E
cient computation of quasi-periodic circuit operating

conditions via a mixed frequency/time approach

Dan Feng
, Joel Phillips
, Keith Nabors
, Ken Kundert
, Jacob Whitey

Cadence Design Systems, San Jose, CA 95134
y Massachusetts Institute of Technology, Cambridge, MA 02139

Abstract approaches(MFT)[2, 3] have been proposed. The methods in [2, 3]

exploit the fact that many circuits of engineering interest have a
Design of communications circuits often requires computing strongly nonlinear response to only one input, such as the clock in
steady-state responses to multiple periodic inputs of differing the case of a switched-capacitor circuit, or local oscillator in the case
frequencies. Mixed frequency-time (MFT) approaches are orders of a mixer, but respond only in a weakly nonlinear manner to other
of magnitude more efficient than transient circuit simulation, inputs. Recent extensions[9] of the MFT algorithms can sometimes
and perform better on highly nonlinear problems than traditional treat circuits with strongly nonlinear responses to multiple inputs.
algorithms such as harmonic balance. We present algorithms for The original MFT algorithms suffered from several drawbacks
solving the huge nonlinear equation systems the MFT approach that prevented their application to practical circuits, particularly
generates from practical circuits. circuits of large size. This paper makes four improvements on the
original algorithms that enable the MFT method to be applied to
1 Introduction large circuits. In section 3 we discuss a quasi-periodic boundary
condition that avoids the ill-conditioning due to a poor choice of
One of the most difficult challenges in circuit simulation is the boundary condition in previous algorithms [3, 2]. In Section 4,
analysis of circuits that operate on multiple timescales. Typical we show how matrix-free Krylov-subspace based iterative schemes
examples of this class are switched-capacitor filters and circuits may be extended to the MFT methods to allow solution of very
used in RF communications systems. Applying standard transient large scale problems. A similar approach is taken in [15]. Section 6
analysis to such circuits can imply simulating the detailed response presents a simple continuation scheme that improves convergence
of the circuit over hundreds of thousands of clock cycles (millions of the Newton iteration. In section 5 we show that, while the
of timepoints), a generally impractical approach. Fortunately many unpreconditioned matrices arising from applying shooting methods
circuits of interest are designed to operate near a time-varying, are often well conditioned and thus suitable for direct application
but quasi-periodic, operating point. Some of these circuits can of Krylov-subspace based iterative solvers, the matrices generated
be analyzed by assuming one of the circuit inputs produces a by the MFT method are never well conditioned even for a well-
periodic response that can be directly calculated by steady-state chosen boundary condition. We present a preconditioning strategy
methods[4], thus avoiding long transient simulation times. Any that gives rapid convergence of the iterative solver.
other (time-varying) circuit inputs are treated as small-signal by
linearizing the circuit around the periodic operating point. Efficient 2 Background on the MFT approach
algorithms[12, 7] now exist to find periodic operating points and to
perform periodic time-varying small-signal analysis[13, 8]. Circuit behavior is usually described by a set of nonlinear
However, many circuits cannot be analyzed with the periodic- differential-algebraic equations (DAEs) that can be written as
operating-point-plus-small-signal approach. For example, predict-
d
dt Q(v(t)) + I (v(t)) + u(t) = 0

ing intermodulation distortion of a narrowband circuit, such as
(1)
a mixer+filter, involves calculating the nonlinear response of the
mixer circuit, driven by an LO, to two high-frequency inputs that
are closely spaced in frequency. The steady-state response of such where Q(v(t)) 2 <N is typically the vector of sums of capacitor
a circuit is quasi-periodic. charges at each node, I (v(t)) 2 < N is the vector of sums of
The situation is further complicated by the fact that many multi- resistive currents at each node, u(t) 2 < N is the vector of inputs,
timescale circuits have a response (again mixers and switched- v(t) 2 <N is the vector of node voltages, and N is the number of
capacitor filters are good examples) that is highly nonlinear with circuit nodes.
respect to at least one of the exciting inputs, and so steady-state We are interested in the case in which the input signal u(t)
approaches such as multi-frequency harmonic balance[7] do not is quasiperiodic. We will say a signal is L;quasiperiodic if it
perform well. can be written as a Fourier series with L fundamental frequencies.
To circumvent these difficulties, mixed frequency-time RF circuits are generally influenced by one periodic timing signal,
often referred to as the LO or the clock, and one or more information
signals. If we let fc denote the clock frequency, and f 1 : : : fS the
S information signals, then the (S + 1)-quasiperiodic input can be
written as
 X1 1
X 1
X
u(t) =  U (k1
  
kS
kc ) (2) Waveform

k1 ;1 kS =;1 kc =;1
Envelope
Sampled Envelope

 e;|2k1 f1 t    e;|2kS fS te|2kc fc t

The fundamental assumption of the MFT method is that the
circuit possessesa quasiperiodic steady-state response. That is, v(t)
is an S + 1 quasiperiodic signal with fundamentals f 1
: : :
fS
fc .
Furthermore, since all physical circuits have a finite bandwidth,
v(t) can be well approximated by taking only a finite number of
terms in the Fourier series, so that

X
K
1 KS
X 1
X
v(t) =  V (k1
  
kS
kc )
k1 =;K1 kS =;KS kc =1
 e;|2k1 f1 t    e;|2kS fS t e;|2kc fc t : (3)
where V (k1
  
kS
kc ) 2 C N . An interesting property of the Figure 1: Sample envelope is the waveform traced out when signal
MFT method is that it is not necessary to truncate to a finite number is sampled with the clock period
of harmonics of fc .
Now suppose that v(t) is sampled at a discrete set of points
0
tn = t0 + nTc , where Tc = 1=fc is the clock period, t0 2 0
Tc )
and n runs over the integers, to obtain a discrete signal v̄(t ). Since 0 which may be written more compactly by introducing the multi-
cycle transition function that is the collection of the K transition
functions from tk to tk + Tc , as
K KS
v̄(t0n ) = v̄Tc = ΦTc (v̄0):
X1 X
 V̄ (k1
  
kS ) (8)
k1 =;K1 kS =;KS Now note that for each node n, the vector of signals on that
 e;|2k1 f1 tn    e;|2kS fS tn node, at the sample time plus one clock cycle, v̄ Tn , is a delayed
0 0
(4)
version of the signals at the sample points (this will be made more
clear below). Therefore there exists a linear operator DTc that maps
v̄0n to v̄Tnc
where
1
V (k1
  
kS
kc )e;|2kc fc t
v̄Tnc = DTc v̄0n :
X
V̄ (k1
  
kS ) = 0
(5) (9)
kc =;1 Note that DTc is a real matrix and independent of node n. Hence (9)
0 holds for each n = 1
  
N . It represents a boundary condition
the “envelope” v̄(t ) is S -quasiperiodic and can be represented as on solution to (1).
a Fourier series in only the “information” fundamentals. The clock
Combining (9) and (8) gives
fundamental has disappeared. For an example of such an envelope,
see Figure 1. The continuous waveform is the waveform that has (DTc  IN )v̄0 ; ΦTc (v̄0) = 0
(10)
V̄ as its Fourier coefficients, or, equivalently, obtained by Fourier
interpolation of the sampled points. where  is the Kronecker product (see [5] for definition) and I N
QS
In principle, since there are only K = s=1 (2Ks + 1) Fourier is the N by N identity matrix. Equation 10 is a system of KN
coefficients to represent v̄, then once the value of K distinct points nonlinear equations and KN unknowns v̄ 0 that can be solved for
t1
  
tK along the sample envelope are known, then the full the envelope sample points. From these sample points and the
transition functions the circuit’s quasiperiodic operating point (in
particular, the spectrum of v) can be recovered, which is discussed
envelope can be recovered. The envelope corresponding to the
quasiperiodic operating point is obtained by obtaining K sample
points that lie on the solution to the DAE given by (1). in Section 7.
We define the state transition function (v0
tk
tf ) = v(tf ) :
v(t) satisfies equation (1) for t 2 tk
tf ] and v(tk ) = v0: In 3 Sample point selection
particular, define the vector
To construct the matrix DTc , referred to as the delay matrix,
v̄0 = v̄T (t1 )
  
v̄T (tK )]T = vT (t1 )
  
vT (tK )]T
(6) consider the Fourier series of v̄0 and v̄Tc . Referring to equation
(4), we have
where superscript T denotes matrix transpose, to contain v̄ at the
K sample points tk = t0 + nk Tc
k = 1 : : : K
nk 2 Z . The value K KS
v̄(t0n + Tc ) =
X1 X
of the K points that follow by one cycle can be obtained from the  V̄ (k1
  
kS )
transition function, k1 =;K1 kS =;KS
e ; |2k1 f1 tn
   e;|2kS fS tn ΩT c (k1
  
kS )
v̄TTc = vT (t1 + Tc )
  
vT (tK + Tc )]T
0 0

(7) (11)
T T
= (v (t1)
t1
t1 + Tc )
  
(v (tK )
tK
tK + Tc ) ]
T where
ΩTc (k1
  
kS ) = e ;|2k f Tc    e;|2kS fS Tc
1 1
(12)
 Thus if Γ is the matrix mapping sample points on the envelope Each iteration of GMRES requires a matrix-vector multiplication.
to Fourier coefficients, then the delay matrix may be constructed as For a vector q 2 <KN , the term (DTc  IN )q is calculated by first
applying a K dimensional DFT N times, then scaling each row
DTc = Γ;1ΩTc Γ: (13) with ΩTc , and finally applying a K dimensional inverse DFT N
times.
In particular Γ may be constructed as the Kronecker product of Let q be partitioned into q = q1T
  
qK
T ]T , qk 2 <N , for
one-dimensional 2K s + 1-point Fourier-transform matrices 1  k  K . Then
Γ(mn
s) = e|2mfs tn =(2Ks +1) @1
2 3
(14)
@ v̄0 (t1 ) q1
@ Φ q = 64 .. 7
:
as
@ v̄ @K
. 5 (23)

Γ = Γ(1)      Γ(S) (15) @ v̄0 (tK ) qK

From the properties of Kronecker products, is likewise a Γ ;1 The calculation of each @ v̄@ k qk can be carried out in the same
0 (tk )
Kronecker product of the inverses of the Γ (s) . In [3], no particular matrix-implicit fashion as discussed in [12].
care was taken in the choice of the sample points t k , so that the
Γ(s) ’s were ill-conditioned matrices corresponding to an “almost- 5 Preconditioning
periodic” Fourier transform. Fortunately, a better choice of points
is possible. For many problems, the GMRES method is not efficient for solving
Assume the K sample points can be arranged into an S - (21) without an effective preconditioner. A simple analysis reveals
dimensional array  (k1
  
kS ), ;Ks  ks  Ks, 1  s  S , why. Consider the case where the state transition function of
such that for a given dimension s, there exists an integer p and the circuit, over one clock cycle, is approximately linear, that is
(x
t
t + Tc ) ' Hx(t). Linear circuits are a trivial example of
 (  
ks + 1
  ) ;  (  
ks
  ) = 2KTs+ 1 + pTs (16) a case where this is true, and while nonlinear circuits will have
s nonlinear state-transition functions, if the method performs poorly
hold. In this case the entries of the Γ (s) matrices are
for linear circuits, it surely will not work well for nonlinear circuits
either. However, many nonlinear circuits have a state-transition
Γ(mn
s) = e|2mn=(2Ks +1) (17) function that is nearly linear[4], a fact we will exploit below to
construct an effective preconditioner. The convergence of the
That is, they are the DFT matrices, and the matrix Γ : C 2K1 +1  GMRES method will depend on the location of the eigenvalues of
    C 2KS +1 =) C 2K1 +1      C 2KS +1 represents an S - the Jacobian matrix, DTc ; J . If H is an eigenvalue of the matrix
dimensional DFT. Thus Γ has a condition number of one; it is H , then ei!Tc ; H , where ! = 2 (k1 f1 + k2f2 +    kS fS )
perfectly well-conditioned. will be an eigenvalue of D Tc ; J , for every k1
k2
: : : in the
MFT analysis. Thus unless all the secondary input frequencies
4 Matrix-implicit solution procedure are nearly commensurate with the clock frequency, the eigenvalues
of DTc ; J will be “fanned out” by delay matrix. This will cause
severe convergence problems for the GMRES solver. Roughly
We employ Newton’s method to solve speaking, the GMRES algorithm in the MFT algorithm with K
(DTc  IN )v̄0 ; ΦTc (v̄0) = 0
(18) total harmonics will take K times as many iterations to converge
than the GMRES iteration for the steady-state problem with only
At iteration i, the Jacobian matrix is given by the clock excitation applied. 1
The following lemmas[5] about the properties of Kronecker
DTc  IN ) ; @@Φv̄T0c

i



products are needed to perform the formal analysis.
( (19)
v̄ =v̄ 0
A1
A2
  
Ap 2 F mm , B1
B2
  
Bp 2
F nn then
0 Lemma 5.1 If
Recall from (13) DTc = Γ;1 ΩTc Γ, which is fixed through all
Newton iterations. Let J = @@v̄Φ jv̄ =v̄i be obtained from the
0 0 A A2    Ap )  (B1B2    Bp )
( 1
0

= (A1  B1 )(A2  B2 )    (Ap  Bp ):

multicycle transition function by


Lemma 5.2 If A 2 F mm
B 2 F nn then
2 3
@1

@ v̄0 (t1 )
v̄0 (t1 )i
@ ΦTc


= 666 ..
7
7
: (a) (A  In )(Im  B ) = (Im  B )(A  In ).
@ v̄0 v̄i 4 . 7 (20)
(b) (A  B );1 = A;1  B ;1 .

5
0
@K

@ v̄0 (tK ) v̄0 (t )i


K
Theorem 5.3 If the system in (1) is linear time-invariant then if H
Note that J is block-diagonal. Defining b = ;(D Tc  IN )v̄0i ; is an eigenvalue of @ Φ=@ v̄ 0 , then e|2!k Tc ; H is an eigenvalue
Φ(v̄0i ), we perform the Newton iteration by solving the equation of the MFT Jacobian matrix.

(( DTc  IN ) ; J )∆v̄0i = b 1
This follows because the eigenvalues of H
are typically inside the unit circle of
(21)
K
the complex plane. The delay matrix replicates the eigenvalue structure times, each
shift being a complex number of order unity, and generally causing the convex hull of
using the iterative solver GMRES [11], and setting the eigenvalues of D Tc ; J to enclose the origin.

v̄i+1
0
i
0
i
= v̄ + ∆v̄ : 0 (22)
 Proof. For linear time-invariant circuits, the diagonal blocks of
@Φ @1 @
=    = @ v̄ (tK ) = H . The Jacobian
2

@ v̄0 are the same, i.e., @ v̄0 (t1 ) 0 K Before Preconditioning

matrix is equal to 1.5 After Preconditioning

(Γ
;1ΩTc Γ)  IN ; (IK  H ) (24)
;1  IN )(ΩTc  IN )(Γ  IN ) ; (IK  H ) (25)
1

= (Γ

= (Γ
;1  IN )f(ΩTc  IN ) ; (26)
0.5

(Γ
;1 ; 1 ;
 IN ) (IK  H )(Γ  IN ) g(Γ  IN )
1 0

= (Γ
;1  IN )f(ΩTc  IN ) (27)
;
−0.5

;(Γ  IN )(IK  H )(Γ  IN ) g(Γ  IN )

1

= (Γ
;1  IN )f(ΩTc  IN ) (28)
−1

;(IK  H )(Γ  IN )(Γ  IN );1g(Γ  IN ) −1.5

= (Γ
;1  IN )f(ΩTc  IN ) (29) −2

;(IK  H )g(Γ  IN ):
−1.5 −1 −0.5 0 0.5 1 1.5

Equations (24) to (25) is because of I N = IN IN IN and Lemma Figure 2: Eigenvalue distribution before and after preconditioning
(27) to (28) due to Lemma 5.2 (a). Since (Γ;1  IN ) is unitary and
5.1. Equations (26) to (27) is due to Lemma 5.2 (b), and equations

its inverse is (Γ  IN );1 , the right hand side of equation (29) has the
same spectrum as (Ω Tc  IN ) ; (IK  H ). It is easy to verify that Preconditioned
Raw

(ΩTc  IN ) ; (IK  H ) is block diagonal, hence its eigenvalues

are the union of eigenvalues of all the blocks, e |2!k Tc IN ; H , for
0
10

k = 1
  
K .
The preceding analysis suggests a good way of preconditioning
||r ||/||r ||

−1
10
0

for solving the Newton equation (21). Solving (21) is equivalent to

solving
k

f(ΩTc  IN ) ; ((Γ  IN )J (Γ;1  IN ))gy = (Γ  IN )b
(30) −2

10

where y = Γ∆v̄i . A good choice of preconditioner is P =

(ΩTc  IN ) ; (IK  H ), where H can be chosen as the Jacobian
matrix from the steady-state analysis in the initial guess stage
discussed in Section 6, or any of the diagonal blocks, @v@ (ti) , for 10
−3

0 50 100 150 200 250 300 350 400 450 500 550
@
i = 1
  
K , of @ v̄0 . In particular, if the single-cycle state-
Φ

transition function is linear and time invariant, then the Newton

equation can be solved in a single GMRES iteration. Note that Figure 3: Convergence of GMRES with (dashed line) and without
the preconditioner presented here will be effective if the Jacobian (solid line) preconditioner for one MFT Newton iteration.
of the state-transition function is nearly constant over multiple
cycles. The circuit behavior inside each clock cycle is hidden
from the preconditioner. This is not the case in, for example, the The eigenvalues are very tightly clustered around unity, indicating
time- or frequency-averaged prenditioners typically used in modern excellent performance of the preconditioner and very rapid GMRES
harmonic balance[7, 6] codes. For this reason the preconditioner convergence.
presented here may perform well under much weaker assumptions Figure 3 shows the effectiveness of the preconditioner in
about the circuit behavior, in particular at higher power levels. reducing the number of GMRES iterations needed to solve each
For each GMRES iteration, a system Pu = v has to be MFT Newton update equation, for the same RF circuit mentioned
solved. Since P is block diagonal, we need to solve a sequence
of K systems (e|2!k Tc IN ; H )uk = vk , for k = 1
  
K ,
;
above. Only three iterations were needed to reduce the residual
by a factor of 10 2, whereas without the preconditioner, over 400
where uT = u1T
  
uTK ]T and v T = v1T
  
vK
T ]T . The iterations are needed to achieve any reduction in the residual at all. 2
preconditioner can be applied very efficiently by incorporating a
Krylov subspace reuse algorithm[13], as the linear systems to be 6 Improving Newton convergence
solved are the same as arise in the small-signal analysis for periodic
time-varying systems. The basic idea of the algorithm is that the Rapid convergence of Newton’s method can only be assured with a
Krylov subspaces associated with the matrices e i!Tc ; H are very good initial guess. To achieve a good initial guess, we first calculate
similar for different !k . Essentially, the Krylov-subspace re-use the periodic steady state response of the circuit with the clock
algorithm allows the preconditioner for the matrix DTc  IN ; J signal applied, while suppressing other non-DC signals. Using
to be applied with only slightly more cost than an iterative solve
with the matrix H .
2
Since the MFT circuit equations are not solved exactly, we have found that, on
average, there is a performance advantage in the MFT method to using approximate
Figure 2 shows the effectiveness of this preconditioner in solves of the Newton update equation, and therefore GMRES is converged to a relatively
compressing the eigenvalues for an example RF receiver circuit. loose tolerance.
the steady state solution as an operating point, a small-signal[13] 20
analysis is performed by treating non-clock fundamentals as small Output
Input
signals. As a result of the small signal analysis, amplitudes at
fs + ks fc , for ;Ks  ks  Ks, 1  s  S , are generated. These 0

amplitudes are transformed into time domain initial conditions via −20
inverse Fourier transform. At higher input power levels, using a
Newton continuation [1] method, with the amplitude of the non-
−40
clock signals as the continuation parameter, is generally effective

dB
in securing convergence.
−60

7 Spectrum calculation −80

After the solution is converged, the values

v̄ v t T
v(t2)T
  
v(tK )T ]T and the integration solution
−100

=  ( 1)
in t
t
 i i+Tc ], i = 1
  
K are available. From these pieces of
information, the spectrum v(t) can be obtained. Let
−120
−100 −50 0 50 100 150 200 250 300 350 400
Hz

Kc
X X1K KS
X
v(t) =  V (k1
  
kS
kc ) Figure 4: Harmonic distortion of a switched-capacitor filter
kc =;Kc k1 =;K1 kS =;KS
 e;|2k1 f1 t    e;|2kS fS t e;|2kc fc t : (31) 8 Test results
Define v̄( ) = v(t1 +  )T
v(t2 +  )T
  
v(tK +  )T ]T . Then 8.1 Switched-capacitor Filter
v̄( ) = f(Γ;1 Ω( ))  IN g (32) The first example is a low-pass switched-capacitor filter of 4kHz
2 3 bandwidth and having 238 nodes, resulting in 337 equations. To
..
. analyze this circuit, the MFT analysis was performed with an 8-
6 P
Kc ;|2kc fc  7
phase 100kHz clock and a 1V sinusoidal input at 100Hz.
6 kc =;Kc V (k1
  
kS
kc )e :
7
4 5 The 1000 to 1 clock to signal ratio makes this circuit difficult
..
. for traditional circuit simulators to analyze. In the MFT method,
three harmonics were used to model the input signal. The eight-
Then for each KN -vector V (
kc ), where ;Kc  kc  Kc, phase clock resulted in the need to use about 1250 timepoints
N -vectors V (k1
  
kS
kc ), where
which is collection of all in each transient integration. This brings the total number of
;K1  k1  K1
  
;KS  kS  KS (the actual order is variables solved by the analysis to slightly less than three million
determined by the Fourier transform), (337  (2  3 + 1)  1250 = 2
948
750). The simulation took a
little less than 20 minutes CPU time to finish, on a Sun UltraSparc1
Z Tc
f(Ω( );1Γ)  IN gv̄( )e|2kc fc  d
workstation with 128 Megabyte memory and a 167MHz CPU clock.
V (
kc ) = T1 (33) Figure 4 shows the output spectrum of the filter.
c 0

Forming f(Ω( ) ;1Γ)  IN gv̄( ) requires the values for v(t1 + 8.2 High-performance receiver
 )
  
v(tK +  ), or synchronized time steps between cycles. The The second example is the high-performance image rejection
total cost is one KN -vector integration and M Fourier transforms, receiver also discussed in [14]. It consists of a low-noise amplifier, a
where M is the number of synchronized time points. splitting network, two double-balanced mixers, and two broad-band
The synchronized time step requirement may not be easily Hilbert transform ouput filters combined with a summing network
met in practice. One alternative is to use interpolation schemes. that is used to supress the undesired side-band. A limiter in the LO
However they potentially lose accuracy. Another alternative is to path is used for controlling the amplitude of the LO. It is a rather
trade integrations for Fourier transforms. Specifically, it is easy to large RF circuit that contains 167 bipolar transistors and uses 378
verify that nodes. This circuit generated 987 equations in the simulator.
To determine the intermodulation distortion characteristics, the
V (k1
  
kS
kc ) (34) circuit was driven by a 780MHz LO and two 50mV closely placed
Z Tc
EpT f(Ω( );1 Γ)  IN gv̄( )e|2kc fc  d
1 RF inputs, at 840MHz and 840MHz+10KHz, respectively. Three
=
Tc 0
harmonics were used to model each of the RF signals. 200
time points were used in each transient clock-cycle integration,
Z Tc S P
= EpT (Γ  IN )( T1c v̄( )e|2( i= ki fi ) e|2kc fc  d )

1
considered to be conservative in terms of accuracy for this circuit.
As a consequence, nearly ten million unknowns (987  (2  3 +
1)2  200 = 9
672
600 ) were generated. It took 55 CPU minutes
0

where Ep is a KN  N block matrix whose pth N  N block is I N to finish on a Sun UltraSparc10 workstation with 128 Megabytes of
and other blocks zero, and p is determined by (k 1
  
kS ) from the physical memory and a 300MHz CPU clock. Figure 5 shows 3rd
Fourier transform. Calculating (34) does not require synchronized and 5th order distortion products.
time points. The total cost of calculating V (
kc ) is K KN -vector To understand the efficiency of the MFT method, consider that
integrations plus one final Fourier transform. However, it might be traditional transient analysis would need at least 80,000 cycles of
more expensive since integrations normally cost more than Fourier the LO to compute the distortion, a simulation time of over two
transforms. days. Additionally, the results would be very inaccurate, because
 0
References
−10
[1] A. ALLGOWER AND K. GEORG, Numerical Continuation
Methods, Springer-Verlag, New York, 1990.
−20
[2] L. O. CHUA AND A. USHIDA, Algorithms for computing almost
−30 periodic steady-state response of nonlinear systems to multiple
input frequencies, IEEE Trans. Circuits and Systems, 28
(1981), pp. 953–971.
dB

−40

−50 [3] K. KUNDERT, J. WHITE, AND A. SANGIOVANNI-VINCENTELLI,

A mixed frequency-time approach for distortion analysis of
−60
switching filter circuits, IEEE J. Solid State Circuits, 24
−70
(1989), pp. 443–451.

−80
[4] K. S. KUNDERT, J. K. WHITE, AND A. SANGIOVANNI-
59.97 59.98 59.99 60
MHz
60.01 60.02 60.03 60.04
VINCENTELLI, Steady-State Methods for Simulating Ana-
log And Microwave Circuits, Kluwer Academic Publishers,
Boston, 1990.
Figure 5: Intermodulation distortion of a high-performance receiver
[5] P. LANCASTER AND M. TISMENETSKY, The Theory of Matrices,
Academic Press, second ed., 1985.
of the large amount of numerical error accumulated by integrating [6] D. LONG, R. MELVILLE, K. ASHBY, AND B. HORTON, Full chip
over so many cycles. In contrast, the MFT method is able to resolve harmonic balance, in Proceedings of the Custom Integrated
very small signal levels, such as the 5th order distortion products Circuits Conference, May 1997.
show in Figure 5.
Solving the MFT equations by direct factorization methods is [7] R. MELVILLE, P. FELDMANN, AND J. ROYCHOWDHURY, Effi-
also impractical, as the storage needed for the factored rank-50,000 cient multi-tone distortion analysis of analog integrated cir-
(987  (2  3 + 1)2 = 48
363 ) MFT Jacobian of Equation 19 is cuits, in Proceedings of the Custom Integrated Circuits Con-
several gigabytes. Forming the Jacobian matrix by direct methods ference, May 1995.
would also require computation time proportional to the cost of
50,000 transient integrations, again a number on the order of days. [8] M. OKUMURA, T. SUGAWARA, AND H. TANIMOTO, An effi-
cient small signal frequency analysis method for nonlinear
9 Conclusion circuits with two frequency excitations, IEEE transactions of
computer-aided design of integrated circuits and systems, 9
(1990), pp. 225–235.
In this paper we have demonstrated that the MFT method is an
efficient approach to analyzing multi-frequency nonlinear effects [9] J. ROYCHOWDHURY, Efficient methods for simulating highly
such as intermodulation distortion. Making the MFT method nonlinear multirate circuits, in Proceedings of the 34 th Design
computationally efficient on problems of engineering interest Automation Conference, Anaheim, CA, June 1997, pp. 269–
required careful construction of the delay matrix, matrix-implicit 274.
Krylov subspace iterative linear solvers, and a preconditioner
tailored to the MFT method and the circuits it typically analyzes. As [10] J. ROYCHOWDHURY, D. LONG, AND P. FELDMANN, Cyclo-
a result, nonlinear systems comprising tens of millions of unknowns stationary noise analysis of large RF circuits with multitone
can be solved in less than an hour with computational resources excitations, IEEE J. Sol. St. Circuits, 33 (1998), pp. 324–336.
commonly available to engineering designers. [11] Y. SAAD AND M. H. SCHULTZ, GMRES: A generalized minimal
One salient advantage of the MFT method as described here residual algorithm for solving nonsymmetric linear systems,
is that the dominant part of the computation is in computing
the functions  and the product of the Jacobian of  with some
SIAM J. Sci. Stat. Comput., 7 (1986), pp. 856–869.
vector. Both computations are essentially the solution of an initial [12] R. TELICHEVESKY, K. S. KUNDERT, AND J. K. WHITE, Efficient
value problem. Each application of the operator D Tc ; J , or steady-state analysis based on matrix-free Krylov-subspace
calculation of the Newton residual, involves solving K such initial methods, in Proceedings of the 1995 Design Automation
value problems, that is, integrating K sets of DAEs forward in Conference, June 1995.
time over one clock period. Each of the K problems, however, is
essentially decoupled. Parallel implementations of the MFT will [13] , Efficient AC and noise analysis of two-tone RF circuits,
therefore enjoy very efficient processor utilization. This decoupling in Proceedings of the 1996 Design Automation Conference,
also assists the implementation of out-of-core solvers. In fact we June 1996.
have observed it is possible to implement the MFT algorithm as an
[14] R. TELICHEVESKY, J. WHITE, AND K. KUNDERT, Receiver
out-of-core algorithm with over 80% average CPU utilization.
Several possible extensions of this work are possible. Of characterization using periodic small-signal analysis, in
particular interest is the computation of poly-cyclostationary[10] Proceedings of the Custom Integrated Circuits Conference,
noise statistics that can be performed by linearizing around the the May 1996.
quasiperiodic steady-state. [15] Y. THODESEN, Two stage method for efficient simulation of
parametric circuits, PhD thesis, Department of telecommuni-
cations, the Norwegian institute of technology, 1996.