Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. ED-30,NO.

9, SEPTEMBER 1983 1207

Univ. of California, Berkeley, 1983. 180-183, Oct. 1982; also IEEE Trans.Circuits Syst., to be pub-
[52] A. R. Newton, “The simulation of large scale integrated circuits,” lished.
Ph.D. dissertation, Univ. of California, Berkeley, July 1978; also [58] M.Guariniand 0. A. Palusinski “Integration of partitioned dy-
Univ. of California, Berkeley, MemoUCB/ERL M78/52, July namic systems using waveform relaxation and modified functional
1978. linearization,” 1983 Summer Computer Simuhtion Proc., July 11-13,
[53] J. Kleckner and A. R. Newton, “Advanced techniques for iterated 1983.
timing analysis,” in preparation. [59] T. Y. Feng, “A survey of interconnection networks,” Comput., no.
[54] D. Senderowicz, “An NMOS integrated vector-locked loop,” Univ. 11, pp. 12-27,1981.
of California, Berkeley, Memo. No. UCB/ERL M82/83, Nov. 12, [60] D. D. Gajski, D. J. Kuck, and D. A. Padua,“Dependence driven
1982. computation,” Proc. IEEE SpringCompcon, pp. 168-172, Feb.
[55] J. White and A. Sangiovanni-Vincentelli, “RELAX2: A new wave- 1981.
form relaxation approach for the analysis of LSI MOS circuits,” in [61] D. A. Padua, D. J. Kuck, and D. H. Lawrie, “High-speed mul-
Proc 1983 Int. Symp. Circuits Syst., May 1983. tiprocessors and compilation techniques,” IEEE Trans.Comput.,
[56] E. Lelarasmee and A. Sangiovanni-Vincentelli, “Some new results vol. C-29, no. 9., pp. 763-776, Sept. 1980.
on waveform relaxation algorithms for the simulation of integrated [62] D. H. Lawrie, “Access and alignment of data in an array processor,”
circuits,” in Proc.1982 IEEEInt. Large-ScaleSyst.Symp., pp. IEEE Truns. Comput., vol. C-24, no. 12, pp. 1145-1155, Dec. 1975.
371-376, Oct. 1982. [63] J. T. Deutsch and A. R. Newton, “Data-flow lbased behavioral-level
[57] J. Kaye and A. Sangiovanni-Vincentelli, “Solution of piecewise simulation and synthesis,” Proc. IEEE ICG4D Con!., Sept. 1983.
linear ordinary differential equations using waveform relaxation and [64] J. T. Deutsch, “Multiprocessorcomputer architecture,” Worcester
Laplace transforms,” in Proc. 1982 Int. Conf. Circuits Comput., pp. Polytechnic Institute, MQP Rep. June 1980.

CAzM: A Circuit Analyzer with


Macromodeling
WILLIAM M. COUGHRAN, JR., ERIC GROSSE, AND DONALD J. ROSE

Abstract -A prototype circuit simulator, based on a revised formulation perform the specified logical function. Circuit analysis is
of the circuit equations, is described. The formulation is basedon func-
often subdivided into three phases called circuit, timing,
tional models of active elements, which can be derived from either equiva-
lent-circuit or black-box models. The implementation is oriented towards and logic simulations; a number of simulation programs
the simulation of digital MOS circuits and takes advantage of the special based on various formulations and algorithms are in use
form of suchcircuits. Some oftheimplementation issues are discussed, ~91.
particularly the hierarchical structuringof the circuit equationsand the use The logic phase simulates the binary behavior of the
of splines to facilitate macromodeling; continuation methods for computing
circuit at the transistor or gate level. Timing simulation
transfercurvesare also discussed.Finally,theresults of computational
experiments are presented. attemptsto analyze the various de1a:ys in the circuit’s
critical paths without solving the differential equations to
I. INTRODUCTION great accuracy. Circuit simulation solves the differential
system accurately enough to predict detailed electrical
D IGITAL integrated circuits may be viewed as a col-
lection of transistor devices interconnected topologi-
cally on asubstrate silicon chip. Although the circuit is
properties.
Circuit analysis provides several interesting numerical
designed to perform two-valued logical functions, the de- problems. This paper tersely describes some of the numeri-
tailed electrical behavior of the individual devices, and cal procedures incorporated in CAzM, a prototype circuit
therefore the circuit itself, is analog. That is, the basic simulator under development. Section I1 focuses primarily
physics describing the devices and the circuit behavior is on the formulation of the circuit equations with special
modeled by differential equations. emphasis on MOS transistor circuits. Since the transient
The task of circuit analysis is to solve these differential aspects of the simulation reduce numerically to solving a
equations accurately enough to predict the circuit’s dy- sequence of static operating-point prob’lems, the formula-
namic analog behavior [3], [13]. This insures that the de- tion of the static case is given. We briefly discuss in Section
vices and interconnections will function as required to I11 a damped-Newton method that has proven effective in
simulation. Continuation methods for computing static
transfer curves are explained in Section IV. Section V
Manuscript received January 4, 1983; revised March 21, 1983.
The authors are with the ComputingMathematics Research Depart- contains a briefoverview of variation-diminishing tensor
ment, Bell Laboratories, Murray Hill, NJ 07974. splines, which we use for functional models of transistors
0018-9383/83/0900-1207$01.00 01983 IEEE

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on February 22,2020 at 15:22:23 UTC from IEEE Xplore. Restrictions apply.
1208 IEEE TRANSACTIONS O N ELECTRON DEVICES, NO.
VOL. ED-30, 9, SEPTEMBER 1983

and gates. The models for MOS transistors and a simple


NAND gate are described in Section VI. The results of
computational experiments are presented in Section F'II.
Finally, in Section VIII, some conclusions are drawn.
11. FORMULATION
Circuits are governed by Kirchhoff's current and voltage
laws as well as constitutive relations. The current law is SUECIRCUIT
Ai = 0 Fig. 1. Hierarchical view of circuit.
where A E R n X mis the circuit's reduced-incidence matrix
[4]and i E Rmis the vector of branch currents. The volt.age derlying circuit [14], [9]; we propose assembling the equa-
law implies that tions in a hierarchical fashion.
A circuit 'may be viewed as a collection of I subcircuits,
u = ATu
some of which may consist of individual devices such as
where u E R" and u E R" are the node and branch vcdt- transistors, whose terminals are connected together at nodes
ages, respectively. The constitutive relations are of the form (see Fig. 1); let u be the vector of voltages at the intercon-
K ( i , u ) = 0 E R" nection nodes. Each subcircuit Sj can beviewed as a
which becomes i = f( u ) for voltage-controlled devices. multiterminal element with its terminals attached to nodes;
If these equations are linearized around some pent each terminal has an associated voltage, which is inherited
( uo, uo, io), the linearized constitutive equations are from the node the terminal is connected to, and current.
Let ij be the vector of 3 ' s terminal currents. Let iT=
KiSi+K,6u=s (ir, * - ,iT) be the blockvector of currents. Then Kirch-
hoff's current law at the interconnections is still Ai = 0
where Kiand K , are m X m matrices given by Ki = dK/ ai
and K , = d K / d u , respectively and s is a source vector. The where A is the appropriate incidence matrix. Moreover,
equations can be summarized in matrix notation as v = ATu where uT = (UT,. -,UT)
. is the block vector of volt-
agedifferences. Finally, there is a constitutive relation
K ( i , u ) = (K,(z,, u,); * . , K / ( i / ,u , ) ) ~ =0. Matrix assembly
can be done with the usual two-terminal elements, devices,
or subcircuits; transient problems can be incorporated by
viewing K as a functional.
Mo is the tableau matrix popularized by Hachtel, Braytcln, Let us consider the special case of an MOS transistor in
and Gustavson [8]. Note that M,, has a trivialblock 1,U some detail. Transistors canbe dealt with individually
factorization, corresponding to the substitution u = ATu, as when we perform Gaussian elimination. We will consider a
block-elimination strategy that treats a particular transistor
T, in the circuit.
Suppose the terminals of transistor T, are attached to
four nodes n,, . * n4, not necessarily distinct. One terminal
a ,

Let is distinguished as the reference terminal; the selection of


the terminal is arbitrary so assume the terminal connected

"=[A
K~ K , A ~
0
We call M , the reduced-tableau matrix.
I. to n , is the reference. Note that the reference terminal need
not be grounded and the reference terminal of one transis-
tor need not be connected to the reference terminal of
another attached transistor. There is a current and voltage
In the tableau approach, the linear system involving .M0
associated with each terminal. However, Kirchhoff's cur-
issolved bya general sparse solver that recognizestile
rent law insures that Cij = 0 so only three currents need
special ( - 1,0,1) entries in A but otherwise takes litlle
enter into the constitutive relation for T,. Similar consider-
account of the nature of the circuit. The "elimination
ations dictate the dependence of the constitutive relation
order" is determined by heuristics that attempt to optimize
upon the terminal voltages; u2 - u,, u3 - u 2 , and u4 - u1
the arithmetic and memory costs of the equation solution;
are the relevantdifferences of node voltages. Thus the
these heuristics may not preserve the "locality" of the
general constitutive relation for T, takes the form
circuit equations.
Most formulations of the network equations can be f(i,, i,, i,, u2 - u l , u g- u,, u4 - u,) = 0 .
derived by starting with the tableau matrix, or the reduccd The part of the reduced incidence matrix associated with
form, and performing various block LU factorizations [dl],
[5]. In particular, the nodal and modified-nodal[ l l ] forrnl-
lations can be obtainedin this way. Moreover, the formula-
tion can be extended to allow state variables. The circuit
equations have hierarchical character, reflecting the u r t -

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on February 22,2020 at 15:22:23 UTC from IEEE Xplore. Restrictions apply.
COUGHRAN et ai.: CAZM: A CIRCUIT ANALYZER WITH MACROMODELING 1209

where the three columnscorrespond to the dependency VDD

edges introduced by the voltage differences.


We can write
P

+INTERNAL NODE n

where Kil and K,, correspond to the constitutive relation


for Tl while K i R and K , , represent the remainder of the
Fig. 2. NAND circuit and logic symbol.
circuit. Let us partition A , i , and u as A = [ A l A , ] ,
ir= (iF, z i
) ,u T =
and (UT,
u ; ) where i, E R 3 and u1 E R4.
We then can write the reduced-tableau matrix as drain-to-source current so there are three internal currents
- ,s3
11, . thatcorrespond to the groundedenhancement,
the other enhancement, and the depletion transistors, re-
spectively. There is also the current i o associated with the
output terminal; since we are considering the static case,
we neglect the zero currents associated with the gates of the
two enhancement transistors. Then Kirchhoff's current law
for the NAND subcircuit is

The formal block factorization indicates how the transistor


"assembles" into the lower right block.
We can select other transistors T/ and eliminate them.
After eliminating 7 transistors, we obtain a further reduced
matrix Note that the second equation states that the output cur-
rent is the difference of two transistor currents. The con-
stitutive relations for the NAND are just the constitutive
relations for the individual transistors.
If each subcircuit or deviceisvoltage controlled, the
L j=l
assembly process will terminate with the nodal matrix.
The 1;'s can share nodes so the - Aj KG1KUjAT'~ can Elements that are not voltage controlled are ignored in the
overlap. Thus,the elimination of transistors involves a assembly phaseandanextended set of equations must
matrix-assembly process not unlike finite-element tech- then be solved. Fortunately, many interesting circuits com-
niques [15]. posed of MOS transistors are voltage controlled.
Let us consider the form of the constitutive relation If the internal voltages and currents of a subcircuit S, are
K j ( i J ,u j ) for a subcircuit S,. Let uj be the vector of not of interest, a macromodel (a function) characterizing
termnal voltages associated with q; these terminal volt- the voltage and current behavior at q's terminals can be
ages can be viewed as boundary conditions for Sj. Simi- used; in this way, unnecessary computaticrnof values inter-
larly, let ij be the vector of terminal currents associated nal to a subcircuit may be avoided. A macromodel is an
with Sj. Then S, satisfies local relations approximation to the constitutive relation K j . The voltage
and current behavior at the terminals may be sampled by
independently analyzing q. Section VI discusses the con-
struction and use of a NAND macromod~:l in CAzM.
qj(ij, sj,uj; u,) = o 111. DAMPED-NEWTON'S METHIOD FOR
where A j , qj,uj, and {,are theappropriate incidence OPERATING-POINT ANALYSIS
matrix, the generalized constitutive relations, and the inter- The operating-point problem is simply a system of non-
nal node voltages and currents, respectively. Note that Aj is linear equations
not a part of the incidence matrix governing the terminal
connections of the circuit. The internal node voltages uj g( u ) = 0.
and currents {,may be thought of as states. The form of g is described in the previous section; g is a
The nMOS-NAND circuit shown in Fig. 2 provides an poorly-scaled sparse nonlinear system. Note that u may
example of subcircuit assembly. Let u l , u 2 , and u,, be the represent currents and voltages in this section. We propose
node voltages that correspond to the two input and one using a damped-Newton algorithm [l]for solvingthese
output terminals, respectively. Each of the transistors has a equations.

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on February 22,2020 at 15:22:23 UTC from IEEE Xplore. Restrictions apply.
1210 IEEE TRANSACTIONSON ELECTRON DEVICES, VOL. ED-30, NO. 9, SEPTEMBER 1983

The Newton step for such a problem is found by solvi.xg differential equations
the Jacobian equations
guu’ + A%, =0
JkXk = - gk
where Jk = g,( uk) = ag( u k ) / d u and gk = g( uk).The nC:Xt
l\u’\\$ + \x12=1.
iterate is taken as The differential equations can be viewed as an initial-value
problem if the operating-point problem g ( u , A,) = 0 is
U k + l = uk + tkxk solved first.
where 0 < tk ,<1is chosen to satisfy the sufficient decrease Suppose we know the solution ( u j , A j ) and its tangent
condition (u;, X j ) for some sj and now wish to advance from sj to
This can be done by making the nonlinear equations
g ( u j + lA, j + l ) = 0 and the pseudo arc-length equation

for fixed S > 0, which we usually take to be the machine u~’~(u~+~-u~)+A’~(A~+~-A~)=s~+~-s~~As~


epsilon. If tk =1, the algorithm is the usual Newton’s obtained by a forward-Euler discretization of the arc-length
method. However, the algorithm damps the Newton slep equation, hold usingNewton’s method. This method,
when tk < 1. popularized by Keller [12],can be viewed as a forward-Euler
The strategy for computing t , is analogous toa llne scheme with Newton as a corrector; s is no longer arc-length
search, but attempts to minimize the search effort. In but is a pseudo arc-length. The initial guess is given by the
particular, if the previous damped-Newton step satisfied forward-Euler equation
the sufficient decrease condition with t k , a trial value for
the next step is selected as

The Jacobian for the augmented nonlinear system is


Iterative application of this formula causes the tk’s to
approach one. If astep does not satisfy the sufficient
decrease condition then we try a sequence of decreasing
steps, based on the last successful step tk-1 This formal block factorization implies that
ti’)= tk-1(~IIUkIl/(lXk\l)j2’~*

-
f o r j = 1, .,I until the decrease condition is satisfied; 1’ is a
given integer, which we usually take as nine. This sch.c:me should be solved by the following steps:
decreases the step slowly for small j but then quic.kly 1) Solve g u y = gh and g,z = r ;
reduces it for largerj. 2 ) SA = ( t - U j ‘ Z > / ( X j - ujTy);
This algorithm with this method of picking tk is globdly 3) S u = z - S A y .
convergent if g satisfies certain technical assumptions (:see
[l]for details); we are investigating these assumption,$ in Note that gu may be ill-conditioned for some A)’s.
the circuit context. Since portions of a circuit may be Once ( u ~ +A~j +,l ) is known, we mustcompute
“off”, isolated rows and columns of the Jacobian mag be A)+l) in order to take another step in s. This can be
zero as well as the corresponding components of gk; defla- done using they = g;lgh computed above and forming
tion of these singular systems suffices to compute xk jn a
bounded way. A’= f Ilyll;
u’ = A‘y
Iv. CONTINUATION FOR STATIC TRANSFER where the sign of A’ is chosen in the appropriate way.
ANALYSIS The means of selecting Asj obviously influences the
We are interested in solving the transfer analysis prob- efficiency of the overall procedure. We find that a modifi-
lem given by cation of algorithm I11 in [ l o ] iseasy to implement and
effective; its main advantage is that no quality measure for
g(u,A)=OforXE[A,,A.]. the corrector iterates need be computed. For brevity, we
In other words, we are interested in finding u ( A ) for a will defer a detailed discussion of the steplength algorithm
range of input voltages A, < A G A,. We have found con- and computational results to a later paper.
tinuation techniques attractive for computing transfer With many circuits, it is possible to enforce only
curves. g(ui+l, = 0 at each point and use an initial guess
Parametrize the system by arc-length based on the forward-Euler equation. (Essentially, any
circuit where a particular value of A gives a unique u can be
g ( u ( s > ,A(sN = 0 treated in this way.) This saves an extra solvefor each
and then differentiate to obtaina system of ordinary Newton iteration. This is an option in CAzM.

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on February 22,2020 at 15:22:23 UTC from IEEE Xplore. Restrictions apply.
COUGHRAN et a[.: CAZM: A CIRCUIT ANALYZER WlTH MACROMODELING 1211

Our experience with continuation for transfer analysis


indicates that the input voltage A should be stepped in
tenths of a volt in some parts of an interval to ten-thou-
sandths of a volt when a digital circuit is changing a
latched state, The user cannot preselect easily an adequate
set of sample values for A to properly resolve a transfer
curve. Continuation does this automatically at a low cost.
I

v. SPLINES FOR MODELING


4
E l
._
Various models have been suggested for active circuit
elements, including analytic and table look-up schemes.
We advocate the use of tensor-product splines that pre-
serve mononicity.
We begin with the problem of approximating a uni-
variate function f on the interval [ a , b]. Suppose we are
given a knot sequence C
a = t , = t , = t , < * * . < tn-2 = t n P l= t , = b . 0 I 2 3
V d s ( VOLTS I
Let
Fig. 3. Transistor static currenti = f( Vds)for several Vgsvalues.
PX(d = {At-”’ 2
ift>x
otherwise. VI. MACROMODELS
Then the B-splines of order three (piecewise quadratics) are Thestatic transistor model, neglecting substrate cur-
defined by rents, is simply given by
= (ti+3 - t i > [ t i ,t i + l , ti+2, ti+,lpx is=-id=f(Uds,Ugs,Ub.~)

where [ti,. . .,ti+ ,] indicates the third-order divided dif- ig=0


ference (see [2] for details). The variation-diminishing spline
of Schoenberg is we call this model a macromodel since it is not an equiva-
lent circuit. We use a semi-analytic enhancement-MOS
S ( X ) = C f (ti*)Bi(x) transistor model [16], [17] as the source of data for the
splines, although experimental data or the results from a
where ti* = ( t i + ti+,)/2. Note that computing S does not semiconductor-modeling program [6] could be used. A
require the solution of a linear system but does require square-root transformation of variables isused in u to
function values at specific points. gs
reduce the number of sample points in that variable since
S has several important properties: 1) [If - SI/ = O(At2) there is quadratic behavior in ugs. Figs. 3 and 4 illustrate
where At = max - til; 2) S preserves the monotonicity the spline approximation to the transistor.
o f f ; 3) S E C1; and 4) constructing S is a linear process. We obtain a depletion-transistor model by adjusting the
B-splines of higher order (piecewise cubics or higher) still flat-band voltage ufb of our enhancement model, that is
give a second-order approximation so piecewise quadratics
= f ( U d s , ugs + u f b , u b s ) .
are sufficient. The facts that S preserves monotonicity and Using the quasi-static approximation, the transient tran-
S is easily differentiated make possible the use of damped- sistor model can be written as
Newton methods [l]; moreover, by linearly extending S
outside of [ a , b ] we avoid constraints in the Newton proce-
dure.
Any linear univariate approximation process can be
extended to several variables through the use of tensor
products; the details of the computational method are Each of the components of q can be modeled with a tensor
described in [7]. For the variation-diminishing spline, a spline.
three-dimensional tensor variant with the same knotsin Let us consider the construction of astatic NAND
each variable is given by macromodel. (The usual nMOS-NAND circuit is shown in
Fig. 2.) The basic idea is to supply voltages on the two
s(x,Y , z > = C f ( t r 9 tj*,t : ) B , ( x ) B , ( y ) B , ( z ) input pins andthe usual output pin. Obviously, if the
ijk
supplied output voltage is not the one: that wouldhave
since the approximation operator is essentially the identity. occurred with the output voltage as an unknown then there
The computational costs of a tensor spline in p dimen- will be a residual current associated wit.h the output lead.
sions is O(nini)space if there are n , sample points for the By running CAzM with various sample voltages for the two
ith variable and O ( 3 P ) time for one evaluation. inputs and one output, we can obtain a tensor spline that

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on February 22,2020 at 15:22:23 UTC from IEEE Xplore. Restrictions apply.
1212 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. ED-30, NO. 9, SEPTEMBER 1983

0.2:
u1=5

u2 = 4
i UI=:

-a
-a
.--
E C
E UI’O
.-

-0.2!
I 5
0
OUT ( V O L T S )

Vg,(VOLTSI Fig. 5. NAND macromodelcurrent.

Fig. 4. Transistor static currenti = f ( V,,) for several vds values.


ple, it takes three transistor evaluations to compute the
models the residual current residual current for a simple NOR circuit whereas a specific
spline macromodel would require the same computational
io = f ( u 1 , u 2 , u o ) . effort as a single transistor evaluation. An XOR can be
Fig. 5 illustrates the NAND macromodel. Note that we can computed via a NAND macromodel anda transistor
writef as a function of node voltages because the NAN13 is evaluation or, as before, a specific spline macromodel can
always connected to ground. be constructed at the cost of a four-dimensional fit.
There is a node between the two enhancement tran.r,is-
tors, node n in Fig. 2, that is an important unknown .lor VII. NUMERICAL EXPERIMENTS
determining the output voltage but does not usually con- Wehave run experiments on several circuits, such as
nect to other devices in the circuit. In other words, nodc n combinational circuits of various sizes, memory cells, trig-
is an isolated node whose value is not used outside of the ger elements, and shift registers. The dramatic effectiveness
NAND. Hence, if the NAND is replicated several times in of macromodeling can be best demonstrated with a one-bit
a circuit, the value of the isolated node is computed each full adder composed of 9 NAND elements (see Fig. 6);
time. The macromodel incorporates the output behawlor larger ripple-carry adders can be constructed from the
once and for all, eliminating the need for computing Ihe one-bit adder.
isolated node’s value. By using the macromodel in operat- We performed operating-point analyses with a series of
ing-point computations, it is possible to only solve for h e larger and larger adders. In the first set of experiments, the
unknown output voltage, reducing the number of utn- input node voltages were taken to be zero volts as were the
knowns by a factor of two. carry-out and output node voltages of each one-bit adder
A further advantage of the NAND macromodel is ihe stage; a constant interior guess of V,,/2 was used for the
character of the nonlinear equations that must be solved. remaining nodes. The results are summarized in Table I.
TheNAND expressedwith three transistors can ca.use In the second set of experiments, the input node voltages
difficulties since enhancement transistors have a dou1,le were taken to be zero volts and the other node voltages
zero at ugs = 0. In particular, linear convergence of New- wereguessedaszero and V,, based on a switch-level
ton’s method is observed when both of the enhancemlcnt simulation. The results are summarized in Table 11.
transistors are operating in the subthreshold range. ‘The The stopping criteria resulted in node voltages accurate
computational experiments described later demonstrate t lis to a minimum of five digits. All computations were per-
point. formed in single precision on a Cray-1 computer.
The NAND macromodel is a three-dimensional temmr The initial guess influences the number of damped-
spline so its computational complexity is the same as [he Newton iterations needed; we believe that the results of a
transistor’s. Hence, the macromodel not only reduces fhe switch-level simulator should be used as the initial oper-
number of unknowns in the nonlinear system but it :e- ating-point guess.However, the method convergeseven
places three function evaluations by one. when the initial guessis poor and we observe similar
Static inverter and NOR macromodels can be ccln- behavior when something other than zero is input to the
structed from simple linear combinations of transismr adders.
models. Alternatively, we could construct a spline macro- The first line in each table above shows the work neces-
model representing an inverter or a NOR gate in order to sary tocompute the operating point when the NAND
reduce the number of spline functions evaluated; for exa m- elements in the one-bit adder are represented by one

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on February 22,2020 at 15:22:23 UTC from IEEE Xplore. Restrictions apply.
COUGHRAN et al.: CAZM: A CIRCUIT ANALYZER WITH MACROMODELING 1213

TABLE I1

-+Y I

OUT
Bits

1
1
Work for full adders with switch-simulator guess
Transistors Unknowns Newtons

27
27
18
9
7+15
4 4
Macromodel?

u 4 - 108 36 4 4
216 8 72 4 4
CIN -I 16 432 144 4 4 Yes
Fig. 6. One-bit full adder circuit. 32 288 864 4 4. Yes

TABLE I connected to inputs and outputvoltages aS the previous set


of NAND’s, and so forth. At the end of this process, the
Work for full adders with interior constant guess
Bits 1 Transistors 1 Unknowns I Newtons I Functions 1 Using NAND
adders’ output voltages would be determined. Also, the
1 Macromodel? question of how the dependency graph of the circuit should
1 1 27 I 18 I 35+9 I 71 No be used in general is still being considered.
1 27 9 10 17 Yes
4 108 36 10 17 Yes ACKNOWLEDGMENT
8 216 72 10 17 Yes
16 432 10 144 17 Yes The authors appreciate the advice and efforts of R.
32 864 288 10 17 Yes Bank, W. Fichtner, J. Gannett, J. Hodgins, and G. Taylor.
REFERENCES
depletion and two enhancement transistors. Thirty-five and
R. E. Bank and D. J. Rose, “Global approximate Newton methods,”
seven damped-Newton steps in Tables I and 11, respec- Numerische Mathematik, vol. 37, pp. 279-295, 1981.
tivelygive the correct answer to three digits at each C. de Boor, A Practical Guide to Splines. New York: Springer,
NAND’s output node; yet it takes another nine and fifteen 1978.
R. K. Brayton, G. D. Hachtel, and A. Sangiovanni-Vincentelli, “A
Newton steps in Tables I and 11, respectively to settle the taxonomy of CAD for VLSI,” in Proc. I981 European Conf. On
values at -theinterior nodes of the NAND’s. The other lines Circuit Theory and Design, pp. 34-57, 1981.
in each table show what happens when the NAND is L. 0. Chua and P.-M. Lin, Computer-Aided Analysis of Electronic
Circuits: AlgorithmsandComputational Techniques. Englewood
represented by a spline macromodel. Obviously, much less Cliffs, NJ: Prentice-Hall, 1975.
work is required with the macromodel since fewer Newton W. M. Coughran, Jr. and D. J. Rose, “The nodal-oriented tableau
steps are taken and there are half as many node voltages to approach to circuit analysis,” unpublished manuscript, 1982.
W. Fichtner and D. J. Rose, “On the numerical solution of nonlin-
solve for. ear elliptic PDEs arising from semiconductor device modeling,” in
Elliptic Problem Solvers, M. Schultz, Ed. New York: Academic
VIII. CONCLUSIONS Press; 1981, pp. 277-284.
E. H. Grosse, “Tensor spline approximation,” Linear Algebra and
We have discussed the formulation of the circuit equa- Its Applications, vol. 34, pp. 29-41, 1980.
tions, a damped-Newton method for computing operating G. D. Hachtel, R. K. Brayton, and F. G. Gustavson, “The sparse
points, a continuation method for calculating transfer tableau approachto network analysis and design,” IEEETruns.
Circuit Theory, vol. CT-18, pp. 101-113, 1971.
curves, and the use of variation-diminishing tensor splines G. D. Hachtel and A. L. Sangiovanni-Vincentelli, “A survey of
for modeling devices and gates. We have not described the third-generation simulation techniques,” Prcc. IEEE, vol. 69, pp.
transient-analysis problem although the formulation easily 1264-1280,1981,
C. den Heijer and W. C. Rheinboldt, “On steplength algorithms for
accommodates the time-dependent problem by viewing the a class of continuation methods,” S I A M J . Numerical Analysis, vol.
constitutive relations K ( i , u ) = 0 as functionals instead of 18, pp. 425-948,1981.
functions. CAzM currently performs transient analysis by C.-W. Ho, A. E. Ruehli, and P. A. Brennan, “The modified nodal
approach to network analysis,” I E E E Trans. Circuits Syst., CAS-25,
using the backward-Euler method to generate a series of pp. 504-509, 1975.
operating-point problems. H. B. Keller, “Numerical solution of bifurcation and nonlinear
We have only begun to research the problem of macro- eigenvalue problems,” in Applicutions of Bifurcation Theory, P.
Rabinowitz, Ed. New York: Academic Press, 1977, pp. 359-384.
models involving time dependence. Thestatic transistor A. R. Newton, “Computer-aided design of VLSI circuits,” Proc.
model is easily extended to a transient model by the IEEE, vO~.69, pp. 1189-1199,,1981.
quasi-static approximation. However, it is not yet clear N. B. Rabbat, A. L. Sangiovanni-Vincentelli, and H. Y . Hsieh, “A
multilevel Newton algorithm with macromodeling and latency for
under what circumstances quasi-static time macromodels the analysis of large-scale.nonlinear circuits in the time domain,”
for gates are applicable. IEEE Trans. Circuits Syst., vol. CAS-26, pp, 733-741, 1979.
Numerical logic simulation giving voltage levels instead G. Strang and G. Fix, A n Analysis of the Finite Element Method.
Englewood Cliffs, NJ: Prentice-Hall, 1973, p. 91.
of switchlevels (“on” or “off”) is another possibility. G. W. Taylor, “A unified model for a short-channel MOSFET,” in
Consider theadders described in ’ the previous section. Proc. 39th Ann. Device Reseurch Con!., Santa Barbara, California,
There are no cycles in the dependency graph of the circuit 1981.
G. W. Taylor, W. Fichtner, and J. G. Simmons, “A description of
so it is possible to evaluate the output voltages of NAND’s MOS internodal capacitance for transient simulation,” IEEE Trans.
connected only to inputs, then output voltages of NANDs Computer-Aided Design of ICAS, vol. CAD-I, pp. 150-156, 1982.

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on February 22,2020 at 15:22:23 UTC from IEEE Xplore. Restrictions apply.

You might also like