~--.---------,.-------,
0.8
ae o.6
!:!
8.
E
~ 0.4
0.2
0.00
(a)
0 · 0 o~-----::csoo~-~-__,I_.,.ooo.,.__~-....,.t-='so,...,o,..._.~--..J
(b) #iterations
Fig. 4. Cooling schedules: (a) exponential decreasing with re-heatings; (b) adapt;ve temperature with given acceptance mtio.
hence, linear movement of this variable would un- parameter space can be viewed as a collection of
derexplore the low bias current range - a draw- hypercubes. Only movements over vertices of this
back which is overcome by using logarithmic scales. multidimensional grid are allowed, being marked
• Discretization of the design parameter space. Many
when they are visited. Thus, if during the optimiza-
design parameters are already discrete in nature, i.e.,
tion process one vertex is re-visited the correspond-
in many microelectronic technologies transistor di-
mensions can only vary over integer multiples of ing simulation need not be performed. Hence, an
the technology grid. Our discretization consists in important number of simulations is saved. When
making discrete those variables which are contin- this optimization process ends, a local optimization
uous, and define a larger size grid for those vari- is started inside a multidimensional cube around
ables which are already discrete in nature. Then, the the optimum vertex for fine tuning of the design.
9
186 Medeiro, Rodrlguez-Madas, Fernandez, Dominguez-Castro, Huertas and Rodriguez-Vazquez
6.0
c:: 2.0
-~
c::;
§ 0.0
u..
~-2.0
E-
-4.0
- 6.0
Fig. 5. (a) Test function with two variables for optimization heuristics comparison; (b) cross section.
In this local optimization, design variables recover heuristics. Its mathematical structure for an N-
their continuous nature or their original grid size. dimensional case is
Together with these heuristics, large efficiency en-
II cos(xk- d) ,
N N
hancements are also achieved by proper control of the f(x) = K · min{-e-eE._,<x•-dl'
DC electrical simulator routines. For this purpose a k=l
dynamic, adaptive, DC initialization schedule is im-
plemented which uses operating point information of
N
- e-eE•=•'"'•-d)
•
II cos(x~; +d)+ 'Y}
N
10
Global Design of Analog Cells using Statistical Optimization Techniques 187
II
Achieved minimum I
(a)
100
90
~ 80
"'c
.8
~
~
Achieved minimum
(b)
number of variables, and not by structural changes in from these tests are shown in the three-dimensional
the cost function. Figure 5(a) shows this function for plots in figure 6. In order to get better insight into the
two independent variables. A cross-section is shown test results, the plot of the test function is allowed to
in figure 5(b). take only integer values. Hence, the minimum achieved
The heuristics described in Section 4.1 and 4.2 have at each test execution is represented by its closest in-
been tested using the test function in (14) with differ- teger value. The X -axis in figure 6 represents the
ent number of independent variables. The test proce- magnitude of the achieved minimum (its closest in-
dure consisted in the repeated execution of the different teger value). TheY-axis corresponds to the number
heuristics on the test function, starting from random of independent variables in the test function /( • ),
points of the parameter space and with a fixed iteration and the Z -axis represents the percentage of iterations
count. For each of these executions the best achieved that achieved that minimum. Figure 6(a) corresponds
minimum was stored. Experimental results arising to a conventional cooling schedule. It had a single
11
188 Medeiro, Rodrfguez-Mac£as, Fernandez, Dom(nguez-Castro, Huertas and Rodrfguez-Vazquez
cooling with fixed scale in variable movements and during the optimization process as a consequence of
variable Markov chain length [ 16]. For a function with the dependance of their statistical variability with de-
a small number of variables most iterations provided vice area and distances between devices [21]. At each
the global minimum of the function but this percentage iteration, design parameters are updated according to
decreased rapidly when the number of variables was the heuristics in Section 4.2 and device model parame-
increased. Figure 6(b) corresponds to our improved ters are changed according to the statistical distribution
cooling schedule with the same number of iterations. of the technological independent parameters [20]. In
The cooling schedule used had four successive coolings addition to enlarging the number of parameters, a new
and re-heatings, variable scale, and a Markov chain addend is incorporated to the cost function to evaluate
length equal to 1. Most iterations provided the global the sensitivity of performance specifications to device
minimum of the function, even when the number of model parameter variations. Such addend is
12
Global Design of Analog Cells using Statistical Optimization Techniques 189
The cost function at some given iteration must be ment of silicon prototypes of the circuits demonstrate
compared with some previous iteration in order to ac- the feasibility of the approach.
cept or reject the current design parameter movement.
A new comparison methodology is introduced adapted
to the new cost function formulation. Each iteration in 6.1. Fully Differential Class-AB Opamp with
a given cooling is compared with the iteration of equal Dynamic Biasing
ordinal from the best of previous coolings, for accep-
tance or rejection, following the statistical optimization Let us first consider the fully differential opamp of fig-
principles of Section 2. If rejected, the design point of ure 7 [22), intended for a 16bit@16-KHz second order
the best cooling is adopted as new point in the current I:..:l modulator. This class-AB opamp includes dy-
iteration and the optimization process continues. Since namic biasing of the output branches to obtain large
(15) is added to the cost function, the optimization pro- output swing and high slew-rate, and uses a dynamic
cess tends to minimize it. That implies small specifi- common-mode feedback network. These advanced cir-
cation increases in the last M iterations and, hence, cuit strategies, and the complexity of the circuit itself
reduced performance statistical deviations. (it contains 48 transistors) renders its sizing a diffi-
A mathematical test function has also been used here cult task, hard to handle for system level designers.
to prove the method's capability and as a benchmark However, the herein proposed methodology was able
for heuristics refinements. Its analytical struture for an to automatically size the circuit for the intended appli-
N -dimensional case is cation after 1 hour CPU time on a 1OOmips sparcstation,
starting from scratch and with no designer interaction
f(x) = K + K' ~ Isin c~Xk) I required. Table 2 shows the sizing obtained.
The first column in table 1 contains the design goals,
c1T6Xk) I)
which includes a design objective on the power con-
+h (dist 2 (X,Xmtn) + ~ lsin sumption and weak restrictions on the gain-bandwidth
product (GBW), phase margin (PM), input white noise
(18) and output swing (OS). Figure 8 shows the evolution of
the cost function during the optimization process. Note
The first addend in (18), sets the mean value off( •) that the vertical axis contains two regions, separated by
and the second one creates variations around that mean a dashed line. The weak region corresponds to the case
value. Statistical deviations are simulated at the third where any of the weak restrictions is violated, while
addend by means of a random variable, h. The global the objective region corresponds to the case where all
minimum of dispersion is located in X min· Sinusoidal weak restrictions are fulfilled; inside this region the
variations set local dispersion minima. A method to optimization process focuses on the design objectives.
reduce variance will be acceptable, if the final solution Note that a good design (meaning one that fulfills all
is close to X min· the weak restrictions) is obtained after 750 iterations.
For a test example with eight independent variables Simulated results corresponding to the obtained sizing
comprised in the interval Xn E [-0.7, 0.7], a mean are shown in the second column of table 1.
value K = I 00 and sinusoidal variations with an am-
Figure 9 is a microphotograph of a CMOS 1.2J.tm
plitude K' = 10, the new heuristics provides a solution double poly n-well prototype of the fully differential
to a distance of 1. 75 from the global minimum, where opamp. Measured results from the silicon prototype
standard deviation is u = 0.25. A conventional statis- are also shown in table 1. The I:..:l modulator CMOS
tical optimization technique ends in distances around prototype, which was built using this opamp, displayed
12 from X min• where standard deviation is u = 12. a measured resolution of 15.7bits@l6KHz.
6. Practical Results
6.2. Fully-Differential Folded-Cascode Opamp
Proposed techniques have been applied to a wide vari-
ety of analog building blocks. Results are shown for the As a second example, let us consider the folded-
design of two fully-differential opamps, a comparator cascade fully-differential opamp of figure 10, which
and an output buffer. Simulated results and measure- displays the sizes provided by the tool. These sizes
13
190 Medeiro, Rodrfguez-Macfas, Ferndndez, Domfnguez-Castro, Huertas and Rodrfguez-Vdzquez
Ao ~ 70 74.9 74.6 dB
GBW ~ 20 19.7 19.4 MHz
PM ~ 60 63.3 6S 0
Table 2. Sizing for the opamp of figure 7. of CPU time on a 100mips sparcstation. Simulation
results for the sized circuit are shown in the second
Mt,2 149.2/2.2 J.&m M21,22 48.2/2.2 pm column of table 3. The opamp has been integrated in
M3,4 22.0/2.2 M23,24 42.8/2.2 a CMOS 1.2mm double poly n-well technology. Ex-
M5,0 80.4/2.2 M2s,20 6.2/2.2
perimental results are given in the third column of ta-
MT,B 11.8/2.2 M21,2s 5.4/2.2
Mg,to 149.8/2.2 M29,30 78.8/2.2 ble 3. The final Et.. modulator prototype displayed
Mu,12 65/2.2 M3t,32 34.2/2.2 16.8bit@40KHz [23].
Mta,tt 78.8/2.2 M33-48 5.0/1.2
Mts,to 34.2/2.2 Mblaa 378.0/5
MtT,t6 121.8/2.2 Ct-4 0.4 pF
Mt9,20 142.8/2.2 lblaa 74 ItA 6.3. Regenerative Comparator
were obtained for the specifications needed in a The comparator used in the same 17bits Et.. modu-
17bit@40KHz fourth order Et!. modulator. The spec- lator was designed using the high-frequency regen-
ifications are given in the first column of table 3. Once erative latch of figure 11 to meet the specifications
again only the power consumption was a design objec- of table 4. The simulated and measured results of
tive. The optimization process started from scratch on a the sized schematics provided by the design tool are
10-dimension design space and required about 45mins. also shown in table 4. As for the previous opamp,
14
Global Design of Analog Cells using Statistical Optimization Techniques 191
3.5
c 2.5
.S!
....
u
c
~
.... 1.5
"'0
u
0.5
0 # iterations
Fig. 8. Cost function evolution for the optimization of figure 7.
measurements correspond to a prototype built in a the measured resolution time and have found that they
CMOS 1.2Jlm double poly n-well technology. We have can be fully explained by taking into account the dy-
analyzed the origin of the slight deviations observed in namics of the measurement set-up.
15
192 Medeiro, Rodrfguez-Macfas, Fernandez, Domfnguez-Castro, Huertas and Rodrfguez- Vdzquez
Ao ~ 70 78.52 76.01 dB
GBW(lpF) 2: 30 34.88 MHz
GBW (12pF,IM!l) 4.17 4.21 MHz
PM(IpF) 2:60 66.28
PM(I2pF, IM!l) 87.2 86.8
Input white noise ~12 13.53 nV/.JHz
SR ~70 74.81 70.5 V/IJ.S
OS ~ ±3 ±3.2 ±3.0 v
Offset 3.35 mV
Power minimlze 1.95 1.93 mW
6.4. High-Frequency Analog Buffer CPU time is shown in table 5. Table 6 shows the spec-
ifications, where the DC gain (A 0 ), output range (OS),
The analog buffer of figure 1 was designed for very low and power consumption are design objectives; the 3-dB
input capacitance. The sizing obtained after 30mins frequency is a weak restriction, and the input capaci-
tance is included as a constrained design objective. As
Table 4. Simulated and measured results for the comparator.
16
Global Design of Analog Cells using Statistical Optimization Techniques 193
shown in the third column of table 6 the tool was able results shown in table 3 for the same example. This is
to obtain a solution with input capacitance as low as due to the change in technological parameters. A tech-
0.07pF and /3dB of 34.35MHz. nology with available data about electrical parameter
correlations was necessary to apply the reduced vari-
Table 6. Simulated results for the analog buffer (output load ance technique. Hence, it was reasonable to compare
IOpP@IMU). the results with the memory-less technique using the
same technological parameters.
Specifications Simulated Units
min. min.
Specs nominal mean variance value mean variance value Units
17
194 Medeiro, Rodrfguez-Macfas, Fernandez, Domfnguez-Castro, Huertas and Rodrfguez-Vazquez
18
Global Design of Analog Cells using Statistical Optimization Techniques 195
de Microelectr6nica. During 1993 he stayed at the Katholieke from the University of Seville, Spain. From 1970 to 1971 he
Universiteit Leuven, Belgium, as a senior researcher. His was with the Philips International Institute, Eindhoven, The
research interests include design and modeling of analog in· Netherlands, as a postgraduate student. Since 1971, he has
tegrated circuits and analog design automation. been with the Department of Electronics and Electromag-
netism at the University of Seville, where he is a professor.
He is also the Director of the Department of Analog Circuit
Design of the Centro Nacional de Microelectr6nica. His re-
search interests are in the field of multivalued logic, sequen-
tial machines, analog circuit design, and nonlinear network
analysis and synthesis.
19
Analog Integrated Circuits and Signal Processing, 6, 197-208 (1994)
Q 1994 Kluwer Academic Publishers, B06ton.
iewel@olympUI.•u.Jhu.•dM; andrroo@olympiU.•c•Jiut.•d•
Abstract. The loseless property of an MOS floating gate is exploited to implement exact summing operations in
the charge domain. Loseless charge sharing in such structures yields circuits with potential applications as building
blocks for analog signal processing. Large signal as well as small signal models of floating-gate transistors are
presented for both above-threshold and subthreshold regions. Experimental data from fabricated devices in a 2
micron double poly, n-well process are in good agreement with the models. A canonical structure, the Multiple
Input Floating-gate Differential Amplifier is proposed and its use in different circuit configurations demonstrated.
Experimental data from a multiple differential input operational amplifier are presented. Limitations of the proposed
circuits are also discussed.
1. Introduction
-- f Ii
makes analog computing methods advantageous over
---
discrete time and digital algorithmic techniques forcer-
tain class of problems. It In
A basic operation in all electrical circuits is that
of current addition. Kirchoff's Current Law (nothing
lout
but conservation of charge) makes this operation exact.
Weighted summation, of some input signals X 1 (a)
vt
ct _i_
• This work was partially supported by a NSF Research Initiation Fig. I. Weighted summation using (a) currents and transconduc-
Award (MIP-9010364). tances and (h) charges and capacitances.
21
198 Yang and Andreou
factor. The output of the circuit is given by standard double-poly CMOS process. By replacing
the two input transistors with multiple control gate
lout = L:: Gmt Vl Floating-Gate MOSFETs, we realize multiple differ-
ential input pairs. Since there are no extra transistors
In the charge based circuit, first introduced as a introduced, problems caused by transistor mismatch
model of a neuron by Shibata and Ohmi [6] (see fig- are no worse than a single input pair MOS differential
ure l(b)), exact summing is implemented by direct amplifier.
conservation of charge on the floating node. A fixed In Section 2 we discuss the physics of the multiple
amount of charge is shared by input voltages through control gate Floating-Gate MOSFET, which is the core
coupling capacitors. C'P and G1 are parasitic capaci- of a MIFDA. Designs of MIFDAs will be presented
tance and leakage conductance, respectively. For an in Section 3 along with measurements on a fabricated
ideal system these components have to be minimized. MIFDA operational amplifier and analysis of its lima-
The transfer function of such a system is given by lions. Finally in Section 4, we show some potential
applications of MIFDAs as analog signal processing
building blocks.
22
A Multiple Input Differential Amplifier Based on Charge Sharing on a Floating-Gate MOSFET 199
(a)
D
Polyl
models for both regimes.
~
Me tall
2.1. Subthreshold
D
Active
face is in weak inversion or even depletion, the surface
potential is a quasi-linear function of the control gate
voltages. We view a MOSFET as symmetrical device
(b) with interchangeable source and drain. The current
equation generally takes the form [2]:
23
200 Yang and Andreou
(b) where
F(VFGB, V) (VFGB - VrH - V) 2
Fig. 4. Output characterisllcs of single control gate Floating-Gate
MOSFETs. (a) N-channel; (b) P-channel. (9)
+ ~~(V + 2</Jp)!
For a regular MOSFET
(5)
24
A Multiple Input Differential Amplifier Based on Charge Sharing on a Floating-Gate MOSFET 201
........ ·········~···· .. ·· ..
1 o_,,
1 o·13
-2 -1 0 2 3 4 5
Fig. 5. Id• vs. V1 and V2 of an N-Channel two control gate Floating-Gate MOSI'ET, W = L = 4p.m, and control gate areas are 3 X 4 p.m2 •
the weighted sum of control gate voltages is given by in figure 5 gives the inside how the transistor works.
Both control gates have equal influence on the tran-
2Cfc lj(TJ (II) sistor's transfer function. If one view Vi as the gate
m c....
9m=--y.tHD
voltage of the transistor, v2
then governs the threshold
The output conductance is also different from that of a voltage of the transistor.
regular MOSFET, To control the floating gate charge, we can use UV
light which temporarily increases the conductance of
Yd = 9d + c2Cfd y'Ki; (12) the oxide associated with the floating gate. The change
aum in conductance allows the charge and discharge of the
floating gate while maintaining the minimal leakage
where 9d = >.ID· >.is the Early effect parameter asso-
current under normal operations. There are other tech-
ciated with the channel length modulation.
niques reported to use a switch transistor connected
In general, a multiple control gate Floating-Gate
to the floating gate to change the floating gate charge.
MOSFET operating in either the subthreshold or
above-threshold region has a transfer characteristic as However, we believe that in an analog system espe-
following: cially where Floating-Gate MOSFETs are used in dif-
ferential mode, even a small leakage current from the
i = 9m(W!V! + W2V2 + · · · + WnVn) (13) floating gate will greatly affect the circuit performance.
The UV light method shows clear advantage over the
where 9m is given by (4) for the subthreshold and (11) switch transistor method.
the above-threshold. One thing we shall mention here
is that the models described above are only good for
3. The Circuit Implementation: MIFDA Design
the first order analysis of the circuit operation, a better
complete Floating-Gate MOSFET model suitable for MIFDA circuit design follows the same principle of
circuit simulator has been developed here and will be regular Opamp or arA design except that the two in-
addressed elsewhere. The charge sharing on the float- put MOSFETs are replaced by multiple control gate
ing gate causes every control gate to have a weighted Floating-Gate MOSFETs. Figure 6 shows a simple
influence on the output current. Measurement results two-stage multiple input Opamp. This circuit is de-
of a two control gate Floating-Gate MOSFET shown signed to work in the above-threshold region. Input
25
202 Yang and Andreou
Vss -5V
=lb-
--< A
--<
--< -
--<
26
A Multiple Input Differential Amplifier Based on Charge Sharing on a Floating-Gate MOSFET 203
27
204 Yang and Andreou
I+ I·
I+ I.
In the subthreshold region, the transfer function is
given by [1]:
t t
v+-i~~v- !:!.I
,.;,/:iV
= Ibia.tanh-2- (18)
.
vbtas _j
I
t !bias
The input range is defined such that the tanh function
-::- (a} is in the quasi-linear region. For ,.;, = 0.6, the input
Fig. 9. A comparison of input stages. (a} DDA; (b) MIFDA. range is approximately l!:i VI ~ lOOmV.
(a) (b)
(c) (d)
(e)
28
A Multiple Input Differential Amplifier Based on Charge Sharing on a Floating-Gate MOSFET 205
For a MIFDA input stage shown in figure 9(b ), the Figure 13 shows a self-compensating sample-and-
above limitation applies only to the differential float- hold circuit using a MIFDA. The leakage current
ing gate voltage, l1 V19 , instead of each differential pair. through the switch transistor causes a decrease in
This means that there is only a range limit for the sum the holding time of a regular sample-and-hold circuit.
of input differential voltages. However, the follower made of a MIFDA responds only
to the differential input. A similar leakage at the neg-
(19) ative node v- virtually compensate fot the leakage at
the positive node v+. In the sample period, the gates
of the two switch transistors aresetto V8 s. v-is con-
Here we assume uniform wei~hts for all inputs. From
nected to ground while v+ is connected to the sample
the device model, we know 0: > n. In the worst
signal. In the hold period, the gates of the switch tran-
case when all differential inputs are the same in both
sistors are changed to Vvv . The leakages of the v-
amplitude and phase, the input range is still slightly bet-
node and the v+ node cancel each other to maintain
ter than that of a current-mode DDA. A small amplitude
a longer holding time. We shall point out that for this
of one differential signal will help increase the ranges
circuit to work, the two leakage currents have to be
of other input signals. Furthermore, phase difference
very similar.
among signals will also help improving the operating
range of input signals.
29
206 Yang and Andreou
w0 =1/8
Wt=l/4
W2=1/2
wa=l
Hold
where r = C/Gm. Figure 14(b) shows a differential lower circuit complexity and better input transistor
differentiator. The transfer function is matching. The input range of a MIFDA is improved
V, = Asr (V+ _ v-) over its DDA counterpart. This provides more flex-
(21)
ST +1 ibility in the use of MIFDAs. The input offset prob-
lem associated with the floating-gate MOSFETs is well
studied. UV light is used for after-fabrication treat-
5. Conclusions
ment of offset caused by the random distribution of
Multiple control gate floating-gate MOSFETs are floating-gate charge. A symmetrical input stage de-
proven to be effective in realizing multiple input dif- sign is also effective in eliminating offsets caused by
ferential amplifiers, which function like regular CMOS the drain feedback effect. It should be pointed out here
DDAs and have some unique advantages including that the circuits we designed for testing are simple and
30
A Multiple Input Differential Amplifier Based on Charge Sharing on a Floating-Gate MOSFET 200
c
T
(a)
(b)
Fig. /4. Dynamic analog signal processing circuits. (a) Differential Integrator; (b) differential Differentiator.
not fully optimized; however, the same design princi- 2. Andreou, A. 0. and Boahen, K. A., "Neuromorphlc Informa-
ples could yield practical amplifier designs. The use of tion Processing: The Current Mode Approach," In Analog VI..SI
MIFDAs in analog computational systems are not lim- Signal and Information Processing, edited by M. Ismail and
T. Plez. McGraw-Hill, 1994.
ited to the examples we show in the last section. We
3. Allstot, D. J. and Black, Jr., W. C., "Teclinologlcal design
believe that MIFDAs have the potential to be the center coRBideration for monolithic MOS switched-capacitor filter-
of many future analog signal processing systems. Ing systems,'' Proc. IEEE, Vol. 71, pp. 967-986, August 1983.
4. Siicldnger, B., Ouggenbiihl, W., "A versatile building block: the
6. Acknowledgments CMOS differential difference amplifier," IEEE J. Solid-State
Cirr:ults, Vol. SC-22, pp. 287-294, Aprill987.
We thank Paul Furth for his suggestions on improved 5. Huang, S.-C., Ismail, M. and Zarabadl, S. R., "A wide range
circuit design. We also thank Mark Martin and peo- differential difference amplifier: a basic block for analog sig-
nal processing In MOS technology,'' IEEE Thzn. Cirr:ult.s and
ple in Analog VLSI group at Johns Hopkins for proof Systems, Vol. 40, pp. 289-301, May 1993.
reading the manuscript 6. Shibata, T., Ohml, T., "A functional MOS transistor featuring
gate-level weighted sum and threshold operations," IEEE Tran.
Electron Devices, Vol. 39, pp. 1444-1455, June 1992.
References
7. Vittoz, B. A. and Pellrath, J ., "CMOS analog Integrated circuits
I. Mead, C. A., Analog VI..SI and N~ural Sy•tems, Addison- based on wealc Inversion operation,'' IEEE J. of So/id...State
Wesley, 1989. Clrr:uits, Vol. SC-12, pp. 224-231, June 1971.
31
208 Yang and Andreou
8. Vittoz, E. A., "The design of high-performance analog cir- and the M.S. E. degree in electrical and computer engineer-
cuits on digital CMOS chips,"IEEE J. of Solid-State Circuits, ing in 1991 from The Johns Hopkins University, where he
Vol. SC-20, pp. 657-665, June 1985. is currently working toward the Ph.D. degree. From 1986 to
9. Fichtner, W. and Piltz!, H. W., "MOS modeling by analytical
1989, as a research engineer in Microelectronics R&D Cen-
appro~imations," Int. J. Electronics, Vol. 46, 1979.
10. Yang, K. and Andreou, A. G., "Subthreshold Analysis of
ter at Chinese Academy of Sciences, he worked on gate array
Floating-Gale MOSFETs,'' in Proc. Tenth Biennial UGlM ASIC design tools including logic, transistor and physical
Symposium, pp. 141-144, Research Triangle Park, NC. May layout cell libraries. He is now a research assistant in the Mi-
1993. croelectronics Laboratory at The Johns Hopkins University,
II. Yang, K. and Andreou. A. G., "Multiple Input floating-gate working on device modeling, floating-gate analog memory
MOS differential amplifiers and applications for analog com· and subthreshold MOS analog VLSI.
putalion,'' Proc. 36th Midwest Symposium on Circuits and Sys·
tems, IEEE, August 1993.
12. Meyer, J. E., "MOS models and circuit simulation,'' RCA Re·
view, Vol. 32, March 1971.
13. Glasser, L. A.," A UV write-enabled PROM,'' 1985 Chapel
Hill Conference on VLSI, pp. 61-65. Computer Science Press,
1985.
32
Analog Integrated Circuits and Signal Processing, 6, 209-217 {1994)
Q 1994 Kluwer Academic Publishers, Boston.
Abstract. This paper presents design techniques for a wide input range CMOS differential difference amplifier
(DDA) and discusses its application as a basic block in the implementation of a simple four-quadrant multiplier celL
The cell can be configured as an amplitude modulator or a one-over circuit, which are widely used in many analog
signal processing applications. The DDA can also be reconfigured as an opamp, and hence can be used to design
many of the opamp-based multiplier circuits. The DDA amplitude modulator (AM) uses a transistor and a resistor
as the only components external to the DDA. A DDA one-over circuit, which provides an output proportional to the
inverse of the input, is also achieved with the same level of simplicity. High-frequency effects due to the DDA's
finite gain-bandwidth (GB) and MOS parasitic capacitances are investigated. Experimental results obtained from a
2 ~tm CMOS MOSIS chip are given.
1. Introduction
The Differential Difference Amplifier (DDA) is a novel
Vpn ---1
analog building block, whose symbol is shown in fig-
ure l(a). The concept of operation is similar to that of Vpp ---1
an opamp with the output given as Vnn ---1
Vnp ---1
Vo = Ao[(Vpp- Vpn)- (Vnp- Vnn)] Ao-+ 00
(I) (a)
tY :
where A 0 is the DC gain of the DDA and (V,p- Vpn)
and (Vnp- Vnn) are the two differential voltage inputs.
~~~
Therefore, if a negative feedback is introduced through Vo
nodes Vpn and/or Vnp• the following design equation
is obtained. Vnn
-Vo
Vnp
Vw- Vpn = Vnp- Vnn (2)
(b)
A fully-differential DDA (FDDA), shown in fig-
ure 1(b), can be designed in a manner similar to a Fig.l. Symbolsof(a)DDAand(b)PDDA.
fully-differential opamp [1]. Furtherm(>re, A DDA can
be reconfigured as an opamp as shown in figure 2 [2], As a result, the DDA can be used to develop opamp-
so that based building blocks. Figure 3 shows examples of
opamp-based multipliers [3, 4]. The output of circuit
(3) (a) is given by
• Parts of this work have been presented at the I Oth Norchip Sem- (.lf) 1 (Xl- X2)y
(4)
inar, Helsinki, Finland, November 3-4, 1992 and Included In the
Seminar's proceedings, pp. 9-14. This work was supported In part (.lf) 2 (Zl- Z2)
by the NSF grant MIP-88%244 and by the Semiconductor Research
Corporation contracts 90-DJ-066 and 91-DJ-066, and In part by the where C!f .lf
)I and ( )2 represent the sizes of tran-
Norwegian NTNF, SINTEP and Nordic VLSI. sistors in the input and feedback sides of the opamp
33
210 Huang and Ismail
Z2
_L
(a) y Vo
-Y
Yo
-Yo
(b)
T
Z2
(a)
Fig. 2. (a) Single-ended and (b) fully-differential opamps.
34
Design and Applications of a CMOS Analog Multiplier Cell Using the Differential Difference Amplifier 211
VM+vm
vM~sv
4
VM=4V
3 VM=3V
Vc VM•2V ·-···
Vo
0
-I
-2
(a) -3
-4
VM+vm I R -sL_~~-~~~~~~-~~~
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
y,~
Vc(V)
:Vo
Fig. 6. Output transfer curves of tbe multiplier with various VM.
~
where A is a DC signal and fm(t) and cos wet are the
-Vo modulating and carrier signals respectively. m is the
modulation index, where m ~ 1 is required for demod-
(b) ulation using an envelope detector [8]. According to
(7) and (8), one can see that Vm operates as the modu-
Fig. 5. DDA-based multiplicationlrnoduiation cell (a) single-ended lating signal and V., as the carrier signal. The DC signal
and (b) fully-differential.
A is
1/2K (VG- Vr). The inverting bias of drain and source A=- [1 + 2KR(VM- Vr)] (9)
nodes can be achieved by a DDA with Vnp = Vin•
and the modulating index m is
Vpp = Vnn = 0 and a negative feedback through node
Vpn• which can be verified easily using design equation 2KR
(2). The output voltage of a multiplier is obtained by m = 1 + 2KR(VM- Vr) (lO)
converting the resulting current into a voltage through
a resistor as shown in figure 5. The output is obtained To satisfy the condition m ~ 1, we must have
as
R ~
2K(VM - Vr- 1)
Vo -Vc- 2KR(VM + Vm- Vr)Vc
for VM- Vr >1 (II)
-[1 + 2KR(VM- Vr)]
(7)
1
[ 2KR ] R ~
1 + 1 + 2K R{VM - Vr) Vm Vc 2K(VM- Vr- 1)
for VM- Vr <1 (12)
where VM + Vm ~IV.: I+ Vr for an NMOS transistor. On the other hand, the amplitude of Vo is limited by
Note that the above expression results in a multiplica- the output swing of the DDA used. If the maximum
tion of Vm and V.,. SPICE DC simulation results of output amplitude is assumed to be Vvv(= -Vss),
Vo versus V., with various VM (vm = 0) are shown in V., = Vcp cos wet and Vm has amplitude VMP. we
figure 6. To be more specific, the multiplier functions have
as an amplitude modulator. Recall that an amplitude
modulated (AM) waveform with a sinusoidal carrier is
[1 + 2KR(VM- Vi)]
given by
2KRVMP ]
x [ 1 + 1 + 2KR(VM- Vr) Vcp ~ Vvv
¢AM(t) = A(l + mfm(t))coswct (8) (13)
35
212 Huang and Ismail
Therefore,
(14)
. [-
mm 1
-::-::-::-:=-:----::c:--,..,.
2K(VM- VT -1)'
Fig. 8. Threshold-voltage-independent DDA-based multiplier.
VDD/Vcp -1 ]
2K(VM- VT + VMP) Since R is inversely proportional to K, a small value of
R can be used by choosing a large K value (large W)
for VM- VT <1 for the transistor. Figures 7(a) and (b) show the SPICE
(15) simulation outputs of the modulator with R =
500 n
and 800 n respectively, where W / L = 50 JLmi3JLm
andVT = 0.975VfortheNMOStransistorandVM =
4 4 V. vm is given as a sinusoidal wave (amplitude IV)
with frequency 1 KHz and V, is a square wave (ampli-
tude 1V) with frequency 20 KHz.
A straightforward way to implement a threshold-
0
voltage-independent multiplier is to use two modula-
tion cells as shown in figure 8. The output is given
-2
as
-4
(16)
0 0.0005 0.001 0.0015 0.002
t(sec)
where transistors M 1 and M2 are assumed matched.
One of the inputs, Vm = Vml - Vm2• and the output
(a) are differential. If a single-ended output is desired, a
differential-to-single-ended converter is easily realized
by a DDA difference amplifier [5]. However, a sim-
pler implementation can be obtained using the opamp
architecture shown in figure 3(a).
Using the same concept, a one-over circuit, which
provides an output proportional to the inverse of the
0
input, can be achieved by the circuit shown in figure 9,
where Va and VB are bias voltages and v;n is the input
-2
of the circuit. This circuit finds many applications, e.g.,
-4
a tracking system where the target velocity is inversely
proportional to an integrated pulse width [9). Note
0.0003 0.001
t(sec)
0.0015 that the FDDA is connected as a differential unity gain
buffer [ 10]. As a result, the outputs are directly con-
nected to the inputs of the FDDA (high input impedance
(b) nodes), and therefore the buffer stage in the FDDA can
Fig. 7. Simulation results of the DDA modulator with (a) R = SOOn
be eliminated. From (2) we have VI-Vo = v2- (- Vo)
and (b) R = soon. which gives VI - v2 = 2Vo. Since VI = VB -1 Rand
36
Design and Applications of a CMOS Analog Multiplier Cell Using the Differential Difference Amplifier 213
37
214 Huang and Ismail
Vdd
Vss
Finite GB Effect
Vo = Vpn-IR
If the finite GB effect of the DDA is taken into ac-
count in the AM circuit, nonlinearities are introduced
in the transistor current I, since the DDA inverter is now
frequency-dependent. Assume the open-loop gain of
the DDA is modeled by A(8) = GB/8. According to
(2), we have + (VG- VT +\'c) 2;B Vo] (22)
which gives
(19)
V0 = -[1 + 2KR(Va- VT)]\'c x €GB(8) (23)
i.e.,
where fG 8 ( s) is an error function resulting from the
GB effect and is given by
(20)
1
(24)
Note that Vpn is no longer equal to - Vc. The current fGB( 8 ) = 1 + crn[1 + KR(VG- VT + Vc)]
I is therefore given by
It results in a phase lag, (),written as
w
() = tan- 1 GB[1 + KR(VG- VT + Vc)] (25)
38
Design and Applications of a CMOS Analog Multiplier Cell Using the Differential Difference Amplifier 215
Fig. 14. AM circuit with aMOS small-signal high-frequency model. 1 + (Ra2C, + f)2w2
Cd and C, are extrinsic capacitances and Cp represents tan-
l WT -tan- l w -
'6 (ReG, + 'T)4
2-
the distributed capacitances in the channel, which are
modeled as a uniform RC transmission line. The first (33)
order y parameters of the transmission line are given
The distributed effect reduces the equivalent resistance
by [12]
of the transistor slightly, and introduces small phase
lag. The output voltage is written as
IT+ 1
Y11 ~ ---2"'=-...,.-
Y22 ~ Rc("f + 1) -V.,- IR
(27)
-1 -Vc[1 + 2KR(Va- VT)/ET(8)] (34)
Y12 = Y21 ~ Rc(1f + 1)
5. Experimental Results
where T = RaCp and Cp = C 0 ., W L. Using the
model, the AM circuit is redrawn in figure 14. Note The DDA proposed in Section 3 was fabricated in a
that the extrinsic capacitance cd is eliminated in the 2 1-lm P-well MOSIS CMOS process. A common cen-
figure, since it is driven by a voltage source Vc. Using troid layout technique was used to improve the match-
the admittance parameters, we may write the following ing of the devices incorporated in the input stages, Fig-
equation. ure 15 shows the measurement of V0 with Vm (100Hz
and amplitude I V), v;, (not shown with I KHz and
amplitude I V) and R = 220 !1. The transistor in the
circuit is built by a series combination of2 SK9160 de-
pletion transistors (which have their sources connected
The equivalent admittance, Y, is therefore given by to the bulks) with their gates and sources connected
respectively and their drains being the new drain and
y Ir /2 source of the composite transistor. This implementa-
Vc = -'V;, tion reduces the body effect of the transistors. The
depletion mode transistors were chosen to simplify the
Y11- Y12 = Y22- Y21 experiments since they are normally on (no DC com-
ponent, VM, is required). It can be seen from figure 15
(29) that the envelope of the output is proportional to the
Rc(~ + 1) modulating signal Vm.
39
216 Huang and Ismail
h,P running
· · · ..........,. . . . . . . . . . . ,.................... ,................. , 1 500 IV I d i v
. J. . . ..
!
..... . ......
V., Joffset:-35.00 mv
.. •····~ . .. .. .. . .... .... . . ~ 1 ' 000 1 cc
:/ ··~ [\
\
J\ v; :, \
[\ r; \
i ·\ .... ... ···~. . ··+ . ·~
..
4 500 W:V/dlV
·~ offset•-35.00 mv
.r··
1 . 0()0 1 cc
....... i
-10.0000 ms 0. 000(10 s 10.0000 ms
2.00 ms/dill
Fig. 15. Experimental results of the DDA modulation cell.
6. Conclusions 4. Khachab, N.J. and Ismail, M., "A nonlinear CMOS analog cell
for VLSI signal and information processing," IEEE J. Solid-
The design of a wide input range DDA and a simple State Circuits, Vol. 26, pp. 1689-1699, Nov. 1991.
5. Huang, S.-C., Ismail, M. and Zarabadi, S. R., "A wide range dif-
DDA-bascd multiplier cell have been discussed, and ferential difference amplifier: A basic block for analog signal
experimental results are given. A DDA can be recon- procerssingln MOS technology," IEEE Trans. Circuits Sy!l.-
figured as an opamp to make use of opamp-based build- II, Vol. 40, pp. 289-301, May 1993.
ing cells. In addition, simple AM and one-over circuits 6. Ismail, M., Huang, S.-C. and Sakurai, S., "Continuous-lime
signal processing," in Analog VLSI: Signal and lnfornuuion
are developed by reconfiguring the basic multiplier cell Procusing, edited by M. Ismail and T. Piez, ch. 3, McGraw-
and taking advantage of the differential input prop- Hill, 1994.
erty of the DDA. This demonstrates that the DDA is 7. Mead, C. and Ismail, M., Analog Vl.SIImplementation of Neu-
a competitive building block to opamps in many signal ral Systems. Kluwer, 1989.
8. Stremler, F. G., Introduction to Communication Systems.
processing applications. AII-MOS implementations of Addison-Wesley, 1982.
the circuits discussed in this paper are easily achieved 9. Etienne-Cumrnings, R., Hathaway, R. and Van der Spiegel,
by replacing resistors with aii-MOS voltage-controlled J., "Accurate and simple CMOS 'one-over' circuit," Electron.
Lett., Vol. 29, pp. 1618-1620, Sept. 1993.
floating resistors, e.g. see [13]. The voltage-controlled
10. de Ia Plaza, A. and Morlan, P., "Power-supply rejection in dif-
resistors will provide electronic programmability of the ferential switched-capacitor filters," IEEE J. Solid-State Cir-
output voltage. cuits, Vol. SC-19, pp. 912-918, Dec. 1984.
II. Sackinger, E. and Guggenbuhl, W., "A versaUle building block:
the CMOS differential difference amplifier," IEEE J. Solid-
State Circuits, Vol. SC-22, pp. 287-294, Aprill987.
References 12. Khachab, N. I. and Ismail, M., "Linearization techniques for
nth-order sensor models in MOS VLS! technology," IEEE
I. Huang, S.-C. and Ismail, M., "A CMOS differential differ- 'fran Circuits Syst., Vol. 38, pp. 1439-1450, Dec. 1991.
ence amplifier with rail-to-rail fully-differential outputs" (to 13. Sakurai, S. and Ismail, M., "A CMOS square-law pro-
be published). grammable floating resistor independent of the threshold volt-
2. Zarabadi, S. R., Larsen, F. and Ismail, M., "A configurable age,'' IEEE Trans. Circuits Syst., Vol. 39, pp. 565-574, Aug.
op-ampfDDA CMOS amplifier architecture," IEEE Tran•. Cir- 1992.
cuits Sy•t., Vol. 39, pp. 484-487, June 1992.
3. Khachab, N. I. and Ismail, M., "MOS multiplier/divider cell
for analogue VLSJ," Electron. Lett., Vol. 25, pp. 1550-1551,
Nov. 1989.
40
Design and Applications of a CMOS Analog Multiplier Cell Using the Differential Difference Amplifier 217
41
Analog Integrated Circuits and Signal Processing, 6, 219-229 (1994)
0 1994 Kluwer Academic Publishers, Boston.
Abstract. This contribution describes developments in the use of numerical optimization techniques as part of a
package whose function is to generate silicon-level layout for general analog functional modules from high-level
specifications. The investigation is using switched-capacitor filters as a case study that is representative, in terms
of its associated physical layout problems, of many classes of analog circuits.
H ~
an efficient technique for the use of an optimization
EDGE
FEEDS PLL
LAYOUT
algorithm in the context of a circuit-simulation-based
H,
ENVIRONMENT
design scheme necessarily involving very CPU-time-
LIN/LOG . ~
MULTIPLIER intensive design objective function evaluations. It is
suggested that these latter results have significance and
1 OTHERS ~ application considerably beyond the immediate area of
this study since numerical optimization based design
------
strategies are commonly used in engineering design
Fig. I. Overview of knowledge-based analog compiler system. where system simulation is expensive but constitutes
43
220 Gustard and Massara
the only practicable way of characterising system per- Biquad structures do however have disadvantages.
formance. Often a single component defines a filter pole, result-
ing in high sensitivity to component tolerance. Given
any reasonably demanding or high-order filter specifi-
2. Choice of SCF Structures
cation, one or more of the biquad sections will be in-
The realization of analog filters in silicon integrated volved in realizing high-Q pole-pairs. Very high sensi-
circuits encompasses a number of design problems tivities to component error are likely to result and, since
which limit the choice of circuit structure. Compo- component tolerances are inevitable in silicon realiza-
nents are nonideal, op-amps have limited gain, finite tion, the frequency response produced by the filter may
output impedance and parasitic poles. Switches have deviate unacceptably from that required by the speci-
finite on-resistance, capacitors have many stray com- fication.
ponents associated with them [4] and significant toler- Figures available for capacitance tolerance vary
ances [2, 5]. A further consideration is that of structure widely, Gregorian and Ternes [2] quote figures of
size; silicon area is expensive, hence any reduction in ± 15% for absolute accuracy and a range of 0.1 - 1%
physical circuit size is advantageous. All these factors for ratio accuracy. Geiger, Allen and Strader [5] quote
have stimulated the development of SCF designs. ± 20% for absolute accuracy and ± 0.06% for ratio ac-
curacy.
To illustrate the sensitivity of a biquad filter to com-
3. Blquad SCFs ponent tolerances, a series of simulations were carried
out where the netlisting routine introduced a variation
Since the 1970s many novel structures have been re-
in component values. Given the above figures, a ran-
ported for the realization of switched capacitor filters.
Each structure has particular advantages. Biquadratic dom error of ± 1%, the extreme of the quoted ratio
blocks (blquads) have low sensitivity to the stray capac- ranges, was introduced in steps of 0.1% across all the
itance inherent in silicon layout, are easy to cascade and passive circuit components. These tests where used
simple to design. However, they are restricted by an rather than the "classical" sensitivity analysis as sen-
inability to produce high Q values, limiting their use sitivity plots are rather open to interpretation. With a
in high-order filters. There have been many biquad series of tolerance plots worst case situations can be
circuits reported, each with various advantages (rep- more easily seen.
resentative examples include [6-11] and [2]). Biquad The specification chosen for this example was that
circuits of the type reported by Gregorian and Ternes of a 7th-order Chebyshev filter with 1dB passband rip-
[2] were chosen for use in this system because of their ple. Passband cut-off frequency was 5kHz and the
low component count and versatility of application to capacitor clock rate was 60kHz. This specification
common filter functions. required three biquad sections of the form of figure 2,
Vin
Vout
Cl"
44
Optimal Design of SC Filters for Analog/Mixed-Signal ICs 221
Table2.
CD
Unit Capacitor Values for Pirst..Qrder Section
..-----II~~
C!>/2 CE ~'
cs CS/2 CD CB
-0t· t·
2.000000 1.000000 1.968542 17.955232
1.5 I I
1 :·.-
0.5
,•. . ..
......... ......
0
Gain '/..·'{:·.= .= _...;~·~.. .~.~~.:~_.:_:.:_:_·!_,, :::·;~,:
dB '~...'~'·.·,,
~~~:~~~ ·. :~-!) .:.f.t: ~
•• .'•"•":·. • • r ::• ,,._
-0.5
..• t- ... .
. ..~'.·.·:::: ·. ·"'.';
0
•.,·_j(-" ··.~.~>.....
-1 ,~~,;i!·I · . .~ ~.::·:':::·r·
., ' v· -.·~·.·-: · · ·.·.:.·: ·.·.•::_..·;,
.....·•._...
. . :: i=
·.·· ..
0 ...
'..
-1.5 ....
·:.··
-2~--~--~--~--~--~--~--~--~--~--~
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Frequency (Hz)
Fig. 4. 7th-order biquad filter with a ± 191> random component tolerance.
45
222 Gustard and Massara
46
Optimal Design of SC Filters for Analog/Mixed-Signal ICs 223
I .\ ,. .
0.2 ,------,.---.,..--.--.---,----,----.---,--.,.----.,
0 h-----------------~~------------~
-~. -t·..... .I
~·r-----~.. .5?
I
~~~
,,
..
"1/ ·::: .. r
-0.2 -~ ?: ·t ..~-
·lj
-0.4 ·v':'i; ?·'.' ·•;.
·;~.
:g
.,
-0.6 .• . ~.... ~
~.·
~~- ·:::~
·::,. l• ·:r· : 'i
Gain _ 0 8
• •~
, ·::;. )),' ·f; ·: •'\.:?
dB l 1
·•I
... ,.
)•
)•. ·,:t:. ·:Jt
''!,!'t. .·:"'!(·...
-1
•, /,1,1,'•,
":·:·· ?;;}
. ... -.
-1.2 ....
-1.4
-1.6
-1.8 ._____._ __.__ __.__ _.__ _,__ _..___..___..__--J'------1
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Frequency (Hz)
Fig. 6. 7th-order ladder filter with a ± 19& random component tolerance.
-0.2
Gain _ 0 4
dB .
-0.6
-0.8
-1 ~---'--~~~--~--~~-----'------~
0 1000 2000 3000 4000 5000 6000
Frequency (Hz)
Fig. 7. Distortion of passband response of a 5th-order chebyshev ladder filter.
The macromodels used for this are shown in figure 8. Top-plate stray-capacitance was approximated by the
Bottom-plate stray-capacitance was calculated as the interconnection value.
typical process value for polyl to substrate, plus an The netlister calculates the value of capacitor for the
approximate interconnection capacitance, caleulated given clock frequency to model finite switch resistance
from process value for metall to substrate of mini- with a value of lOkfl, the simulated impedance of a sin-
mum width and length equal to the capacitor perimeter. gle minimum size MOSFET for the design process.
47
224 Gustard and Massara
c
Cbot yr~ Ctop ~:~r
~~Cr
a) Capacitor b) Switch
Op-arnp macromodels introduce a dominant-pole- The layout of analog circuits requires careful analy-
frequency, input capcitance and output impedance. sis, and many restrictions must be considered. Output
Gain of the op-amp is a capability of SWITCAP II nets must not cross input nets lest positive feedback
and was given a value suitable for the op-amp model. cause circuit instability. Noisy nets such as clock sig-
All the above model values can be adjusted for a given nals must not cross sensitive nets such as op-amp in-
process. puts. Common centroid layout must also be considered
Where biquads are used to achieve a commercial to ensure accurate capacitor ratios [11, 3].
design, a common response to the difficulties of high Circuit schematics are produced with Cadence
sensitivity and the problems added by non idealities and EDGE [17] which has been modified to include SC
strays is to over-design. Thus extra circuit sections will components. It is within this environment that non ideal
be added to ensure that a response template is met de- components are calculated for the netlist macromodels.
spite the high cost in silicon area. For relatively un- After layout of the circuit feedback of the actual layout
demanding specifications this may be acceptable but capacitance can be used for a post-layout optimization
precision requirements could well involve considerable run.
additional circuit complexity.
The SCF Expert produced in this study will provide 6. Optimization Development
either biquad or ladder realizations, selected according
Many methods of numerical optimization have been
to the demands of the specification. For both classes
reported, the most effective examples of which require
of structure, we have explored the effectiveness of us-
function derivatives. Obtaining electronic circuit func-
ing numerical optimization to compensate for realistic
tion derivatives is generally a nontrivial problem exert-
models of the imperfections that will necessarily re-
ing restrictions upon the choice of optimization routine.
sult from IC reali7.ation. These imperfections include
Approximate derivatives of circuit functions can be
stray capacitance; switch resistance; and op-amp non-
evaluated by perturbation of the component variables,
idealities. The result is a system that produces accurate
but this is simulation-intensive requiring n + 1 eval-
designs with no additional silicon area costs.
uations for an n variable problem. Schematic sim-
ulators exert excess loading on the optimization pro-
5. System Operation
cess; SWITCAP II [ 18, 19] typically takes 20 min-
As input to the SCF Expert FEEDS provides a set of utes (elapsed time on SUN SPARC) for a simulation
pole frequencies and Q values for each biquad section involving nonideal circuit components (depending on
and, for ladder functions, a set of continuous analog circuit complexity and high speed clock frequency).
component values. We can see from this that a method which requires
48
Optimal Design of SC Filters for Analog/Mixed-Signal ICs 225
a minimum of simulations per iteration would be an where J is the vector of error values. In LSO, the Hes-
attractive choice for the problem. Newton-Gauss least sian is approximated as
squares optimization (LSO) [20] offers high-speed con-
vergence at the expense of requiring first derivatives. (2)
LSO with real derivatives should theoretically find where J is the Jacobian matrix of first partial deriva-
the minimum of a quadratic problem in one iteration, tives of the elements of f. The algorithm presented
however, on real engineering design problems, the here is developed from a technique reported by Mas-
method may require many iterations to achieve a solu- sara and Fidler [22], and uses the so-called Levenberg
tion and, without suitable damping, may diverge [21]. damping parameter,.>., [23] in the modified expression
LSO approximates the error function by a quadratic for H, where
function. In common with other Newton-based meth-
ods [21] LSO makes use of the Hessian matrix of sec- (3)
ond partial derivatives of the design objective function,
ill which is given by The new algorithm can be seen in figure 9. Note that
re is the vector of design variables and a is the line
search parameter [22].
(1)
If the design objective function is in fact well ap-
proximated by a quadratic then the evaluated approx-
imation to the Hessian should be approximately the
same regardless of the position in the error space. This
led to investigations into the similarity of the matrix at
each iteration. Investigations were extended to a 5th-
order low-pass Chebyshev ladder SCF circuit of the
same structure as figure 5, (this being of reasonable
complexity but sufficiently small to minimize investi-
gation time.)
In this problem, as in most engineering design prob-
lems, it must be supposed that the design objective
function is very far from a simple quadratic form. We
were therefore not surprised to find that, in general,
the Hessian matrix elements varied widely across the
functional surface. On the other hand, it was found
that there were regions in which sequences of large re-
ductions in the value of the sum-of-squares. objective
occurred with relatively little change in the value of
the Hessian. This indication that the design objective
function is, in such regions, well approximated as a
quadratic form can be used to signal the temporary fix-
ing of the Hessian matrix and its use as the basis for a
number of iterations without re-evaluation.
A method was designed which retains the Jacobian
matrix, and hence the approximate Hessian matrix.
The basis is the goodness of the new move; if the matrix
yields a new minimum with better than a given percent-
age improvement, the matrix is retained. Divergence
or an improvement below the percentage criterion re-
sults in the matrix being re-evaluated. In practice, this
has proved to reduce optimization time considerably by
decreasing the number of circuit simulations required
Fig. 9. Flow chart of new algorithm. very considerably.
49
226 Gustard and Massara
Table 4. Comparison of performance of new algorithm compared to a ±1% tolerance ratio variation as in previous experi-
standard algorithm.
ments, however the stray capacitance was subjected to
Percentage improvement of new algorithm a ±50% tolerance. As stray capacitance is not usually
Filter Order 5th 7th 9th
considered a major feature in integrated-circuit design,
figures for the accuracy of stray capacitance are not
Ideal Circuit 15.6% 21.6% 21.3% readily available. However, typical figures are quoted
Nonideal Circuit 28.3% 20.7% 6.9%
Pre-distorted
for the absolute value. Given the tolerance of designed
Nonideai Circuit SO.O% 45.6% 19.1% capacitance and applying this to the stray components
we could expect a worst case of ±20% to the designed
capacitor and =t=20% to the stray component. Taking
From examination of figure 7 we can see that the into consideration the expected inaccuracy of the stray-
nonideal components yield two distinct forms of dis- capacitors a value of ±50%, though high, would seem
tortion: a reduction of approximately 50% in the pass- suitable to examine the limits of the design.
band ripple level, and a frequency shift of approxi- Stray capacitance does not have a first-order effect
mately 10%. Optimization of the nonideal circuits is upon the circuit as it is not relied upon to achieve filter
the most time-consuming operation hence any method poles, because of this factor it is not so much the value
which minimizes this process will considerably reduce of the stray capacitance but rather its presence in the
design time. By pre-distorting the ideal filter circuit to
circuit.
achieve a 100% increase in the ripple and 10% lower
Despite the wide range in values, the structure per-
passband frequency it was found that optimization time
formed well. As can be seen in figure 10 the Cheby-
was reduced by as much as 66%. Savings from the new
shev function is well maintained but with an increase
algorithm also increased to as much as 50%, however,
in passband ripple at higher frequencies. All the values
the generality of this pre-distortion technique has not
are normalised to OdB as the tolerance exerts a range
yet been investigated.
of up to ldB error in the DC gain.
The optimization algorithm was used with approx-
If a more lenient tolerance of ±25% is used, only a
imate stray capacitance included. After layout within
small amount of deviation in the pasband occurs. This
the EDGE environment, actual stray capacitance was
is illustrated in figure 11.
extracted and the circuit submitted to a final optimiza-
tion run. As the starting point for the design is near
to the solution, we would expect the design objective
function to be well approximated by a quadratic form. 8. Further Work
Predictably, then, it was found that the Jacobian matrix
had to be calculated once only, after which the algo- Biquad and ladder structures each have their own ad-
rithm effected a solution using the initial Jacobian ma- vantages and disadvantages within a filter design. Each
trix with only five iterations. Optimized passive circuit can be used to good effect for particular applications.
components had the values shown in table 5. The limiting factor upon biquads is the sensitivity to
After extraction of the final layout-component val- component tolerance. High-order filters contain pole-
ues, the design was subjected to a series of rigorous pairs with high-Q values which are more difficult tore-
tolerance analyses. Unit capacitors where subjected to alize accurately using biquad structures. Alternatively,
50
Optimal Design of SC Filters for Analog/Mixed-SignaiiCs 227
0.5
-0.5
Gain
dB
-1
-1.5
-2
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Frequency (Hz)
Fig. 10. Bft'ect of ±191> nndom component and ±5091> stray-capacitance tolerance upon an optimized 7th-order ladder filter.
0.5 I I
-0.5
Ga.in
dB
-1 •,'
·~.~ ~ ::·.
: :
: :
::
-1.5
-2
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Frequency (Hz)
Fig. II. Bft'ect of ± 191> nndom component and ±2591> stray-capacitance tolennce upon an optimized 7th-order ladder filter.
ladder structures can easily assimilate high-order struc- structures to achieve the best of both worlds. Biquads
tures but are more difficult to design with; further, can be used for the low-Q sections, which they can
they are sensitive to stray capacitance. As we have efficiently achieve, whereas ladder structures can be
shown, this can be corrected with optimization, but an used for the high-Q sections. Direct methods to de-
alternative is the amalgamation of biquads and ladder sign ladder filters for such a use do not exist however,
51
228 Gustard and Massara
optimization should be effective for such a situation. 10. Szentirmai, G . and Ternes, G. C ., "Switched-capacitor build-
Ing bloclcs," IEEE Transactions on Circuits and Sy.flems,
Work is in progress on this area at the present time.
Vol. CAS-27, pp. 492-501 , June 1980.
II. Gregorian, R., Martin, K. W. and G. C. Ternes, K. W.,
"Switched-capacilor circuli design," Proceedings ofthe IEEE,
9. Conclusions Vol. 71, pp. 941-966, Augusll983.
12. Jacobs, G. M., Allstol, D. J., Brodersen, R. W. and Gray, P. R.,
"Design techniques for MOS swilched-capacilor ladder fil-
A modified algorithm for enhanced optimization of lers," IEEE Transactions on Clrcuir. and Systems, Vol. CAS-
simulation-intensive problems has been presented in 25, pp. 1014--1021, December 1978.
13. Brodersen, R. W., Gray, P. R. and Hodges, D. A., "MOS
this paper. Experimental results applied to three ladder
swilched-capacilor filters," Proceedings of the IEEE, Vol. 67,
structures have been discussed to verify the improve- pp. 61- 75, January 1979.
ments. Pre-distortion techniques for nonideal struc- 14. Hoslicka, B. J. and MoschyiZ, G . S., "Swilched-capacilor fil-
tures have been suggested which further enhance the ters using FDNR-Lilce super capacitances," IEEE 7>-ansactions
on Circuits and Systems, Vol. CAS-27, pp. 569-573 June
speed of the algorithm. 1980.
Further results demonstrating the use of the routine IS. Orchard, H. J., "lnduclorless fillers," Electron. lLII. , p. 224,
applied to a 7th-order ladder-filter have been presented, 1966.
where effective compensation for the nonideal compo- 16. Mietec n.v., WeslerringlS, B-9700, Oudenaarde, Belaium.
17. Cadence design systems, Inc. San Jose, CA.
nents present in silicon realization has been success- 18. Fang, S. C., Tsividis, Y. P. and Wing, 0., "SWITCAP: a
fully achieved. switched-capacllor network analysis program Part 1: basic
Component tolerance has been applied to biquad features," IEEE Circuits and Systems Magazine, pp. 4-10,
September 1983.
structures and ladder structures, the results of which 19. Fang, S. C., Tsividis, Y. P. and Wing, 0., "SWITCAP: a
have been presented and compared. swilched-capacitornelworlc analysts program Part II: advanced
applicalions," IEEE Circuits and Systems Magazint, pp. 41-
46, December 1983.
20. Gauss, K. F., "Theoria motus corporum coelistiam,"ln Weru,
References Vol. 7, pp. 240-254, 1809.
21. Massara, R. E., Optimization Methods in Electronic Circuit
I . Wu, K. K. and Mack, R. J ., "FEEDS: a fledble and expandable Design. New York: Wiley, 1991.
expert design syslem for analogue design automalion," 5'" 22. Massara, R. E. and Fidler, J. K., "An efficient damping method
Int. Symp. on Integrated Circuit De.ign, Singapore, Septem- forleasl squares algorithms," Electron. l.Ltt, Vol. II, pp. 33-34,
ber'l993. 1975.
2. Gregorian, R. and Ternes, G. C., Analog MOS Integrated Cir- 23. Levenberg, K., "A method for solution of certain nonlinear
cuits for Signal Processing. New York: Wiley-lnterscience, problems in leasl squares," Quart. Appl. Math., Vol. 2, pp.
1986. 164-168, 1944.
3. O'Leary, P, "Practical aspec!B of mixed analogue and digital
design," In Analogue-Digital ASICs-Circuit Techniques, De-
sign Tools and Applications, edited by R. S. Soin, F. Maloberti
and J. Franca, chapter 10, pp. 213-238. London: P. Peregrinus
Ud, 1991.
4. Ghausi, M. S. and Lalccr, K. R., Modem Filter Design: Active
RC and Switched Capacitor. Prentice-Hall, Englewood Cliffs,
NJ, 1990.
5. Geiger, R. L., Allen, P. E. and Strader, N. R., Vl..SI Design Tech-
niques for Analog and Digital Circuits. McGraw-Hill, New
York. 1990.
6. Gregorian, R., "Switched-capacitor filter design using cas-
cade sections," IEEE Transactions on Circuits and Systems,
Vol. CAS-27, pp. 515-521, June 1980.
7. Martin, K., "Improved circuits for the realization of switched- Robert E. Massara (born in Sheffield, England, 1947) ob-
capcaitor filters," IEEE Transactions on Circuits and Systems, tained a first clas s honours degree in Electronic Engineering
Vol. CAS-27, pp. 237- 244, April 1980. from the University of Essex In 1972, and a Ph.D. in 1977.
8. Martin, K. and Sedra, A. S., "Exacl design of swltched- He worked for GEC Semiconductors and lhe University of
capcaitor bandpass fillers using couplcd-biquad structures," Keele before laking a post as a Lecturer at Essex. He was
IEEE Transactions on Circuits and Sysr.ms, Vol. CAS-27, pp.
promoted 10 Senior Lecturer in 1984 and to Professor in 1991.
469-475, June 1980.
He is currently Head of lhe Department of Electronic Sys-
9. Nossek, J. A. and Ternes, G . C ., "Switched-capacitor filler de-
sign using bilinear element modeling," IEEE 1>-ansactions on tems Engineering, and of lhe Department's Neural and VLSI
Circuits and Systems, Vol. CAS-27, pp. 48~-491, June 1980 Systems Research Group, and is Director of lhe University's
52
Optimal Design of SC Filters for Analog/Mixed-Signal ICs 229
Centre for VLSI Systems Design. His research Interests In- Systems Engineering from the University of Essex in 1991.
clude analogue VLSI design automation; IC filter design and Since 1991 he has been working towards a Ph.D. degree in
numerical engineering design methods. He Is a Fellow of the the area of switched-capacitor filters for integrated circuit
(UK) Institution of Electrical Engineers. realization at the University of Essex. Mr. Oustard Is an
Associate Member of the (UK) Institution of Electrical En-
gineers.
53
Analog Integrated Circuits and Signal Procesalng, 6, 231-241 (1994)
0 1994 Kluwer Academic Publishers, Boston.
caloba@cot.ufrj.br
Abstract. This paper presents a method for the overdesign of gain-shaping filters in an optimal way. The passband
minima and the stopband maxima of a filter approximation are adjusted using an iterative method, with the objective
of obtaining a circuit that satisfies exactly the specifications when the statistical deviation in the filter transfer function
due to errors on its components is considered. The result is a filter that can be significantly more selective than what
is obtained by a simpler overdesign method that results in an equal-ripple ideal design. Two different approaches
are examined, one based in the development of a special approximation followed by synthesis, and the other by
direct optimization of an initial classical design.
55
232 de Queiroz and CaMba
i..,\
: \
l ~\\
1 \.
: : ·~ .
.. ., ......(••······;···'-'~.:···.
dB
Fig. 2. Passband gain curve of a normalized elliptic 7th-order low-pass filter, and the gain curve with the statistical deviation due to S% errors
In all the element values subtracted, with the flat-loss error discounted. The maximum gain was normalized to 0 dB, as in the other figures.
56
Optimal Gain Overdesign in Analog Filters 233
Fig. 3. Stopband (beginning) gain curve for the same filter, shown also with the statistical deviation added, discounted of the flat-loss error.
In an LC doubly terminated filter, it is easy to elim- to obtain a more selective overdesigned filter. This can
inate from the errors the flat-loss errors that affect only be done by computing the overdesigned approximation
the gain of the filter and are easy to compensate. They with irregular passband and stopband ripples, in a way
are caused only by the errors in the terminations at any that makes the worst-case resulting filter equal-ripple.
frequency Wm where maximum power transfer occurs. The idea is exemplified in the example below.
With the sensitivities of the terminations discounted of
this value (i. e., replacing si,<iw) by si,<Jw>- si,Uwm)
in (1)), the gain statistical deviation ~G(w) in (I)
3. Example 1
reduces to zero at the maximum power transfer fre-
quencies, that is, at the passband gain maxima for a
A 7th-order elliptic filter realized as a doubly termi-
well-designed filter. With this discounted error used,
nated LC ladder presents the passband and stopband
the passband gain limits are obtained by subtracting
gains and errors shown in figures 2 and 3. The objec-
the error curve from the nominal gain curve, since the
tive is to obtain a similar filter that in the worst case is
maximum power transfer prevents any increase in the
equal-ripple in the passband and stopband. An algo-
passband gain.
rithm for obtaining an optimally overdesigned filter is:
In the stopband, the maximum gain (minimum at-
I. Compute the worst-case gain curve, as the filter gain
tenuation) limit is obtained by adding the error curve curve (dB), minus the flat-loss (-6.0206 dB in the
to the nominal gain curve. The discount of the flat- example), minus the gain statistical deviation (dB,
loss error can also be used. The infinite sensitivities at with flat-loss errors discounted) in the passband, and
the transmission zeros can be ignored, by considering plus the gain statistical deviation in the stopband, as
only the error at the stopband maxima and borders, or discussed in the last section. If it is equal-ripple,
by using "slope-normalized" sensitivities [2]. stop. Else, call E; the gains at the passband minima
When the errors in an overdesigned equal-ripple fil- of this composite curve:
ter are considered, clearly the worst-case (in this case
in statistical sense) resulting filter is not equal-ripple
at all. It is possible, however, to use the fact that er- E; = G(w)- flat-loss - ~G(w;) dB
rors are higher at the passbands and stopbands limits (2)
57
234 de Queiroz and Cal6ba
Table 1. Nominal passband minimum gains A;, stopband maximum gains A;, and corresponding gains with worst-case error considered, E;
and E;, for the six first steps of the optimization algorithm of the example.
-AJ/ -Et -A2/- E2 -Aa/- Ea -A~/- E; -A;/-E~ -A;/- E~
I I I 50 50 50
1.154 1.341 2.382 40.63 47.51 49.08
0.866 0.746 0.420 61.78 52.62 50.93
2
1.012 1.049 1.154 48.64 50.10 50.00
0.856 0.711 0.364 63.51 52.51 50.93
3 1.037
1.001 1.009 49.23 50.01 50.00
0.855 0.705 0.351 64.50 52.50 50.93
4 1.008
1.000 1.002 49.52 50.00 50.00
0.855 0.703 0.348 65.13 52.49 50.93
5
1.000 1.000 1.001 49.68 50.00 50.00
0.855 0.703 0.348 65.55 52.49 50.93
6 1.000 49.77 50.00 50.00
1.000 1.000
and call E:
the gains of the composite curve at the is slower near the passband and stopband borders. The
stopband gain maxima: final element values, with the same structure of figure I,
are listed in table 2(b).
E; = G(w)- flat-loss - ~G(w;) dB (3)
Table 2. Element values for the example filters.
58
Optimal Gain Overdesign in Analog Filters 235
Fig. 4. Optimum overdesigned filter passband gain, with flat-loss discounted, shown also with the gain statistical deviation subtracted.
Fig. 5. Bqoal-rlpple 7th-order and 8th-order overdesigned filters passband gains, as in figure 4.
for the two filters are are in table 2(c) (7th-order) and The optimum overdesigned filter is more selective
2(d) (8th-order). The structure for the 8th-order filter than the equal-ripple one of the same order, and slightly
is the one in figure 1 with an additional inductor, L 1 , less selective that the 8th-order uniform-ripple overde-
in series with the input termination Rt. signed filter, as can be observed in figures 6 and 7.
59
236 de Queiroz and CaMba
o -\~---··r··-----~-~r··~---··:···-·-·····:···-·---·--r·······---!-·-·······r···------·~---····--r
. -~ ·\" --···:· ······;·::-· ··::·· ... :···· .. ··-·:· ........... ··r······· ··-~·-···· ...... ~.. -· ..... ····:·-· ...... ··:
\ : : :: :: . : ; : : ; : :
\ ~ i i; i \ ; ; i . . ~ i
Fig. 6. Optimum overdesigned filter stopband gain, with Oat-loss discounted, shown also with the gain statistical deviation added.
Fig. 7. Uniform-ripple 7th-order and 8th-order overdesigned filters stopband gains, as in 6gure 6.
A frequency scaling shall be done to adjust the tran- ing, the method would find a solution for any values of
sition band of the overdesigned filter in the specifica- Amax and Amln· The limit solution would be a Butter-
tions (not done in the example). This will be discussed worth approximation, frequency scaled as needed tore-
in the next section. Note that, with this frequency seal· duce the worst-case passband and stopband errors. Of
60
Optimal Gain Overdesign in Analog Filters 237
course, the order could have to be increased to satisfy If the resulting selectivity does not satisfy the filter re-
the transition band width specification, and the order quirements, a higher order must be used.
increase would increase the error at the passband and
stopband borders. In critical cases a solution can be
inexistent, by this or by any other overdesign method. 5. Direct Optimization
j
corrections.
n, = 1 + AG(1) d~1 > (8)
3. For usual filter structures, the number of extremes
needed for the complete characterization of the filter
gain curve, plus values that define frequency scal-
When the passband border is renormalized to ing and bandwidth, is precisely equal to the number
1 radls, the worst-case stopband border is n = n,;n•. of reactive elements. In these cases the system of
61
238 de Queiroz and CalOba
equations (9) is directly solvable (if not singular). If ripple, and an inferior rejection band with 2 transmis-
there are more reactive elements than needed, or if sion zeros and minimum attenuation of 50 dB was com-
the terminations are included among the x,, the sys- puted directly by the optimization algorithm described,
tem (9), in the form Sv = e, has more unknowns from an initial approximation with the correct shape,
than equations, and there are many solutions. A obtained by estimating empirically the element values
convenient solution is obtained by transforming the of the structure in figure 8.
system Of equations into SST Z = e, with V (the
!:ix1 fx 1 vector) recovered as v = sT z. It can be
shown that this transformation computes the vector
v with the minimum possible modulus that satisfies
Sv=e.
4. Updatetheelementvaluesbyx, .- x,(l+!:ixtfxi),
and return to 1. Fig. 8. Overdesigned irregular bond-pass filler. Pinal element val-
This algorithm is simpler, and less computationally ues in table 2(e}.
intensive than the previous one, since it avoids the syn-
thesis of a new approximation, that is also done by The filter was overdesigned by the consideration of
optimization, and the filter synthesis. It presents, how- the errors at the two stopband maxima, the two pass-
ever, some numerical problems. The most serious is band minima, and the passband borders, with the ter-
that the LC doubly terminated realization with max- minations fixed, and 5% tolerances assumed for all the
imum power transfer is an "extreme" solution for an elements. The final element values are listed in table
approximation. The gain sensitivity matrix S becomes 2(e). The obtained gain curve is shown in figure 9, with
singular at the solution. It is obvious that the lines the passband detailed.
of the matrix corresponding to passband maxima are
null, but there are also other relations that tum the ma-
trix singular even if the maxima are substituted by other 7. Conclusions
frequencies. The problem can be avoided by ignoring
the equations of the maximum power transfer extremes 1\vo methods for optimized overdesign of gain-shaping
when the gain is close enough to the desired value, with filters with maximum-ripple passbands and stopbands
the transpose matrix technique (step 3 above} used to was presented, with the algorithms described. The first
correct the number of unknowns in the system of equa- algorithm is more precise, although more computation-
tions. This problem does not exist for filters without ally intensive. The second algorithm is simpler and
maximum power transfer, but another problem appears: more general, but somewhat numerically problematic.
Because there is no relation among gain minima and The examples used the statistical deviation as a mea-
error maxima, and the discounting of the flat-loss er- sure of the worst-case error. A true worst-case error,
ror can't be made, the number of extremes in the gain or any other kind of error could also be used exactly in
curve and in the gain-error curve in the passband can the same way. The method, specially with the second
be different. For any filter, the overdesign of pass- algorithm, can easily be extended to the overdesign of
band and stopband borders is also problematic, since phase characteristics, and can also be applied to other
the assumed proportionality between attenuation and purposes than to only compensate for random errors in
error is not valid in these regions. The multiplication the component values. Deterministic effects, as the ef-
of the computed !:ixtfx, by a reduction factor (about fects of losses and parasitic poles in active realizations
0.5) is usually necessary to force the convergence in can also be compensated.
these cases.
6. Example2 Appendix
This example illustrates the overdesign by direct opti- In this appendix, the algorithm for the design of low-
mization. A normalized 6th-order irregular band-pass pass filters with irregular passband and stopband rip-
filter, with passband between I and 2 rad/s, with I dB ple described in [4] is summarized. The algorithm,
62
Optimal Gain Overdesign in Analog Filters 239
10 !l . '
. '
\ j li
j \ ~l
-2
implemented in a computer program, was used in the to be determined, and Y(w) has m 11 = (n- p11 )/2 + 1.
generation of the first example. The general form of the normalized characteristic func-
The desired approximation has a purely real of imag- tion K(jw)/e is:
inary characteristic function K(s) for s = jw, and is
in a form that can be called "asymmetrical rational ap- aX(w)
proximation." Ignoring a possible -1 or ±jw multi- Yr(w)
plying factor, for an nth-order approximation, K(jw)
can be written as: · · · + axp.+2wP•+ 2 + axp.wP• (12)
· · · + aYn-2W 2 + ayn
K( 'w) = eaX(w) = eaX(w) ( 10)
J wnY(1/w) Yr(w) K(jw) presents the maximum possible number of
ripples in the passband and in the stopband. The
The polynomial Yr(w) is simply Y(w) with the co- normalized characteristic function presents m,-1 ex-
efficients in reverse order. The polynomials X(w) and treme values f xi at the passband, each occurring at a
Y(w) are of the same degree n. The two multiplying frequency wx; (see figure 10). The values of the fxi
factors are functions of the maximum passband atten- are related to the specified passband gain minima Ai
uation Amtn and of the minimum stopband attenuation by:
Amax• and given by:
E: = VlOO.lAmox _ 1 j a=
J yflQO.lAmtn - 1
e
/x; = ±.!.Jw-o.lA, _ 1
e
(13)
63
240 de Queiroz and CaMba
-6~--------~--------~ ~--------~----------~
r ~d/s; 1
0
Fig. 10. Normalized characteristic function of a 7th-order low-pass elliptic filter, showing the meaning of the extreme values fx; and /IIi· 1be
vertical scale is compressed by 11 = ±(1 + Ln IK(jw)fe I) for IK(jw)l > 1.
decreasing in frequency for the others) by: (d) Repeat (a), (b), and (c) until convergence.
The convergence is not particularly fast, but is very
ea 2
fy; = ± VlO-O.lA,' - 1
(14) robust. Once obtained X(w) and Y(w), the charac-
teristic function K ( s) is obtained by replacing w in
K(jw) by sjj, and ignoring a possible -1 or ±j mul-
The center of the transition band can be fixed at
tiplying term. The approximation synthesis is com-
w = 1 by fixing X(1) = 1 and Y(1) = 1. (As im- pleted in the classical way by the solution FeldtKeller's
plemented, a frequency scaling is done after the initial
equation H(s)H(-s) = K(s)K(-s) + 1 to find
design to move the passband border tow = 1.) As
the transducer function H (s ). Note that the algo-
initial approximations for the polynomials, X(w) =
rithm can generate, as particular cases, all the clas-
wP·- 1 Cn-p,+l(w) and Y(w) = wP•- 1 Cn-v.+l(w),
sical low-pass approximations with a real character-
where Ck(w) is the kth-order Chebyshev polynomial,
istic function. The elliptical approximations, for ex-
are used. The iterative algorithm to refine the initial
ample, are obtained with Px = p 11 = 1 for odd
approximation is:
orders, or p,. = p11 = 0 for even orders, and the f x;
(a) Findwx;, k = 1, ... , m, -1, them, -1 roots and fy; alternating between ±1, with the last equal to
immediately smaller than I of the polynomial -1, which results in X(w) = Y(w). By increasing
Yr(w)- X'(w)- X(w)Y:(w), frequencies of p,. and p 11 , the same algorithm can generate Cheby-
the maxima and minima of aX(w)jY,.(w) for shev approximations (p11 = n ), Inverse Chebyshev ap-
o:::;w:::;l. proximations (Pz = n), Butterworth approximations
the system of linear equations: (p,. = p 11 = n), and many intermediate variations. In
(b) Solve
Example I, the fx; and fy; of initially elliptic approx-
aX(wx;) = fx;Yr(wx;), k = 1, ... , m, -1; imations were computed by (13) and (14), using the
X(1) = 1, finding. the new approximation for
gains in table 1.
the coefficients of X (w).
(c) Interchange X(w) with Y(w), wx; with wy;, References
fx; with fy;, p, with p 11 , and m, with m 11 , I. Orchard, H. J., "Inductorless filters," Electronics Letters,
and repeat steps (a) and (b). Vol. 2, pp. 224-225, June 1966.
64
Optimal Gain Overdesign in Analog Filters 241
2. Fidler, J. K. and Nightingale, C., "Slope-nonnalized sensitiv- Electronic Engineering and at the Electrical Engineering Pro-
Ity: A new sensitivity measure," Electronics utters, Vol. IS, gram of the Coordena~iio dos Programas de P6s-gradua~lo
pp. 54-56, January 1979. de Engenharia (COPPE). His main interests are in general fil-
3. de Quelroz, A. C. M. and CaiOba, L. P., "Physically symmetri-
cal and antimetricalladder fillers with finite transmission u:- ter theory, monolithic filter realizations, and computer-aided
ros," Proc. 30th MWSCAS, Syracuse, USA, pp. 639--643, 1987. synthesis and design of electronic circuits.
4. de Queiroz, A. C. M. and CaiOba, L. P., "An approximation al-
gorithm for irregular-ripple filters," Proc. STBREEE Interna-
tional Teleco1111t1UIIications Symposium. Rio de Janeiro, Brazil,
pp. 430-433, 1990.
65
Analog Integrated Circuits and Signal Proci!BIIing, 6, 243-263 (1994)
0 1994 Kluwer Academic Publishers, Boston.
M.ATARODI
UMGr T<clrnology Corporation, Milplttu, CA
J. CHOMA, JR.
Unlverdry of Southern California, DepartrMnt of Electrical &glneerlng-Eiectrophyslc.r, Unlverrlt)l Pork, MC: 0271, Los Angeks, CA, 90089-0271
}ohnc(lmi:Jlr.usc.edll
Abstract. Using a complementary bipolar junction transistor process having NPN transistors with a maximum
short circuit common emitter gain-bandwidth product (/T) of 7.2 GHz and PNP transistors with a maximum IT
of 4.5 GHz, an operational transconductance amplifier has been designed for a 3-dB bandwidth of 7.2 GHz. The
design process invokes new phase compensation strategies and develops innovative new ways of exploiting existing
broadbanding techniques. The utility of the design is confirmed by demonstrating its application in two operational
transconductance amplifier-capacitance filters. One of these examples is a 225 MHz lowpass filter, while the other
is a bandpass filter with a center frequency of 250 MHz.
67
244 Atarodi and Choma, Jr.
effective low frequency forward transconductance is The neutralization afforded by the capacitors, Gel
largely determined by the emitter degeneration resis- and Cc2 , is imperfect because transistors Ql and Q2
tance, R... Four broadbanding schemes are incorpo-
have non-zero base resistances, and each has a net emit-
ter impedance comprised of internal emitter resistance
rated into the amplifier.
appearing in series with the parallel combination of R ••
The first of the broadbanding mechanisms is shunt
peaking via the capacitances, C.e~ and Cee2. which
and c••. It follows that the object of neutralization is
not merely a base-collector junction capacitance. It is,
are nominally the same value, say c••. These capaci-
in fact, a complicated impedance whose intrinsic ele-
tances create a left half complex plane zero in the dif-
ments are functions of device parameters that are diffi-
ferential transconductance response at a frequency of cult to predict analytically. In order to offset the Miller
1/R ••c... The capacitance, c•• , is chosen so that multiplication that resultantly prevails despite the use
its associated zero lies close to the dominant pole es- of Gel and Cc 2 , a third frequency compensation scheme
tablished in the absence of c••. thereby effecting at in the form of a folded common base cascode is used.
least partial cancellation of this dominant pole. A sec- This cascode, which is made up of transistors Q6 and
ond broadbanding vehicle is afforded by the shunt- Q7, is used to couple the differential output current of
anti phase shunt connection of the capacitances, Cc 1 Ql and Q2 to the load. In the subject amplifier, the
and Cc2· If these nominally identical capacitances are load is a constant current sink formed of transistors QS
equal to the intrinsic transistor capacitance between through Q II. Observe that the differential resistance
each base and corresponding collector of the differ- presented at low signal frequencies to the emitters of
ential pair, the effective base-collector capacitances of transistors Q6 and Q7 is extremely high owing to the
Ql and ofQ2 are ideally nulled, or neutralized. Since current sinking emitter degeneration of QS and Q9 that
Miller multiplication in a phase inverting amplifier typ- is afforded by the transistor pair, Q 1Q...Q 11.
ically postures the base-collector capacitance as the The fourth broadbanding vehicle utilized in the OTA
most significant contribution to the amplifier dominant is feedforward compensation via the capacitances,
pole, such neutralization can substantially broadband C11 1 and C112· These capacitances, which are nom-
the frequency response of the uncompensated ampli- inally equivalent, bypass the PNP transistors, Q6 and
fier. Q7, at high frequencies. They therefore circumvent the
68
A 7.2 GHz Bipolar Operational Transconductance Amplifier for Fully Integrated OTA-C Filters 245
response degradation problems incurred by a PNP tran- that the differential voltage between the collectors of
sistor gain bandwidth product, fT, that is measurably Q8 and Q9 is kept small by the high voltage gain de-
smaller than its NPN counterpart, despite a comple- veloped by the transistor quad, Q14 through Q 17. This
mentary fabrication process capability. high gain is realized by the active current source, Q18,
The output signal response is extracted as a volt- that loads the collectors of transistors Ql5 and Ql6.
age developed across a single ended capacitive load Three noteworthy design-oriented points can be of-
that terminates the collector of transistor Q7 to ground. fered with respect to the common mode feedback sub-
The current that flows through, and thus charges, this circuit. First, the specific value of the quiescent volt-
capacitive load is essentially the signal current flowing age developed at the collectors of Q8 and Q9 can be
through the collector of Q7, since the differential and adjusted by applying a suitable static voltage, Vcm• at
common mode impedances seen at the collectors of the junction of the bases of transistors Q15 and Ql6.
transistors Q8 and Q9 are very large. Accordingly, the Under normal operation of the subject orA, Vem = 0.
capacitively loaded OTA behaves as an integrator; that Second, the use of degeneration resistances in the emit-
is, the output signal is the integral of the differential ters of Ql4 through Ql7 ensures minimal capacitive
input voltage. loading of the collector nodes at Q8 and Q9. This
The final major design issue is biasing. In particular, minimization is essential if the pole established by the
transistors Q4 and Q5 supply the biasing current com- integrating external load capacitance is to be predicted
mensurate with the operation of transistors Q6 through accurately. To this end, it might be interjected that a
Qll in their linear active regimes. But the quiescent fabrication process featuring an insulating or a semi-
voltages developed at the collectors of transistors Q6 insulating substrate is desirable since the substrate ca-
and Q8 and at the collectors of transistors Q7 and Q9 pacitances of transistors Q7 and Q9 also act as para-
are difficult to define reliably because these voltages sitic load susceptances. Finally, the capacitances used
appear at the junction of a current source (Q6 or Q7) across the four emitter degeneration resistances in Q14
through Q 17 provide slight gain peaking at high signal
and a current sink (Q8 or Q9). The voltage predictabil-
frequencies. Such peaking is useful in that high com-
ity problem is exacerbated by relatively poor matching
mon mode subcircuit gain is essential to minimize the
among like model parameters in the PNP and the NPN
offset in the differential voltage established between
units of a complementary BIT process.
the collectors of Q8 and Q9.
To rectify the foregoing bias voltage uncertainty,
common mode feedback is implemented with the dif-
ferential to single ended converter formed of transistors 3. OTA Analysis and Simulation
Ql4 through Ql9. In an attempt to demonstrate the 3.1. Transconductance
effectiveness of this feedback, suppose that the quies-
cent voltage at the collector of Q8 or at the collector of The low frequency forward transconductance, Gmo• of
Q9 rises due to thermal drift or parametric uncertainty. the orA is the ratio of the signal current, 10 , , flowing
Since the collector of Q8 and the collector of Q9 are through the extrinsic capacitive load to the differential
connected to the base of Ql4 and Ql7, respectively, input voltage signal, V.. Assuming balanced operating
and since the collector of Q 14 is tied to the collector conditions, a conventional analysis of the differential
of Ql7, the voltage at tl)e collector of QI7 decreases. mode half circuit of the orA yields [25]
It follows that the voltage at the collector and base of
the diode-connected transistor, Q 19, and at the base of G _ lo, _
mo - V. - 2
(a:6) ( Ree a:1+ Tibl ) (l)
transistor Q 18 decreases by the same amount. But this
decrease is met be an increase in the voltage at the node where 0:1 is the small signal, short circuit, common
at which the collector of Q18 and the bases of Q4 and base current gain of transistor Ql (and Q2), a:6 is the
Q5 are incident. In tum, the voltages at the collectors corresponding current gain parameter for the PNP tran-
of Q4 and QS decrease. Since Q6 and Q7 operate as sistors, Q6 and Q7, and r ib1 symbolizes the small signal
common base structures, nearly the same voltage de- common base input resistance of transistors Q 1 and Q2.
crease is seen at the collectors of Q6 and Q7, thereby For large R ••• this transconductance is largely deter-
offsetting the original increase in voltage assumed at mined by R •• itself. On the other hand, small R.e ren-
the latter collector nodes. Equally important is the fact ders Gmo strongly dependent on r;b1• whose value can
69
246 Atarodi and Choma, Jr.
-30
-40
--------------------- ........ .
i
'-J -50
~~.-
~
-60
-70
-80L--L-L~LUil-~~-LLUliL-~~~~-~-L~~
be adjusted through variations in the quiescent current, 3-dB passband of the OTA-C filter in which the OTA
Ix, conducted by the current sinking transistor, Q3. is utilized. This noise dilemma is serious in that VinT
Using BIT parameters provided by the Texas Instru- imposes a lower bound on the minimum signal that
ments Corporation, simulations confirm the accuracy can be processed reliably by the OI'A. Moreover, for
of (1). Moreover, and as depicted in figure 2, these low frequency noise analysis purposes, Ree effectively
simulations suggest 3-dB bandwidths well in excess of appears in series with the internal transistor base resis-
a gigahertz. tance, Tb, whose thermal noise in the absence of R ••
generally dominates the value of V;nT [27]. The sim-
ulated data itemized in table 1 underscores the delete-
3.2. Noise rious impact exerted by R •• on the noise properties of
theai'A.
In addition to increasing the 3-dB bandwidth and sta-
bilizing the low frequency forward transconductance
of the OTA, the emitter degeneration resistance, R.., 3.3. Distortion
increases the linear range of OTA operation; that is,
An especially important figure of merit for an OI'A uti-
the negative series-series feedback implicit to R.e re-
lized in filtering applications is its dynamic range, DR,
duces circuit distortion [26]. Unfortunately, increases
which is defined as
in Ree are accompanied by increases in the integrated
total root mean square (RMS) input referred noise volt-
age, VinT• where the integration is conducted over the
DR=20log V.m)
(v,-
mT
(2)
70
A 7.2 GHz Bipolar Operational Transconductance Amplifier for Fully Integrated OTA-C Filters 247
3.16 148
2000 10 301
316 2120 If the frequency and phase responses of the uncom-
pensated (meaning c•• = 0) OTA are known, c•• can
where Vim is the maximum RMS input signal ampli- be selected so that the phase contribution of the intro-
tude that produces a total harmonic distortion (THO) duced zero offsets the phase angle associated with poles
of less than 1%. Note that DR is essentially a loga- of the uncompensated OTA. Implicit to this selection
rithmic comparison of the maximum allowable input process is the requirement, R.. :;$> r ibl• so that the
signal amplitude (for THO < 1%) and the minimum pole time constant in {3) is significantly smaller than
detectable input signal amplitude. both the time constant, R•• c•• ,and the dominant time
Since the OTA designed in this work is a fully differ- constant of the poles implicit to the uncompensated
ential circuit topology, even order output harmonics of OTA. Note, however, that large R.e increases the input
a single frequency input test sinusoid are theoretically referred noise and diminishes Gmo. Accordingly, ca-
eliminated. With Ree = 2 KO, and at an input signal pacitive shunt peaking is not completely effective, and
frequency of 3 MHz, the simulated THO is about 1% additional phase compensation schema are mandated.
for Vim = 3.53 volts. For R.e = 150 nand at an input
signal frequency of 300 MHz, V.m is 99 mV. Using
table 1, it follows that the dynamic range at 3 MHz and
for R •• = 2 KO is 87.5 dB. On the other hand, DR=
54.3 dB at 300 MHz and for R•• = 150 0.
71
248 Atarodi and Choma, Jr.
72
A 7.2 GHz Bipolar Operational Transconductance Amplifier for Fully Integrated OTA-C Filters 249
-38~-------------------------------------,
,~·;;·:-~- .. :~
-42
,,}~"/ \ \\
_____ ..... ..,.~ \\
-46 \\\1
""\I
-SO 'I
"
'I
'.I
\I
\I
'""' -54 '.I
\\
~ •I
I
I
c§ -58
I
I
I
I
I
I
I
I
-62 I
I
I
I
I
I
-66
1-( I
I
I
I
I
I
I ) -- .. ( 2) ---· ( 3)
-70
-74L--L~~~~~_L~~ll-~~~~~~~~~~
73
250 Atarodi and Choma, Jr.
(b)
0.6
0.4
0.2
I
0
§ -0.2
f
-0.4
-0.6
-0.8
-I
IE7 1E8 1E9
frequency (Hz)
Fig. 5. Phase response, with respect to a -90° reference (conesponding to an ideal Integrator) of the OTA after: (a) capacitive shunt peaking;
(b) both capacitive shunt peaking and feedforward capacitive compensation.
features, the phase response of the capacitive termi- 8. Meyer, R. G., Eschenbach, R. and Chin, R., "A wideband ul-
tralinear amplifier from DC to 300 MHz," IEEE J. Solid-State
nated OTA must be maximally flat at -90° (the phase
Circuits, Vol. SC-9, pp. 167-175, Aug. 1974.
angle of an ideal integrator) to within about ± 1° over
9. Allen, P. and Terry, M. B., "Use of cunent amplifiers for high
the passband of the desired filter. performance voltage applications," IEEE J. Solid-State Cir-
cuits, Vol. SC-15, pp. 155-162, 1980.
10. Meyer, R. G. and Blauschild, R., "A four terminal wideband
References monolithic amplifier," IEEE J. Solid-State Circuits, Vol. SC-
17, pp. 634-638, Dec. 1981.
I. Grebene, A. B., Bipolar and MOS Analog Integrated Circuit
II. Choma, J., Jr., "Gain and bandwidth characteristics of a vari-
Design, chap. 8. New York: Wiley-Interscience, 1984.
able gain, actively neutralized, differential pair," IEEE Trans.
2. Solomon, J. E. and Wilson, G. R.. "A highly desensitized
Circuits and Systems, Vol. CAS-33, pp. 66-71, Jan. 1986.
wideband monolithic amplifier," IEEE J. Solid-State Circuits,
Vol. SC-I, pp. 19-28, Sept. 1966. 12. Choma, J., Jr., "Simplified design guidelines for dominant pole
3. Gilbert, B., "A new wideband amplifier technique," IEEE J. amplifiers peaked actively by emitter or source followers,"
Solid-State Circuits, Vol. SC-3, pp. 353-365, Dec. 1968. IEEE Trans. Circuits and Systems, Vol. 36, pp. 1005-1010,
4. Mataya, J. A., Haines, G. W. and Marshall, S. B., "IF am- July 1989.
plifier using Cc-compensated transistors," IEEE J. Solid-State 13. Armijo, C. T. and Meyer, R. G., "A new wide-band Darling-
Circuits, Vol. SC-3, pp. 401-407, Dec. 1968. ton amplifier," IEEE J. Solid-State Circuits, Vol. SC-24, pp.
5. Camenzind, H. R. and Grebene, A. B., "An outline of design 1105-1109, Aug. 1989.
techniques for linear integrated circuits," IEEE J. Solid-State 14. Beall, W. G. and Choma, J., Jr., "Charge-neutralized differen-
Circuits, Vol. SC-4, pp. 110-122, June 1969. tial pair," J. of Analog Integrated Circuits and Systems, Vol. I,
6. Wooley, B. A., "Automated design ofDC-coupled monolithic pp. 33-44, Sept. 1991.
broad-band amplifiers," IEEE J. Solid-State Circuits, Vol. SC-
IS. Degrauwe, M.G. R., "CMOS voltage references using lateral
6, pp. 24-34, Feb. 1971.
bipolar transistors," IEEE J. Solid State Circuits, Vol. SC-20,
7. Couglin, J. B., Gelsing, R. J., Jochems, P. J. and van der Laak,
H. J. M., "A monolithic silicon wideband amplifier from DC to pp. 1151-1157, December 1985.
I GHz," IEEEJ. Solid-State Circuits, Vol. SC-8, pp. 414-419, 16. Wyszynski, A., Schaumann, R., Szczepanski, S. and Halen,
Dec. 1973. P. V., "Design of a 2. 7 GHz linear OTA in bipolar transistor
74
A 7.2 GHz Bipolar Operational Transconductance Amplifier for Fully Integrated OTA-C Filters 251
-5
-10
-15
-20
..-.. -25
.....
_e.
-30
=8
>
-35
-40
-45
-50
l_l
-55
1E7 1E8 1E9 lElO
frequency (Hz)
Fig. 6. Magnitude response of a lowpass OTA-C filter; the 6-dB loss at low signal frequencies is incurred by source and load termination
resistances.
22. Jost, S. R., "An 850 MHz current feedback operational ampli·
tier," Proc. Bipolar Circuits and Tech. Mtg., pp. 71-74, 1992.
23. Toumazou, C. and Lidgey, F. J., "E•tending voltage-mode am-
plifiers to current-mode performance," lEE Proc. G. Electron.
u~·
Circuits and Systems, Vol. 137, pp. 116-130.
• + -+ 24. Atarodl, M. and Choma, J., Jr., "High frequency fully inte-
3 4 grated OTA-C filters using a 7.2 GHz bipolar OTA," Proc.
- - IEEE 36th Midwest Symp. on Circuits and Systems, 1993.
25. Witherspoon, S. and Choma, J., Jr., "The analysis of balanced
C!L_ ~2 linear differential circuits," IEEE Trans. on Education (to be
published in Dec. 1995).
26. Moree, J. P., Groenwold, G. and van den Broeke, L.A. D., "A
Fig. 7. Basic schematic diagram of a two pole OTA-C
bipolar integrated continuous-time filter with optimized dy-
biquadratic bandpass filter. namic range," IEEE J. of Solid-State Circuits, pp. 954--961,
array technology wilh lateral PNPs." Proc. IEEE International Sept. 1993.
Symp. on Circuits and Systems, pp. 2844-2847, 1992. 27. Moinian, S. and Choma, J., Jr., "The frequency response of
17. Gray, P.R. and Meyer, R. G., Analysis and Dt.dgn of Analog bipolar transistor noise figure," IEEE Trans. on Circuits and
Integrated Circuits. New York: John Wiley & Sons, chap. 6, Systems, Vol. CAS-33, pp. 72-76, Jan. 1986.
1977. 28. Choma, J .. Jr. and Witherspoon, S. A., "Computationally ef-
18. Franco, S., De.rign with Operational Amplifiers and Analog ficient estimation of frequency response and driving point
ICs. New Yodc McGraw-Hill, 1988. Impedance In wideband analog amplifiers," IEEE Trans. on
19. Franco, S., Current Feedback Amplifiers Benefit High Speed Circuits and Systems, Vol. CAS-37, pp. 720--728, June 1990.
Designs. EDN, 1989. 29. Clelger, R. L. and Sanchez-Sinenclo, E!., "Active fllter de-
20. Koulllas, I. A., "A wideband low-offset current-feedback op sign using operational transconductance amplifiers: a tutorial,"
amp design," Proc. Bipolar Circuits and Tech. Mtg., pp. 120- IEEE Circuits and Devices Magazine, pp. 20--32, 1985.
123, 1989. 30. Atarodi, M., Analysi.r and Design of Operational Tronscon·
21. Lidgey, F. J. and Toumazou, C., "Current-mode analogue sig- ductance Amplifiers for Ota-C Filter Applications, Ph. D. Dis-
nal processing," Proc. Bipolar Circuit.• and Tech. Mtg., pp. sertation, University of Southern California, Los Angeles, Cal-
224--232, 1991. ifornia, Aug. 1993.
75
252 Atarodi and Choma, Jr.
O r-------------~----------------~
-5
-10
:0
"C
c: - 15
~
-20
-25
-30 L _L.....I--L...J...-'---L_L_L_J_.I____j_.JL...J...-'---L_L_L_J_.I____j_L__L_1___,_--L.J..J
Fig. 8. Frequency response of the OTA-C bandpass filter; the center frequency is 250 MHz, and the quality factor Is 2; that Is the 3-dB bandwidth
is about 125 MHz.
Mojtaba Atarodl received the B.S.E.E. from Amir Kabir John Choma, Jr. earned his B.S., M.S., and Ph.D. degrees
University of Technology (Tehran Polytechnic) in 1985, and in electrical engineering from the University of Pittsburgh in
the M.S. degree in Electrical Engineering from the Universi- 1963, 1965, and 1969, respectively. He is currently a Profes-
ty of California, Irvine, in 1987. He received the Ph.D. de- sor of Electrical Engineering at the University of Southern
gree from the University of Southern California (USC) on the California, where he teaches undergraduate courses in elec-
subject of analog IC design in 1993, and is currently with the trical and electronic circuits and graduate courses in analog
Linear Technology Corporation working on the same subject. electronics and network theory. Prof. Choma consults In the
76
A 7.2 GHz Bipolar Operational Transconductance Amplifier for Fully Integrated OTA-C Filters 253
areas of wideband analog and high speed digital integrated edited electronic circuit texts, and he is an area editor of
circuit analysis, design, and modeling. Prior to joining the a forthcoming CRC Press handbook on circuits and filters.
USC faculty in 1980, Prof. Choma was a senior staff design Prof. Choma is currently a member of the Board of Gover-
engineer in the TRW Microelectronics Center in Redondo nors for, and the Vice-President- Administration of, the IEEE
Beach, California. His earlier positions include technical Circuits and Systems Society. He is also a Regional Editor
staff at Hewlett-Packard Company in Santa Clara, Califor- of the Journal of Circuits, Systems, and Computers, an As-
nia, Senior Lecturer in the Graduate Division of the Depart- sociate Editor for the Journal of Analog Integrated Circuits
ment of Electrical Engineering of the California Institute of and Signal Processing, and a former Associate Editor for the
Technology, and part time or full time positions with the Uni- IEEE Transactions on Circuits and Systems. Prof. Choma's
versity of Pittsburgh, University of Santa Clara, University of research interests include wideband analog and high speed
Pennsylvania, and University at Los Angeles. Prof. Choma, digital integrated circuit design, behavioral analysis of elec-
the author or co-author of 80 journal and conference papers, tronic systems, integrated device modeling, and engineering
is the 1994 recipient of the Prize Paper Award from the IEEE education in the circuits and systems areas. Prof. Choma, a
Microwave Theory and Thchniques Society. He is the author Fellow of the IEEE, is the recipient of several teaching awards
of a Wiley Interscience text on electrical network theory, and and is a ''Teaching Fellow" in the USC Center for Teaching
a forthcoming Richard D. Irwin text on electronic circuit de- Excellence.
sign. Prof. Choma has contributed several chapters to two
77
Analog Integrated Circuits and Signal Processing, 6, 255-263 (1994)
C 1994 Kluwer Academic Publishers, Boston.
pbaronDIOulfcyvx./oulsvllk.•du
Abstract. This paper derives node expansion methods by which a given passive network and its nodal admittance
matrix are modified by expanding a node into two nodes and introducing a nullor or a dependent source between the
newly created nodes. Node expansion provides a systematic method to introduce active elements in a network. The
elements of the admitance matrix are modified, but the dimensions of the matrix are unchanged. These methods,
which can be applied repetitively, are used to derive filters and oscillators from parental passive networks in a
systematic manner.
be realized using a current conveyor. Assume there exists one node k which has three or
Other active devices require two or more nullors more admittances connected to it. If the node k is ex-
panded into two distinct nodes k' and k" by applying
in their network models. In some cases, these devices
the following procedure:
may be inserted into a network by repetitive application
Procedure ( nullor)
of node expansion. However for simplicity and con-
l. Replace node k with two distinct nodes k' and k"
venience, node expansion methods are developed so such that at least one branch of the original network
that dependent sources can be directly inserted while (which was connected to node k) is connected to
keeping the dimension of the NAM constant. node k' and at least two branches to node k",
79
256 Desai and Aronhime
Ykt
Ytp
k Yk(p+tl
•
•
Fig. 2. Results after node k has been expanded.
= - L"
3. Connect a norator between node k' and ground,
then the new nodal admittance matrix is obtained from
-lout Vq(Ykq) + vk' (3)
q=l
the theorem given below. q#k
1 The nodal admittance matrix of the net-
t
THEOREM
work modified by the procedure (nullor) is obtained by
replacing the admittances Yik (or Yk;) which are now 0 =- Vq(Ykq) + Vk' (q =~+ 1 Ykq) (4)
connected between node k' and ground and between q;~tl q#k
node k' and node i, i = 1, 2, ... , k- 1, k + 1, ... , n,
with zero in the kth row of the nodal admittance matrix Thus the ( n + 1) square intermediate admittance ma-
of the original network. trix has two rows, corresponding to nodes k' and k",
replacing the kth row, and two columns replacing the
Proof: Figure I shows a general node k, k = 1, kth column. Since Vk' = Vk"• the elements of the
2, ... , n, in a passive network that contains n +1 nodes (k + 1)th column are added to the elements of the kth
including the ground node with n ::::>: 3. It is convenient column. There are now (n + 1) equations in (n + 1)
to assume the network's graph is completely connected. unknowns (n node voltages and the norator current),
Thus there are at least three admittances connected to and if we are not interested in the norator current, the
each node. Applying KCL to node k we obtain kth row and the (k + l)th column are eliminated to
obtain ann x n nodal admittance matrix [10]. Hence,
n n+l
o= - L: Vq(Ykq) + vk L: Ykq (2)
the new nodal admittance matrix has for its kth row
the nodal equation for node k". As a result, the admit-
q =I q =I
q#k q#k tances y;,, 1 :::; p:::; (n- 2), which were connected to
node k' are eliminated from the kth row of the nodal
Node k is expanded into nodes k' and k" such that admittance matrix of the modified network.
at least one admittance is connected to node k' and at •
80
Current-Mode Synthesis Using Node Expansion Techniques 257
l
fied so that the norator is connected to node j, j f. k'
Y= or k", instead of to the ground nade, then the modi-
fied nodal admittance matrix is obtained by replacing
Yt2 + Yta + Yt4 -y12 admittances Yllc in the kth row with zeros and adding
[ -Yta
-Yl2 Yt2 + Y2a + Y24 -Y23 them with their signs to the corresponding elements in
the jth row.
-Yta -Yzs Yts + Y2s + Ya4
(5)
THEOREM 2 If the nullator and norator are inter-
If node 3 is now expanded into two nodes 3' and 3" as changed In the procedure (nullor), the nodal admit-
shown in figure 3(b), with admittance y 2a connected tance matrix is modified along the columns instead of
to node 3' and admittances Yts and Y34 connected to the rows as given in Theorem 1 and in the corollary.
node 311, the new nodal admittance matrix is obtained
by replacing Y23 by zero in row three of eq. (5) resulting Proof" From [4), interchanging nullators and norators
in: in a network in which all active elements are modeled
81
258 Desai and Aronhime
by nullors yields an adjoint network whose nodal ad- and k". There are still only n unknown node volt-
mittance matrix is the transpose of the nodal admittance ages as Vk = Ko. Vpq· Since Vk is immediately known
matrix of the original network. • if Ko. and Vpq(Vpq = Vp(n+I) - Vq(n+I)) are known,
a node equation for node k' is not required to solve
The nullor introduced in the passive network has the
for the n node voltages. The new NAM will contain
form of a negative second generation current-conveyor
the node equation for node k" in place of the equation
(CCII-). Other active elements, viz., transconduc-
for node k in the original NAM, and hence, the admit-
tance amplifiers, positive second generation current-
tances Yik, i = 1, 2, ... , k - 1, k + 1, ... , n + 1 which
conveyors (CCII+) and dependent sources, are mod-
are now connected to node k' are replaced with zero
eled using more than one nullor. Dependent sources,
in the kth row of the NAM. Additionally, these same
for example, are modeled with two and sometimes three
admittances Yik are multiplied by Ko. in the kth col-
nullors together with one or more admittances [10].
umn, deleted from the kth column, and then added and
Nullors may be introduced into the network repetitively
subtracted to the corresponding entries of the pth and
provided the procedure (nullor) is satisfied, but, with
qth column respectively, because Vk = Ko. Vpq.
more than two nullors, this technique becomes cum-
It should be noted that after the expansion procedure
bersome. Hence, for convenience and simplicity, node
any admittances Yk(n+l) which are now connected to
expansion theorems are presented which employ de-
node k' will be completely eliminated from the admit-
pendent sources directly.
tance matrix and may be viewed as load admittances
Consider the procedure given below for introducing •
on the controlled source.
a dependent voltage source in a network.
Procedure (VCVSICCVS) The proof for node expansion employing a CCVS
I. Replace any node k with two distinct nodes k' is similar.
and k" such that there are at least two admit- A procedure and a corresponding theorem are now
tances connected to node k" and one admittance provided for introducing a VCCS or a CCCS into a
to node k'. It is assumed that there are m admit- network.
tances connected to node k with m ~ 3. Procedure (VCCSICCCS)
2(a). For a VCVS the controlling voltage is taken be- 1. Replace node k with two distinct nodes k' and k"
tween any two nodes p and q, where p f k' or such that at least one ungrounded (floating) ad-
ground and q can be any node including (triv- mittance which was connected to node k of the
ially) p. original network is connected to node k', and at
2(b). For a CCVS the controlling current is a branch least one ungrounded admittance is connected to
current flowing through admittance Ypq· node k". Grounded admittances that were con-
3. Connect the dependent voltage source between nected to node k are connected to node k".
node k' and ground. 2. Connect node k' to ground.
Then the new nodal admittance matrix is obtained from 3(a). For a VCCS the controlling voltage is taken be-
the theorem given below. tween nodes p and q, where p f k' or ground and
q can be any node including (trivially) p.
THEOREM 3 The NAM of the network modified by 3(b). For a CCCS the controlling current is taken as
procedure (VCVS/CCVS) is obtained by replacing the that current flowing through admittance ypq or
admittances Yik (or Yk;), i = 1, 2, ... , k - 1, k + admittance Yp(n+I)·
1, ... , n + 1, connected to node k' with zero in the 4. Connect the dependent current source between
kth row, multiplying these same admittances still re- node k" and ground.
maining in the kth column by 'Ko. 'for a VCVS and Then the new nodal admittance matrix is obtained from
by 'K&Ypq 'for a CCVS, eliminating these admittances the theorem given below.
Yik in the kth column, and adding and subtracting them
to the pth and from the qth columns, respectively. Note THEOREM 4 The NAM for the network modified by
that 'Ko.' is unitless while the units for 'Kb' are ohms. the procedure (VCCS/CCCS) is obtained by eliminat-
Proof (VCVS): During the process of node expan- ing those admittances Yik• i = 1, 2, ... , k - I, k +
sion node k is replaced with two distinct nodes k' 1, ... , n + I, connected to node k' (now ground) from
82
Current-Mode Synthesis Using Node Expansion Techniques 259
(a)
One way the 's' term can be eliminated is by ampli-
tude scaling Yt2a or Yt2b in Yt2 in (8). Thus, expanding
node 2 and introducing a CCVS as described in the pro-
l
cedure (CCVS) we get the circuit in figure 4(b) which
has a node admittance matrix given by
l
is given by
Yt3 + Yt2a + Y12b -Yt2a
Y= [ (13)
-Yt2a -Kc Y12a + Y2a
(c) Using the same choice of components as before we
obtain the characteristic equation
Fig. 4. (a) A passive circuit with two nodes plus a ground node;
(b) expansion of node 2 using procedurt (CCVS); (c) expansion of s 2c1c2 + s(Ct92 + c2(91 + 92 - Kc)) + 9192 = 0
node 2 using procedurt (VCCS).
(14)
the kth row and the kth column, and by adding a con-
stant ~Kc (coefficient of the dependent source) to the With Kc = K92 and c2(9t + 92) + CtY2 = c292K,
admittances in the kth row and the pth and qth columns, we can eliminate the 's' term. All oscillator circuits,
respectively. If the controlling variable is current, re- obtainable with the parent passive circuit shown in fig-
place Kc with K:iYpq or K:iYp(n+l)· ure 4(a), whether voltage- or current-mode, can be ob-
The proof for Theorem 4 is similar to that for The- tained by a systematic application of these theorems.
orem 3. As an example of the use of node expansion Note that the CCVS in figure 4(b) can be replaced by
using dependent sources in the design of sinusoidal os- its nullor model to get the circuit in figure 5. Here the
cillators consider the circuit shown in figure 4(a). Its conductance G can be adjusted to set the coefficient of
l
node admittance matrix is given by the dependent source.
83
260 Desai and Aronhime
Yt2b
r-----------------------------------~w
-,
I
I
I
I
Y23 I
I
I
I
y-----------------}zj
Fig. 5. CCVS in figure 4(b) replaced with a two-nullor representation.
Fig. 6. Network of figure 3(b) with 1/14 set to 0, a current excitation applied, and a load admittance attached.
in figure 3(b) and the node admittance matrix in (6). If The circuit in figure 6 is the basis for the circuit pre-
Yt4 is set to zero, node l is excited by a current source, sented by Liu, Tsao and Wu [II]. With suitable choices
and the output is taken through a load admittance YL• of types and values of components, they obtain a cas-
the circuit in figure 6 is obtained which has the transfer cadable filter section which can realize all five second
function: order filter types.
Other filter circuits can be obtained from figure 3(b).
lout Y2a(Ya4Yl2 - Y24Yta)
(15)
If instead of setting Yt4 = 0, we make Y24 = 0 and
I,n = IYI Yt2 = oo, apply. the excitation to node l, and con-
nect the load YL as in figure 4, with a suitable choice
where of components, we can obtain lowpass, highpass and
bandpass circuits. These circuits were presented by
(16) Higashimura et al. [ 12].
84
Current-Mode Synthesis Using Node Expansion Techniques 261
(a)
A'
t Klo
(b)
Fig. 7. (a) A second order biquadratic network N excited through a unity gain current buffer; (b) node 'A' expanded.
As an example of the use of Theorem 2, assume Node expansion techniques using dependent current
that figure 3(a) is expanded at node 3 with the norator sources can be used to generate several oscillator cir-
connected between nodes 3' and 3" and the nullator cuits. Figure 7(a) shows a second-order passive net-·
between nodes 3' and 2. Also assume admittance y 34 work N with an output current / 0 and an input current
is connected between 31 and ground. Applying an ex- excitation It. The transfer function 10 //1 of the pas-
sive network N can be expressed as
citation at node 1 or node 2 and connecting the load
YL between 311 and ground, we obtain, with the choice
of suitable types and values of the components, a uni- (17)
versal filter. These circuits have been developed and
presented by Liu, Kuo and Tsay [13]. In the above
examples, the null or can be replaced by a CCII-. Node A is expanded using procedure (CCCS) and The-
orem 4, yielding the network shown in figure 7(b). The
Many existing circuits can be generated using this
resulting circuit is a hinged network, but the portion
procedure. Higher order filter sections can be devel- of the network containing the excitation in figure 7(b)
oped starting from a completely connected passive net- is removed since by adjusting the gain of the depen-
work. dent source the circuit will oscillate. The oscillation
85
262 Desai and Aronhime
condition and the frequency of oscillation are given by 5. Stevenson, J.. "Use of reciprocity and duality to generate equiv-
alent active-RC networks," in Proc. IEEE Int. Symp. Circuit<
and Sysums, Kyoto, Japan, pp. 821-822, May 1985.
(18) 6. Guo-hua, W., Watanabe, K. and Fukui. Y., "An extended dual
transformation approach to current-mode synthesis," in Pmc.
IEEE Int. Symp. Circuits and Systems, New Orleans, LA, pp.
2 boat- btao 2294-2295, May 1990.
(19) 7. Guo-hua, W., Fukui, Y. Kubolo, K. and Watanabe, K.,
Wo = b2a1 - b1a2
"Voltage-mode to currenl-mode transformation using the ex-
tended dual transformation," in Proc.IEEE Int. Symp. Circuits
where K is the gain of the dependent source. In devel- and Systems, Singapore, pp. 1833-1834, June 1991 .
oping different oscillators the only conditions forced 8. Davies, A., "Matrix analysis of networks containing nullators
on the network N are a 1 :I 0, and b2a1 :I b1a 1 from and norators," Electronics utters, Vol . 2, pp. 48-50, Feb. 1966.
(18) and (19). 9. Davies, A., "The significance of nullators, norators and nullors
In active network theory," The Radio and Electronic Engineer,
pp. 259-267, Nov. 1967.
10. Bruton, L., RC-Activt Circuits Theory and De.<ign. Englewood
4. Conclusions Cliffs: Prentice-Hall, 1980.
II. Liu, S., Tsao, H. and Wu, J., "Cascadable current-mode single
Procedures, four theorems and a corollary have been CCII-biquads," Electronics Leturs, Vol. 26, pp. 2005-2007,
presented which enable a designer to directly modify Nov. 1990.
the entries of the NAM of a given network. Theorems I 12. Higashimaura. M., Fukui, Y. and Ishida, M., "Synthesis of
current-mode transfer functions using a single currenl con-
and 2 and the corollary show how the NAM is modi- veyor," Microelectronics Journal, Vol. 22, pp. 35-46, Nov.
fied while keeping the dimensions constant by properly 1991.
inserting a nullor into the given network. 13. Liu, S., Kuo, J. and Tsay, J., New CCII-based current-mode
In Theorems 3 and 4, dependent sources are em- biquadratic filters.'' Int. J. Electronit's, Vol. 72, pp. 243-252,
1992.
ployed to modify the NAM.
Several examples of the applications of these the-
orems are provided in which universal current-mode
filters which have been previously published in the lit-
erature are derived. In addition, a current-mode oscil-
lator circuit is obtained by applying Theorem 4.
Node expansion is shown to provide a systematic
approach to introducing nullors and dependent sources
into a circuit. In general, node expansion can be
used repetitively, and the resulting nullors or dependent
sources can be replaced by a variety of active devices.
86
Current-Mode Synthesis Using Node Expansion Techniques 263
87
Analog Integrated Circuits and Signal Procetl!!ing, 6, 265-274 (1994)
C1 1994 Kluwer Academic Publlshel'll, B011ton.
RICHARD J, TRIHY
CatkiiCt Dulgn Systenu, 555 Rlwr Oaks l'arlwa)l Building 3, MS 3D I, San Jose, CA 95134
RONALD A. ROHRER
SRC-CMU Center for Compuur-Aided Dtngn. Dept. of Electrical and Compour Engineering, Camegle Mellon Unlwrllty. P/Jtrburglo, PA 15213-3890
Abstract. Asymptotic Waveform Evaluation (AWE) [1, 2] is an efficient and general technique for simulating
linear(ized) circuits. This paper discusses strategies for macromodeling nonlinear circuits with AWE. One approach,
multi-region AWE macromodels, represents an extension of piecewise linear models, with the addition of internal
states. Each region represents an AWE approximation to a linearization (at some bias point) of the nonlinear circuit
of interest. In addition a technique is presented for initializing the internal states when the model transitions from
one linearization to another during a transient simulation. The second approach is to treat nonlinearity as a second
order effect that is superimposed on a linear solution as a post-processing step. A relaxation algorithm that exploits
the reuseable AWE solution is employed to modify the linear solution so that it accounts for the macromodel
nonlinearity.
89
266 Trihy and Rohrer
(1)
AWE is an efficient and general technique for Analogous to PWL models, a nonlinear AWE macro-
frequency- and time-domain simulation and modeling model needs to incorporate information from multiple
oflarge linear(ized) circuits [1, 2). The AWE approach linearizations. These multi-region AWE macromod-
is to approximate the response of the circuit with a low- els consist of a number of dominant pole approxima-
order dominant pole model that captures the behavior tions, corresponding to a number of different linearized
of the circuit. The advantage of AWE over classical equivalent circuits at the different bias points of the
simulation strategies is that the evaluation of this dom- nonlinear element of interest [6).
inant pole model can be, depending on circuit size, Hence, to construct a multi-region AWE macro-
significantly faster than both classical transient- and model, the component of interest is biased at a number
frequency-domain simulations. Furthermore the re- of different operating points. At each of these oper-
sultant AWE model forms a macromodel that can be ating points the circuit is linearized and a dominant
reused repeatedly, at incremental cost. pole AWE macromodel is evaluated. Each of these
AWE employs a form of Pade approximation [8) to linearizations represents a new segment in the model.
evaluate the reduced order model. This approximation Consider the CMOS inverter characteristic curve in
is obtained by matching characteristics of the original figure 2. If the corresponding circuit is biased at three
circuit (namely moments) to a reduced order model. operating points, and linearized at these points, then an
Hence the steps required in obtaining an AWE approx- AWE dominant pole macromodel can be constructed
imation are first the evaluation of the circuit moments, for each case. For the example of figure 2, the de gains
and then Pade approximation. of these three macromodels are superimposed on its
To obtain a reduced, qth, order approximation to nonlinear characteristic. Clearly the de characteristic
an nth order system, 2q circuit moments are required. of the multiregion AWE macromodel is similar to what
These 2q circuit moments are obtained with a single would be expected from a PWL macromodel. How-
LU factorization of the circuit matrix, followed by 2q ever, it differs from it by including internal states, the
forward and backward substitutions. Assuming that effect of which can only be seen in a transient simula-
the cost of the LU factorization is dominant, this is tion and not in the de plot of figure 2.
approximately equal to the cost of a single timepoint Similar to memoryless PWL models, the AWE
evaluation in SPICE. Once the 2q moments have been macromodels must transition from model segment
obtained, Pade approximation converts them to a qth to model segment as the transient analysis proceeds.
90
Nonlinear Macromodeling with AWE 267
Vout
Yin
However, because the AWE macromodels have mem- response in the ith region is known and has two com-
ory in the form of internal state variables, these internal ponents, an initial condition response and a forced re-
states must be initialized at each transition. (This prob- sponse,
lem does not arise for the PWL models since, in this
case, there is a one-to-one correspondence betv.-een the
state variables in the different equivalent circuits.) The
y;(t) = t a
kaeP•t +( t
a
kaep.t) * V.n(t) (2)
91
268 Trihy and Rohrer
==-
u.;.1::····•oo•
p.,. ~ .11
I I I
I I
\
Jj \I
92
Nonlinear Macromodeling with AWE 269
Now the task of solving the nonlinear circuit (5) has 4.1. Load Line Interpretation
been transformed to the task of solving for the wave-
form J(t) under the constraints imposed by (6) and An alternative view of this approach for handling non-
(7). The linear circuit described by (7) can be solved linearity can be seen by considering the graphical solu-
directly using AWE where the expressions for the state tion of a de circuit. Consider the simple diode circuit of
variables in the circuit (1) are augmented with addi-
tional impulse response terms with respect to I (t). Ini-
tially the independent source is assumed to have a zero
contribution to the circuit. The final waveform takes £
the form shown in figure 5. This augmenting current
Zero error
/ I
t
Fig. 5. Waveform of augmenting source. Fig. 6. Spline.
93
270 Trihy and Rohrer
figure 7 and the graphical representation of the problem H(x) = Mx + b + F(x) = 0 (9)
in figure 8.
In a circuit simulator such as SPICE, Newton- where M is a matrix contributed to by linear elements,
Raphson iterations are employed to converge to the b is a vector of independent sources, F(x) is a matrix
de solution. This involves repeated relinearizations of contributed to by nonlinear elements, and x is a vector
the nonlinearity. An initial guess is assumed for the of solution variables. Employing a single linearization
answer and the nonlinearity is linearized at this point. to solve this equation (9) means approximating F(x)
This corresponds to a linearized resistance of Reqi in as
figure 8. The resultant linearized circuit is solved for
a new solution estimate and the diode is relinearized (10)
at this new solution estimate (Req2)· This process is
continued until convergence. where K is a matrix of constants; and approximating
However, the new strategy allows the use of a single H(x) as
linearization of the nonlinear device, thereby avoiding
the repeated linearizations and circuit changes. The (11)
solution process for the circuit of figure 7 using this
approach is shown in figure 9. The problem with this where A = M + K, a matrix of constants representing
approach is that convergence is slow compared with the linearized circuit. Hence the iteration process may
the SPICE approach. However, as described above, be written
Newton-Raphson iterations can be employed to speed
x; - A - i H(x;)
up the convergence without performing any relineariza- (12)
tions. Furthermore, since AWE solves the linear circuit x 1 - A- 1 [Mx; + b + F(x;)]
For convergence of ( 12) it is required that
94
Nonlinear Macromodeling with AWE 271
v
4.4. Examples
95
272 Trihy and Rohrer
" "
2.00
l.RO
f1 71
uo
1.40
1.20
1.00
0.80
0.60
0.40
0.20
0.00
-0.20
-0.40
-0.60
-1).80
-1.00
-1.20
-1.40
1.60 Linear
,..r ,..r ,..... s"iewiiii-··············
1.40
1.20
..
1.00
r-
r··
,.:
i If
!
0.80
i l i
i i I!
0.60 ~ J: r
A- J-
0.40 ~ I i
i i
I
0.20
I l
0.00
x x to·6
0.00 20.00 40.00
Fig. 12. Slew rate limiting in switched capacitor integrator.
96
Nonlinear Macromodeling with AWE 273
one with and one without opamp slewing. The simu- 3. Raghavan, V., An Integrated Methodology for E/utronlc Sys·
tem Interconnect AMiysis, Ph.D. dissenation, Carnegie Mellon
lation results are shown in figure 12. University, Pittsburgh, PA, 1992.
4. Griffith, R. and Nakhla, M., "Mixed frequency/time domain
analysis of nonlinear ciruits,"lEEE Transactions on Computer-
5. Conclusion Aided Design, pp. 1032-1043, August 1992.
5. Dllcmen, C. T., Alaybcyi, M. M., Topcu, S., Atalar, A., Sezer,
Strategies for handling nonlinearity with Asymptotic E., Tan, M. A. and Rohrer, R. A., "Piecewise linear symptotic
waveform evaluation for transient simulation of electronic cir·
Waveform Evaluation have been presented. Piecewise cuils;• Proc. of the IEEE Int. Symp. on Circuits and Systems,
linear (PWL) models are general and flexible and allow pp. 854-857, June 1991.
the simulation accuracy to be adjusted by changing the 6. Trihy, R. J. and Rohrer, R. A., "AWE macromodels for non-
number of segments in the PWL models. However, ev- linear circuils," Midwest Symposium on Circuits and Systems,
1993.
ery time a model changes segment a new linear circuit 7. Trihy, R. J. and Rohrer, R. A., "Simulating sigma-delta modu·
results. The resulting large number of linear circuits lators in AWEswit," Proc. of the IEEE Int. Conf on Computer
compromises the simulator's efficiency. Aided Design, pp. i41 - 144, November, 1993.
The multi-region AWE macromodels allow the cir- 8. Baker, G. A., Jr., Essentials ofPadl Approximations. Academic
Press, 1975.
cuit components to be modeled at a higher level of 9. Meta-Software, HSP/CE User's Manual. 50 Curtner Avenue,
abstraction. The incorporation of internal states per- Suite 16, Campbell, CA 95008.
mits greater flexibility in the modeling process. Sim- 10. Ortega. J. M. and Rheinboldt, W. C., Iterative Solution ofNon·
ilar to the PWL models, additional macromodel seg- linear Equations in Several Variables . Academic Press, 1970.
ments increase the number of linear circuits that need
to be solved; however, the greater model abstraction
allowed reduces this overhead. Furthermore both of
these techniques can have accuracy problems at steady
state. PWL models tend to give accurate results in the
transient portion of the waveform, but the PWL model's
IV characteristic does not match the true IV character-
istic exactly. This can manifest itself in a steady state
error.
The steady state error problem is overcome by using Richard J, Trlhy was born in Cork, Ireland, In 1966. He
the relaxation approach, since the true (SPICE) non- received the B.E. and M.Eng.Sc. degrees In electrical engi-
linear model can be employed. Furthermore the relax- neering from the National University of Ireland at University
ation technique fits elegantly with AWE and does not College Cork (UCC) in 1988 and 1990, respectively, and the
require any additional linear circuits. It does however Ph.D. degree from Carnegie Mellon University, Pittburgh, in
have the limitation that it can have convergence prob- 1993. From 1988 to 1990 he was at the National Micro-
lems if the nonlinearity is too severe. Nevertheless, it electronics Research Centre, UCC. He is currently a Senior
has been found to work well for the nonlinearity associ- Member of the Technical Staff at Cadence Design Systems,
San Jose, California.
ated with the mosfet switches in the switched capacitor
circuits. This suggests that employing the relaxation
approach and the multi-region AWE macromodels to-
gether would result in better accuracy at steady state
and improve convergence performance.
References
I. Huang. X.. Raghavan, V. and Rohrer, R. A., "AWEism: A pro-
gram for the efficient analysis of linear(ized) circuits," Proc. uf
the IEEE Int. Conf. on Computer Aided Design, pp. 534-537.
November 1990.
2. Pillage, L. T. and Rohrer, R. A., "Asymptotic waveform eval- Ronald A. Rohrer received the B.S.E.E. degree from MIT,
uation for timing analysis:' IEEE Transaclions on Compuler· Cambridge, MA, in 1960, and the M.S.E.E. and Ph.D. de-
Aided Design, Vol. 9. pp. 352-366, Aprtl1990. grees from the University of California, Berkeley, in 1961 and
97
274 Trihy and Rohrer
1963, respectively. His 1960s pioneering work in electronic Carnegie Mellon University, Pittsburgh, PA. A Fellow of the
design automation formed the basis for several advances in IEEE since 1980, Dr. Rohrer was President of the Circuits
circuit simulation, sensitivity, and optimization. He is the and Systems Society in 1987, and received the CAS Society
founding editor ofthe/EEE Transactions on Computer-Aided Award in 1990. Among other honors, in 1989 he was elected
Design and the author and coauthor of three textbooks and to the US National Academy of Engineering, and awarded
numerous technical papers. He is currently the Wilkoff Uni- the IEEE Education Medal in 1993.
versity Professor of Electrical and Computer Engineering at
98
ANALOG INTEGRATED CIRCUITS AND SIGNAL PROCESSING
An International Journal
Special Issue on
With the newer generation of integrated circuits incorporating larger analog portions, verification of mi
analog-digital systems is essential to ensure the performance of these systems at critical specifications. '
digital domain being fairly well developed, most of the bottlenecks are posed by the analog sections and
analog-digital interface. Thus, modeling the analog sections and the mixed-mode interface at higher-Ie·
provides an efficient means for analysis of such systems. A key element in this development is behavi
modeling and simulation applied to various stages of the design process, from synthesis to testability
yield estimation.
The goals of this special issue are to report on recent developments and advances in the area of mode
and simulation of analog and mixed analog-digital circuits and systems, and provide a forum for addres:
critical issues and highlighting potential new problem areas.
Topics of interest for this special issue include, but are not limited to:
To be considered for this Special Issue of Analog Integrated Circuits and Signal Processing, prospec
authors should submit six copies of their complete manuscript by March 1, 1995 to one of the two G
Editors:
Instructions for Authors are regularly published in the Journal and may also be obtained from the Guest Editors or by contac
Karen S. Cullen, Kluwer Academic Publishers, 101 Philip Drive, Assinippi Park, Norwell, MA 02061, or by Tel: (617) 871-6
Fax: (617) 871-6528 or Email: <Karen@world.std.com>. Analog Integrated Circuits and Signal Processing is an archival,
reviewed journal publishing research papers on design and applications of analog integrated circuits and signal processing cw
and systems.
Analog Interfaces
for Digital Signal
Processing
Systems
by Frank Op 't Eynde, Willy Sansen
New
THE KLUWER INTERNATIONAL SERIES IN ENGINEERING
AND COMPUTER SCIENCE 225
Publication
The ever-increasing complexity and speed of digital
circuits has considerably modified the architecture of
integrated signal processing systems, resulting in the
analog parts of the system being pushed towards the
boundaries of the signal processing chain. The specifi-
cation requirements of these analog interface circuits
are becoming very strict, in order to fully benefit from
the speed performance and the high dynamic range
offered by digital circuits.
Analog Interfaces for Digital Signal Processing Sys-
tems analyzes the analog interfaces of a digital signal
processing chain, and presents techniques to obtain
maximum performance for various technologies and
applications. The book serves as a general introduc-
tion and as a reference work in the fields of low-distor-
tion analog circuits and oversampled data converters.
It can also be used as the text for advanced courses
covering these topics.
Contents
Foreword. Preface. 1. The Power Consumption of
KLUWER
CMOS Wideband Amplifiers. 2. Low-Distortion CMOS
ACADEMIC
Amplifier Design. 3. Oversampled A-to-D and D-to-A PUBLISHERS
Converters. 4. Higher-Order Sigma-Delta A-to-D Con-
verters. 5. The Practical Implementation of Sigma-
Delta D-to-A Converters. Index.