See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/333999298

Simulation of High-Speed Interconnects

Article in Proceedings of the IEEE · May 2001

DOI: 10.1109/5.929650

CITATIONS READS
614 545

2 authors:

Ramachandra Achar Michel S. Nakhla

Carleton University Carleton University
250 PUBLICATIONS 4,303 CITATIONS 365 PUBLICATIONS 8,318 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Michel S. Nakhla on 27 August 2013.

The user has requested enhancement of the downloaded file.

Simulation of High-Speed Interconnects
RAMACHANDRA ACHAR, MEMBER, IEEE AND MICHEL S. NAKHLA, FELLOW, IEEE

Invited Paper

With the rapid developments in very large-scale integration High-speed interconnect problems are not always handled
(VLSI) technology, design and computer-aided design (CAD) appropriately by conventional circuit simulators, such as
techniques, at both the chip and package level, the operating fre- SPICE [23]. If not considered during the design stage, these
quencies are fast reaching the vicinity of gigahertz and switching
times are getting to the subnanosecond levels. The ever increasing interconnect effects can cause logic glitches that render a
quest for high-speed applications is placing higher demands fabricated digital circuit inoperable or they can distort an
on interconnect performance and highlighted the previously analog signal such that it fails to meet specifications. Since
negligible effects of interconnects, such as ringing, signal delay, extra iterations in the design cycle are costly, accurate pre-
distortion, reflections, and crosstalk. In this review paper, various diction of these effects is a necessity in high-speed designs.
high-speed interconnect effects are briefly discussed. In addition,
recent advances in transmission line macromodeling techniques Hence, it becomes extremely important for designers to sim-
are presented. Also, simulation of high-speed interconnects using ulate the entire design along with interconnect subcircuits
model-reduction-based algorithms is discussed in detail. as efficiently as possible while retaining the accuracy of
Keywords—Asymptotic waveform evaluation, AWE, CFH, cir- simulation [23]–[139].
cuit simulation, complex frequency hopping, distributed networks,
high-speed interconnects, Krylov-subspace methods, macro- A. What is High-Speed?
modeling, model-order reduction, moment-matching techniques, Speaking on a broader perspective, a “high-speed inter-
multiconductor transmission lines.
connect” is the one in which the time taken by the propa-
gating signal to travel between its end points cannot be ne-
I. INTRODUCTION glected. An obvious factor that influences this definition is
the physical extent of the interconnect; the longer the inter-
The recent trend in the VLSI industry toward miniature
connect, the more time the signal takes to travel between its
designs, low power consumption, and increased integration
end points. Smoothness of signal propagation suffers once
of analog circuits with digital blocks has made the signal in-
the line becomes long enough for the signal’s rise/fall times
tegrity analysis a challenging task. The quest for high-speed
to roughly match its propagation time through the line. Then
applications has highlighted the previously negligible effects
the interconnect electrically isolates the driver from the re-
of interconnects (Fig. 1), such as ringing, signal delay, distor-
ceivers, which no longer function directly as loads to the
tion, reflections, and crosstalk. Interconnects can exist at var-
driver. Instead, within the time of the signal’s transition be-
ious levels of design hierarchy (Fig. 2) such as on-chip, pack-
tween its high and low voltage levels, the impedance of in-
aging structures, multichip modules, printed circuit boards,
terconnect becomes the load for the driver and also the input
and backplanes. It is predicted that interconnects will be re-
impedance to the receivers [1]–[12]. This leads to various
sponsible for majority of signal degradation in high-speed
transmission line effects, such as reflections, overshoot, un-
systems [1]–[22].
dershoot, crosstalk, and modeling of these needs the blending
of EM and circuit theory.
Alternatively, the term “high-speed” can be defined in
Manuscript received July 31, 2000; revised February 16, 2001. This work terms of the frequency content of the signal. At low frequen-
was supported in part by the Natural Sciences and Engineering Research cies an ordinary wire, in other words, an interconnect, will
Council of Canada (NSERC), by Micronet, a Canadian Network of Centers effectively short two connected circuits. However, this is not
of Excellence on Microelectronics, Communication and Information Tech-
nology Ontario (CITO), by Canadian Microelectronics Corporation (CMC), the case at higher frequencies. The same wire, which is so
by Nortel Networks, Ottawa, and by Gennum Corporation, Toronto. effective at lower frequencies for connection purposes, has
The authors are with the Department of Electronics, Carleton Uni- too much inductive/capacitive effects to function as a short
versity, Ottawa, ON K1S 5B6, Canada (e-mail: [email protected];
[email protected]). at higher frequencies. Faster clock speeds and sharper slew
Publisher Item Identifier S 0018-9219(01)03967-6. rates tend to add more and more high-frequency contents.

0018–9219/01$10.00 © 2001 IEEE

PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001 693

Fig. 1. High-speed interconnect effects.

sidered” are interconnect length, cross-sectional dimensions,

signal slew rate, and the clock-speed. Other factors that also
should be considered are logic levels, dielectric material, and
conductor resistance. Electrically short interconnects can be
represented by lumped models whereas electrically long in-
terconnects need distributed or full-wave models.
The rest of the paper is organized as follows. In Section II,
a brief description of high-speed effects and interconnect
models is provided. Section III provides a detailed analysis of
transmission line equations and derivation of a generic mul-
ticonductor transmission line stamp, suitable for inclusion in
Fig. 2. Interconnect hierarchy.
an MNA analysis. Section IV provides a review of circuit
equations in the presence of distributed elements. Review of
efficient techniques for discretization of Telegrapher’s equa-
An important criterion used for classifying interconnects
tions is given in Section V. Sections VI–VIII give a detailed
is the electrical length of an interconnect. An interconnect
account of simulation of interconnects using model-reduc-
is considered to be “electrically short” if, at the highest
tion techniques. Section IX provides references to related ad-
operating frequency of interest, the interconnect length
vanced topics.
is physically shorter than approximately one-tenth of
the wavelength (i.e., length of the interconnect/ ,
II. HIGH-SPEED INTERCONNECT EFFECTS
). Otherwise, the interconnect is referred to as
“electrically long” [1], [12]. In most digital applications, High-speed effects influencing a signal propagating on
the desired highest operating frequency (which corresponds an interconnect could be multifold, such as delay, rise time
to the minimum wavelength) of interest is governed by degradation, attenuation, crosstalk, skin effect, overshoots,
the rise/fall time of the propagating signal. For example, undershoots, ringing, and reflection. In this section, we give
the energy spectrum of a trapezoidal pulse is spread over a detailed account of each of these high-speed effects (which
an infinite frequency range; however, most of the signal are also known as transmission line effects).
energy is concentrated near the low-frequency region and
decreases rapidly with the increase in frequency. Hence, A. Propagation Delay
ignoring the high-frequency components of the spectrum A signal traversing from one end of a transmission line to
above a maximum frequency, , will not seriously alter the other end takes a finite amount of time; in other words, it
the overall signal shape. Consequently, for all practical experiences a certain amount of delay . Fig. 3 illustrates
purposes, the width of the spectrum can be assumed to be the case of an ideal delay line. In addition, the signal may
finite. A practically used relationship between the desired encounter rise time degradation as shown in Fig. 4, where
and the (rise/fall time of the signal) can be expressed the rise time at the receiver end is larger than the rise
as [2], [4], [36], [66], [90] time at the source end [2], [4]. Rise-time degradation
further adds to the overall delay experienced by the signal,
(1)
as it influences the maximum and minimum attainable logic
This implies that, for example, for a rise time of 0.1 ns, the levels between the switching intervals.
maximum of frequency of interest is approximately 3 GHz or
the minimum wavelength of interest is 10 cm. In some cases, B. Attenuation
the limit can be more conservatively set [90] as . The signal through an interconnect may suffer attenua-
In summary, the primary factors that influence the decision tion, due to ohmic or conductance losses. This is illustrated
“whether high-speed signal distortion effects should be con- in Fig. 4. Ohmic losses are more pronounced at higher fre-

694 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Fig. 3. Illustration of propagation delay.

Fig. 4. Illustration of attenuation and rise-time degradation.

quencies due to the uneven current distributions. Conduc-
tance losses are proportional to the dielectric loss factor of the
dielectric material and are also a function of the frequency. the phenomenon of overshoots, which occur when the termi-
If the losses are high, the signals may not retain the spec- nating impedance is larger than the characteristic impedance
ified logic levels during the transit through an interconnect of the line. As seen, the undershoots, overshoots, and the
and may result in false switching of digital circuitry. ringing experienced by the signal increases with the delay
of the interconnect. Fig. 6 illustrates the ringing associated
C. Signal Reflection and Ringing with a lossy line, for various cases of terminations.
Signal reflection and the associated ringing can severely 1) Mechanism of Reflections: Consider the interconnect
distort signal propagation at higher frequencies. The prime system shown in Fig. 7, which shows the simplest case of
cause of reflection-related signal degradation is the discon- impedance variation from to . Such a variation re-
tinuity in characteristic impedance of the transmitting line. sults in part of the onward propagating signal getting
Such a discontinuity can be either distributed or lumped in reflected . The coefficient of reflection is given by
nature. In the case of distributed discontinuity, the impedance . As seen, the reflection
variation on a line takes place over a certain length. For ex- will vanish when there is no mismatch . Care must
ample, this can occur due to the change in the medium along be taken in high-speed designs to minimize the reflections,
the length of the signal trace, which may have to traverse which otherwise may result in false switching.
several layers on a printed circuit board (impedance may
not be well controlled from layer to layer). Following are D. Crosstalk
some common causes of discontinuities: connectors between Crosstalk refers to the interaction between signals that are
card-to-board, cable-to-card, leads between chip and chip propagating on various lines in the system. An analogy of
carriers, or between card wiring and chip carriers, long vias, crosstalk could be the “the interference from other lines while
orthogonal wiring, flip-chip soldier balls, wire bonds, and re- talking on the phone.” Crosstalk is mainly due to the dense
distribution lines, etc. wiring required by compact and high-performance systems.
Another major contributor to the reflection related signal High-density and closely laid interconnects result in electro-
degradation is the impedance mismatch between the line magnetic coupling between signal lines. The active signal
characteristic impedance and source/terminating imped- energy is coupled to the quiet line through both mutual ca-
ances. Fig. 5 illustrates these effects for the case of a lossless pacitance and inductances, resulting in noise voltage–cur-
line. Fig. 5(b) and (c) shows the undershoots for relatively rents. This may lead to inadvertent switching and system
small and large delay lines. In general, undershoots occur malfunctioning. Crosstalk is a major constraint while routing
when the terminating impedance is less than the character- in high-speed designs. An example of crosstalk is given in
istic impedance of the interconnect. Fig. 5(d) and (e) shows Fig. 8. By its very nature, crosstalk analysis involves systems

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 695

Fig. 5. Illustration of undershoots, overshoots, and ringing in lossless interconnects.

Fig. 6. Illustration of ringing in lossy interconnects.

of two or more conductors. Such systems are studied on the E. High-Speed Interconnect Models
basis of dominant propagating modes. System behavior in re-
sponse to any general excitation is then a linear combination Depending on the operating frequency, signal rise times,
of modal responses. and nature of the structure, the interconnects can be mod-

696 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

 the signal conductor. To account for these effects, modeling
based on frequency-dependent p.u.l. parameters may be
necessary. An illustration of frequency-dependent variation
of and parameters for an example microstrip are given
in Fig. 9.
Fig. 7. Reflection due to impedance mismatch. 4) PEEC and rPEEC Models: As switching speeds
extend into the gigahertz range, two-dimensional (2-D)
eled as lumped, distributed (frequency independent/depen- transmission line models become inadequate due to the
dent RLCG parameters, lossy, coupled), or full-wave models. spatial EM effects of three-dimensional structures. The mod-
1) Lumped Models: At lower frequencies, the intercon- eling of these structures has been successfully accomplished
nect circuits could be modeled using lumped or using partial element equivalent circuit (PEEC) models.
circuit models. circuit responses are monotonic in na- PEEC models are circuits where individual resistances
ture. However, in order to account for ringing in signal wave- and capacitances are extracted from the geometry using a
forms, circuit models may be required. Usually lumped quasi-static (nonretarded) solution of Maxwell’s equations.
interconnect circuits extracted from layouts contain a large The rPEEC models include the retardation and provide
number of nodes that make the simulation highly CPU inten- full-wave solution. Simulation of these models is relatively
sive (more details about lumped circuit modeling are given CPU intensive as they involve large resultant networks
in Section V). [15]–[22].
2) Distributed Transmission Line Models: At relatively
higher signal-speeds, electrical length of interconnects III. DISTRIBUTED TRANSMISSION LINE EQUATIONS
becomes a significant fraction of the operating wavelength, Transmission line characteristics are in general described
giving rise to signal distorting effects that do not exist at by Telegrapher’s equations. Consider the transmission line
lower frequencies. Consequently, the conventional lumped system shown in Fig. 10(a). Telegrapher’s equations for such
impedance interconnect models become inadequate and a structure can be derived by discretizing the line into infin-
transmission line models based on quasi-transverse electro- itesimal sections of length and assuming uniform p.u.l.
magnetic mode (TEM) assumptions are needed. The TEM parameters of resistance , inductance , conductance
approximation represents the ideal case, where both and , and capacitance . Each section then includes a re-
fields are perpendicular to the direction of propagation sistance , inductance , conductance , and ca-
and it is valid under the condition that the line cross section pacitance [Fig. 10(b)]. Using Kirchhoff’s current and
is much smaller than the wavelength. However, the inho- voltage laws, one can write [12]
mogeneties in practical wiring configurations give rise to
or fields in the direction of propagation. If the line
cross section or the extent of these nonuniformities remain
(2)
a small fraction of the wavelength in the frequency range of
or
interest, the solution to Maxwell’s equations are given by
the so-called quasi-TEM modes and are characterized by (3)
distributed , , , per unit length (p.u.l.) parameters
[12] (discussed in detail in Section III). Taking the limit , one gets
In practical situations, owing to complex interconnect ge-
ometries and varying cross-sectional areas, the interconnects (4)
may need to be modeled as nonuniform lines. In this case,
the p.u.l. parameters are functions of the distance, along the Similarly, we can obtain the second transmission line
length of the transmission line [96]–[98]. equation
3) Distributed Models with Frequency-Dependent Pa-
rameters: At low frequencies, the current in a conductor is
distributed uniformly throughout its cross section. However, (5)
as the operating frequency increases, the current distribution
gets uneven and starts getting concentrated more and more Substituting (2) in (5), we have
near the surface or edges of the conductor. This phenomenon
can be categorized as follows: skin, edge, and proximity
effects [12], [30], [99], [100]. The skin effect causes the
current to concentrate in a thin layer near the conductor sur-
face and this reduces the effective cross section available for
signal propagation. This leads to an increase in the resistance
to signal propagation and other related effects [9]. The edge
effect causes the current to concentrate near the sharp edges
of the conductor. The proximity effect causes the current to
concentrate in the sections of ground plane that are close to (6)

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 697

Fig. 8. Illustration of crosstalk.

Fig. 9. An example of frequency-dependent variation of resistance and inductance.

Fig. 10. Transmission line system.

or where and represent the p.u.l. impedance and admit-

tances of the transmission line, given by
(11)
The set of equations represented by (9) and (10) can be solved
if they can be written in terms of one of the unknowns [either
or ] as follows:

(12)
(7)

Taking the limit , one gets (13)

where is the complex propagation constant, given by

(8)
(14)
Taking the Laplace transform of (4) and (8), one can write
where represents the real part of the propagation constant
and is known as the attenuation constant, whose units are
(9) expressed in nepers/m. represents the imaginary part of
the propagation constant and is known as the phase constant,
(10) whose units are expressed in radians/m. The solution of (12)

698 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

 where and represent the impedance and admittance ma-
trices, given by
(22)
The , , , and matrices are obtained by a 2-D solution
of Maxwell’s equations at appropriate positions, along the
propagation axis. For this purpose, depending on the nature
and geometry of the structure, and the desired accuracy, tech-
niques based on quasi-static or full-wave approaches can be
used. The , , , and matrices are symmetric and posi-
tive definite [12], [90].

B. Multiconductor Transmission Line Stamp

In this section, we derive a stamp relating the terminal cur-
rents and voltages of MTL structures, suitable for inclusion
in SPICE-like simulators. The transmission line stamp [63]
Fig. 11. Multiconductor transmission line system. is derived through decoupling of MTL equations.
Differentiating the partial differential equations given in
and (13) can be obtained as a combination of forward-re- (20) and (21) with respect to , we have
flected waves traveling on the line as
(23)
(15)
(16) (24)
The phase shift and attenuation experienced by the traveling Substituting (21) in (23) and (20) in (24), we get the fol-
waves are given by and , respectively. If lowing two sets of coupled wave equations:
the lines are lossless, the propagation constant is given by
. The line in this case rep- (25)
resents a pure-delay element.
(26)
A. Multiconductor Transmission Line System
Consider the multiconductor transmission line (MTL) Decoupling of equations in (25) or (26) can be achieved
system, with coupled conductors, shown in Fig. 11. through the use of suitable modal transformation matrices
Using steps similar to the case of single transmission line, [11]. For this purpose, introduce a transformation relating
we can derive the multiconductor transmission line equa- the circuit voltages and modal voltages as
tions. Per-unit-length parameters ( , , , and ) in this (27)
case become matrices and voltage–current variables become
vectors represented by and , respectively. Noting these Hence, (25) can be rewritten as [for simplicity, we omit the
changes, we can rewrite (4) and (8) as accompanying term ]

(17) (28)
or
(18)
(29)
The MTL equations represented by (17) and (18) are a set of
coupled first-order partial differential equations (PDE) For effective decoupling of equations to take place, the ma-
and they can be put in a more concise form as trix product in parenthesis must lead to a diagonal matrix as

(30)

where the diagonal matrix contains the eigenvalues of the

product , which corresponds to the roots of the charac-
(19)
teristic equation
For the case of multiconductors, (9)–(11) are modified as
(31)
(20) where represents the unity matrix (we assume the general
case that there exist distinct eigenvalues). Next, as is evi-
(21) dent, the transformation matrix , which relates the circuit

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 699

voltages and modal voltages, consists of linearly indepen-
dent column vectors , corresponding to the eigenvectors
of product , given by

(32)

[Similarly, we can write a transformation relating the cir-

cuit currents and modal currents as . Hence,
(26) can be rewritten as . The di- Fig. 12. Example circuit for MNA formulation.
agonalization of leads to the same diagonal matrix ,
represented by (30). (This can be easily proved by noting that A relationship between the near-end and far-end
. In other can be derived using (38) as
words, matrices and are similar or they have the
same eigenvalues.) The transformation matrix consists of (39)
the eigenvectors corresponding to the product .]
Having obtained the propagation constants, the solution of Using (36) and (39) and eliminating the constants and
(29) can be written in the standard form as , we get

(33)

where represents the th modal voltage and ,

are the corresponding constants, pertaining to incident and (40)
reflected waves, respectively. Equation (33) can be written
in the matrix form as Assume that the -parameter-based stamp of multiconductor
stamp is required in the standard form, where current
flows inwards. In this case, the expression for in (40)
must be multiplied by 1. Noting this and simplifying (40),
we can write the MTL stamp in terms of -parameters as

(34) (41)

where

Defining and premulti-

plying both sides of (34) by the modal transformation matrix
[from (27)], we can write (34) in terms of circuit voltages (42)
as
Matrix Exponential Stamp: An alternative form of the
(35) MTL stamp is also quite popular and it has the matrix
exponential form [72], which is explained below. Equations
where and are constant vectors, which can be deter- (20) and (21) can be written in the hybrid form as
mined from the terminal currents and voltages (i.e., at
and ).
A relationship between the near-end and far-end
voltages can be derived using (35) as

(36)
(43)
Next, substituting (35) in (20), we have
Using the terminal conditions, the solution of (43) can be
written as

(37) (44)

A relationship between the forms represented by (41) and

or (44) can be obtained as follows: Define as

(45)
(38)

700 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Using some algebraic manipulations, we can express the re- The above equation, representing a simple three-node cir-
lationships between the hybrid parameters (44) and the -pa- cuit, has the same form as any other MNA matrix repre-
rameters (41) as senting a large linear lumped network. Hence, MNA and
output equations for lumped linear networks can be written
using a generic notation as

(49)
where and are selector matrices, with entries (0 or 1),
(46) and the superscript “ ” denotes the transpose. Let
. From (49), MNA equations in the frequency-domain
can be written as

(50)
For the case of nonlinear elements, MNA equations in (49)
can be modified as

(47) (51)
Similarly, another useful representation of the MTL stamp is where is a nonlinear function of .
in terms of ABCD parameters, which can be written as 1) Formulation of Linear Subnetworks Containing
Distributed Elements: Consider a linear subnetwork
containing distributed elements. Using (41), the fre-
(48) quency-domain equations of a distributed subnetwork
In the next section, we will review a generic formulation of containing coupled conductors can be written as [63]
distributed interconnect circuit equations, suitable for gen- (52)
eral purpose circuit simulators. where and represent the Laplace-domain
terminal voltages and currents of the distributed element,
IV. FORMULATION OF CIRCUIT EQUATIONS respectively, represents the admittance matrix having
Prior to introducing interconnect simulation algorithms, complex dependency on frequency, which are described in
it would be useful to review a generic formulation of cir- terms of line parameters. Equation (50) representing the
cuit equations. For both frequency or time-domain analysis, lumped linear network can be combined with (52) as
the first step is to set up the modified nodal analysis matrix
(53)
(MNA) [140]. For example, consider the circuit in Fig. 12. Let
be the desired output. Using Kirchhoff’s current law, the where
time-domain MNA and the output equations can be written • are constant matrices describing
as the lumped memory and memoryless elements of sub-
network , respectively, and is the node-space of
subnetwork ;
• is the selector matrix that maps the terminal currents
of the distributed subnetwork to the nodal space of the
linear subnetwork , and is the unity matrix;
• is a constant vector with entries determined
by independent voltage–current sources of subnetwork
, and is the vector of node voltage
waveforms appended by independent voltage source
currents, linear inductor current waveforms of linear
subnetwork .
Equation (53) can be concisely written as
(54)
2) Generic Formulation of Nonlinear Circuits with Dis-
tributed Elements: Consider a general network containing
an arbitrary number of nonlinear and linear (lumped and dis-
tributed) components. For simplicity, let the linear compo-
nents be grouped into a single linear subnetwork as shown
in Fig. 13.

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 701

Fig. 13. Nonlinear network  containing linear subnetwork  with distributed elements.

Fig. 14. Example circuit with lumped/distributed/measured and nonlinear devices.

Using (51), without loss of generality, the circuit equations The linear multiterminal subnetwork can be character-
[140] for the network can be written as ized in the frequency-domain by its terminal behavior as

(56)
(55) where is the -parameter matrix of subnetwork ,
where is the vector of terminal voltage nodes that connect
• are constant matrices describing the subnetwork to the network , and is the Laplace
the lumped memory and memoryless elements of net- transform of .
work , respectively, and is a constant vector 3) Example: To illustrate the formulation scheme de-
with entries determined by the independent voltage and scribed in this section, consider the circuit shown in Fig. 14.
current sources; The network equations can be written as follows.
• is a function describing the nonlinear elements The entities in (55) with respect to the given nonlinear net-
of the circuit, is the vector of node work can be obtained as
voltage waveforms appended by independent voltage
source current, linear inductor current, nonlinear ca-
pacitor charge, and nonlinear inductor flux waveforms,
is the total number of variables in the MNA for-
mulation, and is the total number of ports in linear
subnetwork ;
• with elements where
, with a maximum of
one nonzero in each row or column, is a selector matrix
that maps the vector of currents entering
the linear subnetwork , into the node space of
the network .

702 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

 using a traditional ordinary differential equation solver such
as SPICE [23], [140]–[151].
CPU Expense: Frequency-domain simulation of large
linear networks is conventionally done by solving (50) or
(53) at each frequency point using LU decomposition and
forward–backward substitution. For time-domain simu-
lation, integration techniques are used to convert a set of
time-domain differential equations into a set of difference
equations. For example, application of the trapezoidal rule
to (51) leads to a nonlinear set of difference equation [151]

(57)
The entities in (53), (54), and (56) characterizing the given To solve (57) at each time point, Newton iterations are
linear subnetwork can be obtained as required, which may need several LU decompositions. This
causes the CPU cost of a time-domain analysis to be expen-
sive (note that and matrices for interconnect networks
are usually very large).
The objectives of interconnect simulation algorithms are
to address both mixed frequency/time problem as well as
to handle large linear circuits without too much of CPU
expense. There have been several algorithms proposed for
this purpose, which are broadly classified into two main cat-
egories, as follows. 1) Approaches based on macromodeling
each individual transmission line set. Techniques such as
“method of characteristics” are grouped in this category and
are discussed in detail in Section V. 2) Approaches based
on model-order reduction (such as AWE, CFH, PRIMA)
of the entire linear subnetwork containing lumped as well
as distributed subnetworks and are discussed in detail in
Sections VI–VIII. It is to be noted that the second approach
can also be used in conjunction with the first approach.

V. SIMULATION TECHNIQUES BASED ON

TRANSMISSION-LINE MACROMODELS
In this approach, transmission-line networks described by
Telegrapher’s equations (partial differential equations) are
translated into a set of ordinary differential equations (known
as the macromodel), through some kind of discretization.
The conventional approach [12], [36] for discrete mod-
eling of distributed interconnects is to divide the line into
segments of length , chosen to be small fraction of the
A. Interconnect Simulation Issues wavelength. If each of these segments (assume that the line
Simulation of large interconnect networks is associated is discretized into “ ” segments) is electrically small at the
with two major bottlenecks: mixed frequency/time problem frequencies of interest (i.e., ), then each
and CPU expense. segment can be replaced by lumped models. Generally the
Mixed Frequency/Time Problem: The major difficulty in lumped structures used to discretize transmission lines con-
simulating high-frequency models such as distributed trans- tain the series elements and , and shunt elements
mission lines is due to the fact that, while described in terms and . The parameters ( are the p.u.l.
of partial differential equations, they are best represented in inductance, resistance, conductance, and capacitance of the
the frequency-domain (56). As seen, they do not have a di- line, respectively (Fig. 10).
rect representation in the time-domain. On the other hand, Distributed versus Lumped: Number of Lumped Segments
nonlinear devices can only be described in the time-domain Required: It is often of practical interest to know how many
(55). These simultaneous formulations are difficult to handle lumped segments are required to reasonably approximate a

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 703

Fig. 15. Macromodel using Method of Characteristics.

distributed model. For illustration, consider segments, A sample of such techniques are compact finite differ-
which can be viewed as low-pass filters. For a reasonable ences, integrated congruent transform, and exponential
approximation, this filter must pass at least some multiples Padé-based matrix-rational approximation, and are dis-
of the highest frequency of the propagating signal (say cussed in Section V-D–V-F.
ten times, ). In order to relate these parameters,
we make use of the 3-dB passband frequency of the LC filter A. Method of Characteristics
given by [2], [4] The method of characteristics (MC) [43]–[45] transforms
partial differential equations of a transmission line into or-
(58)
dinary differential equations containing time-delayed con-
where is the length of the line and represents trolled sources.
the delay p.u.l. From (1), we have and using Consider the case of two conductor transmission lines, as
(58), we can express the relation in terms of the shown in Fig. 15(a). An analytical solution, in terms of -pa-
delay of the line and the rise time as , rameters for (9) or (10) can be derived [43] as
or

(59)

In other words, the delay allowed per segment is approx- (61)

imately . Hence, the total number of segments ( ) where is the propagation constant, is the characteristic
needed to accurately represent a total delay of is given by impedance, and are the terminal voltage and current
(60) at the near end of the line, and and are the terminal
In the case of segments, in addition to satisfying (59), voltage and current at the far end of the line. The -parame-
the series resistance of each segment must also be accounted. ters of the transmission line are complex functions of , and
Example: Consider a digital signal with rise time of 0.2 in most cases cannot be directly transformed into an ordinary
ns propagating on a lossless wire of length 10 cm, with a differential equation in the time domain. The MC succeeded
p.u.l. delay of 70.7 ps (this can be represented by a distributed in doing such a transformation, but only for lossless transmis-
model with p.u.l. parameters of nh/cm and sion lines. Although this method was originally developed in
pF/cm). If the same circuit were to be represented by lumped the time domain using what was referred to as characteristic
segments, one needs curves (hence, the name), a short alternative derivation in the
35 sections. It is to be noted that using more sections does not frequency domain will be presented here. By rearranging the
clean up ripples completely but helps to reduce the first over- terms in (61), we can write
shoot (Gibb’s phenomenon). Ripples can be reduced when
losses are properly taken into account. (62)
One of the major drawbacks of the above conventional
Next, (62) can be rewritten as
discretization is that it requires a large number of sec-
tions, especially for circuits with high operating speeds
and sharper rise times. This leads to large circuit sizes (63)
and the simulation becomes CPU inefficient. In order to where
overcome these difficulties, several techniques for effi-
cient discretization were proposed in the literature. These
methods can be broadly classified, based on the passivity (64)
property (details concerning passivity of macromodels can Using (62) and (64), a recursive relation for and can
be found in Section VIII-F) as follows. 1) Macromodels be obtained as
with no guarantee of passivity: A sample of such techniques
are method of characteristics, transfer function approxi-
mation and Chebyshev polynomial-based approximation, (65)
and are discussed in Section V-A–V-C. 2) Macromodels A lumped model of the transmission line can then be de-
with guaranteed passivity by construction of macromodels: duced from (62) and (65), as in Fig. 15(b).

704 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

 If the lines were lossless (in which case the propagation The least square solution of (70) is given by
constant is purely imaginary; ), the frequency domain
(71)
expression (65) could be analytically converted into time-
domain using the inverse Laplace transform as Poles of the system are obtained by computing the roots
of the denominator polynomial and they belong to the
LHS plane only, since they are obtained from an even func-
(66)
tion (68). If any poles are purely imaginary, they are rejected.
where is replaced by a time shift (or delay). Each trans- Next, the residues are obtained by matching the real and
mission line can, therefore, be modeled by two impedances imaginary parts of (67) to the sampled data, as follows:
and two voltage controlled voltage sources with time delay.
Since this transmission line model is in the time domain, it
can be easily linked to transient simulators.
For lossy lines, the propagation constant is not purely
imaginary and, hence, cannot be replaced by a pure delay. .. .. ..
In that case analytical expressions for and cannot be . . .
found in the time domain. To handle such cases, classical
MC can be extended through Padé synthesis of character-
istic impedance and complex propagation constant [44]. In
the case of multiconductor transmission lines, MC can be
applied through decoupling of MTL equations [12].
.. .. .. ..
. . . .
B. Transfer Function Approximation
Least square approximation-based techniques [111] de-
rive a transfer-function representation for the frequency re-
sponse of transmission line subnetworks. The method fits
data from sample frequency points, to a complex rational
..
function , where .

(67)
Re
where and are the th pole–residue pair, is the total Re
number system poles, and is quotient. In order to obtain a ..
stable-model (poles restricted to the left-half plane), the real .
part [even part of ] is fitted to the real part of data Re (72)
samples. Let the real part of (67) be approximated as
Im
..
.
Re (68)
Im
Here, the solution for the residues are obtained by solving the
Writing (68) at several frequency points, , least square approximation, similar to the equations in (71).
and expressing it in a matrix form, we get the matrices in Once the pole–residue model is obtained, it can be easily
(69) at the bottom of the page. Expressing (69) in a simple converted to a time-domain macromodel described in terms
notation of ordinary-differential equations.
Least square approximation provides higher flexibility in
(70) modeling all types of interconnect models. However, the so-

.. Re
.
Re Re Re
.. .. .. .. .. .. (69)
. . . . . .
..
Re Re . Re

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 705

lution of (71) can be ill-conditioned. Also, the algorithm does (10). The variations in space for voltages and currents of a
not guarantee a passive macromodel. transmission line system can be expressed as

C. Chebyshev Polynomials
One of the efficient approaches for discretization is to ex-
press the variations in space for voltages and currents of
a transmission line system in terms of known basis func- (76)
tions, such as Chebyshev, [91], [37] or Wavelet polynomials.
For example, consider the single transmission line equations
(4) and (8). Assume that the voltage and the current where and are the degrees of freedom of approxima-
can be expanded in the form tion, while and are the known expansion func-
tions. Assume that the line is divided into equal segments
of length . The unknown voltages are represented
in terms of values at nodes corresponding to ;
. The current distribution is described
(73)
in terms of its values at the centers of the segments,
where is the th degree Chebyshev polynomial; ; . Next, spatial derivatives
and are the unknown variables. The derivatives of and can be approximated using compact
of and with respect to are also expanded central difference operator as
using Chebyshev polynomials as

(77)

(74) where denotes the node where the operator is centered,

represents either or or . The unknown coefficients
and are computed such that the desired truncation error
where , are related to and as criteria is satisfied. For example, fourth order approximation
is achieved when , . Performing dis-
cretization operation on TL equations (9), (10) results in a
discrete form [90]

(75)

Using (73) and (74) and the orthogonal properties Cheby-

shev polynomials, the Telegrapher’s equations (4), (8) can be
converted to a set of ordinary differential equations in terms (78)
of the unknown coefficients and .
One of the advantages of the algorithm is that it can also be where and .
applied for interconnects with nonuniform line parameters Using suitable corrections for end-points, the set of equations
by expanding line parameters as Chebyshev polynomials represented by (78) can be converted to the standard MNA
with respect to position . Chebyshev approximations are form.
able to achieve better accuracy with fewer variables when One of the advantages of the algorithm is that it can
compared to direct lumped segmentation. However, achieve better accuracy with fewer variables when com-
the algorithm does not guarantee the passivity of the re- pared to direct lumped segmentation. Also, the
sulting macromodel. Similarly, an alternative approximation algorithm guarantees the passivity of the macromodel by
strategy can be adopted based on expansions of and construction [90].
in terms of wavelet basis functions [88], [89].
E. Integrated Congruence Transform
D. Compact Finite-Differences-Based Approximation Consider an -conductor system and the TL equations
Compact finite-differences (CFD) [150]-based ap- represented by (19), which can be expressed after slight mod-
proximations were suggested in the literature to convert ification, in the Laplace-domain as
Telegrapher’s equations into ordinary differential equations
[90]. For the purpose of simplicity, consider the case of
(79)
a single transmission line system, represented by (9) and

706 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

where proximation. Also, it guarantees the passivity of the resulting
macromodel.
Matrix-Rational Approximation: Consider the exponen-
tial form of Telegrapher’s equations describing the multicon-
ductor transmission lines, given by (44)
(80)

and are , the vector of currents and voltages

along the length of the line, of dimension each. Next, (85)
define a transformation as

(81) where is the length of the line. The matrix is approxi-

mated using matrix-rational function as
where
(86)

where and are polynomial matrices ex-

pressed in terms of closed-form Padé rational functions [39]
as
(82)

Note that the transformation matrix is a function of

only, and not dependent on . Substituting (81) in (79), pre- (87)
multiplying by the transpose of and integrating them
with respect to the normalized variable (from 0 to 1), we
get With some mathematical manipulations, (86) can be trans-
lated into a macromodel represented by a set of ordinary dif-
(83) ferential equations, in a closed form. Since all the coefficients
in the macromodel are known a priori, in terms of the p.u.l.
where parameters, the macromodel can be easily stenciled into a cir-
cuit simulator as the stamp of the transmission line. The fact
that the coefficients and are known a
priori in closed form, provides substantial computational ad-
vantage for this algorithm.
It can be proved that the matrix-rational function-based ap-
proximation preserves the passivity of reduced model [39].
(84) Also, the extension of the above matrix-rational approxima-
tion-based technique to handle frequency-dependent param-
The transformation defined by the set of equations eters can be found in [40].
(81)–(84) is called the integrated congruence transform
[86]. Expanding the expressions for and with some VI. MODEL-REDUCTION BASED SIMULATION ALGORITHMS
mathematical manipulations, (83) can be translated into Interconnect networks generally tend to have a large
a set of ordinary differential equations. It can be proved number of poles, spread over a wide-frequency range. Even
that integrated congruence transform-based approximation though the majority of these poles would normally have
preserves the passivity [86]. very little effect on simulation results, however, they make
the simulation to be CPU extensive by forcing the simulator
F. Exponential Padé-based Matrix-Rational Approximation to take smaller step sizes.
This algorithm directly converts partial differential equa- Dominant Poles: Dominant poles are those that are close
tions into time-domain macromodels based on Padé rational to the imaginary axis and significantly influence the time as
approximations of exponential matrices [39], [40], [87]. In well as the frequency characteristics of the system. The mo-
this technique, coefficients describing the macromodel are ment-matching techniques (MMTs) [59]–[75] capitalize on
computed a priori and analytically, using closed-form Padé the fact that irrespective of the presence of a large number of
approximant of exponential matrices. Since closed-form re- poles in a system, only the dominant poles are sufficient to
lations are used, this technique does not suffer from the usual accurately characterize a given system. This effect is demon-
ill-conditioning experienced with the direct application of strated in Fig. 16, where it is clear that pole will have little
Padé approximations. Hence, it allows a higher order of ap- effect on the final transient result.

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 707

Fig. 16. Illustration of dominant poles.

A brief mathematical description of the underlying con- where the superscript denotes the th derivative. Using
cepts of MMTs is given below. Consider a single input/single a simpler notation, we can rewrite (93) as
output system and let be the transfer function. can
be represented in a rational form as

(88)
(94)
where and are polynomials in . Equivalently,
(88) can be written as The coefficients of Taylor series expansion ( ) are also
identical to the time-domain moments of the impulse re-
sponse . This can be easily seen by using the Laplace
(89) transform of [77]

where and are the th pole–residue pair, is the total

number of system poles, and is the direct coupling con-
stant. The time-domain impulse response can be computed
in a closed form using inverse Laplace transform as

(90)

In case of large networks , the total number of poles can

be of the order of thousands. Computing all , the poles (95)
will be highly CPU intensive even for a small network and
for large networks it is completely impractical. Model-re-
duction techniques address the above issue by deriving a re- Due to this analogy, the coefficients of Taylor series expan-
duced-order approximation in terms of dominant poles, sion, ( ), are generally referred to as moments.
instead of trying to compute all the poles of a system. As- It has been shown that the moments provide an estimation
suming that only dominant poles were extracted, (88) can of delay and rise times [57], [58]. Elmore delay [57], which
be rewritten to obtain approximate frequency/time responses, approximates the midpoint of the monotonic step response
as waveform by the mean of the impulse response, essentially
matches the first moment of the response. This can be con-
sidered as one of the basic forms of approximation. However,
(91) in order to get accurate prediction of interconnect effects, it
is essential that the reduced-order model must match (or pre-
serve) as many moments as possible.
(92)
Several algorithms can be found in the literature for re-
duction of large interconnect subnetworks [59]–[94]. They
Moments of the Response: Consider the Taylor series ex- can be broadly classified into two categories: 1) approaches
pansion of a given transfer-function, , at point, based on explicitly matching the moments to a reduced-order
model and 2) approaches based on implicitly matching the
moments. The techniques such as AWE belong to the first
category and are discussed in Section VII. Techniques such
as PVL, PRIMA, which are based on Krylov subspace for-
mulation, belong to the second category and are discussed in
(93) Section VIII.

708 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

VII. MODEL-REDUCTION BASED ON EXPLICIT Comparing from (94) and (100), we note that
MOMENT-MATCHING
These techniques employ Padé approximation, based on
explicit moment-matching to extract the dominant poles and
residues of a given system [8], [59]–[63].
(101)
A. Padé Approximation
Consider a system-transfer function that is approxi- Residues can be evaluated by writing the equations in (101)
mated by a rational function as
in a matrix form as

(96)
where , are the unknowns (total of
variables). Consider the Taylor series expansion
of at , in terms of its moments. Matching
to the rational function approximation given in (96) (hence,
the name moment-matching techniques, which is also known
as Padé approximation), we get
(102)

(97)
In the above equations represents the direct coupling be-
Cross multiplying and equating the coefficients of sim- tween input and output. More accurate ways to compute
ilar powers of starting from to on both sides can be found in [8].
of (97), we can evaluate the denominator polynomial coeffi-
cients as B. Computation of Moments
Having outlined the concept of MMTs, we need to
evaluate the moments of the system, which are required by
(98)–(102). Consider the simple case of lumped circuits
and the corresponding MNA equations represented by (50).
Expanding the vector using the Taylor series, we have
(103)
(98) where represents the th moment-vector. Equating co-
efficients of similar powers of on both sides of (103), we
obtain the following relationships:
The numerator coefficients can be found by equating the re-
maining powers of (from to ) as
(104)
The above equations give a closed form relationship for the
computation of moments. The moments of a particular output
of interest [which are represented in (96)–(102)] are picked
from moment-vectors . As seen, (104) requires only one
(99) LU decomposition and few forward–backward substitutions
during the recursive computation of higher order moments.
Equations (98) and (99) yield an approximate transfer func- Since the major cost involved in linear circuit simulation is
tion in terms of rational polynomials. due to LU decomposition, MMTs yield very high-speed ad-
Alternatively, an equivalent pole–residue model can be vantage (100 to 1000 times) compared to conventional sim-
found as follows. Poles are obtained by applying a root- ulators.
solving algorithm on denominator polynomial . In order Generalized Computation of Moments: In the case of net-
to obtain , the approximate transfer function given by (91) works containing transmission lines, moment-computation
is expanded using Maclaurin series as is not straightforward. A generalized relation for recursive
computation of higher order moments can be derived as fol-
(100) lows [62], [63], [67], [69], [72]. Considering the MNA equa-
tions containing MTL stamps (54) and expanding and

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 709

 in Taylor series at an expansion point , we get From the property of matrix exponentiation of an arbitrary
[72] matrix , we have

(110)
Let
(105)
(111)
where denotes the th derivative of and de-
Hence, (110) can be rewritten as
notes the th moment of at . Equating coeffi-
cients of similar powers of on both sides of (105), we
have
(112)
Expanding the RHS of (112) further, and collecting the terms
in powers of , we have

(106)

Generalizing (106), a recursive relation for any th higher

order moment can be obtained as (113)
Equating (109) with (113) gives

(107)

It can be seen that the coefficient on the left-hand side of

(107) does not change during higher order moment compu-
tation. Hence, it requires only one LU decomposition and
forward–backward substitutions to compute moments.
Also, it is easy to note that the lumped networks are a special
case of (107) [where for , in which case (107)
reduces to the form given by (104)]. Next, (107) requires the
derivatives of . These can be obtained using (53) as (114)
and so on. From the above results, a recursive relationship
for generating transmission line moments can be obtained as

(108)

The derivatives can be obtained as a function of the

derivatives of the entries on the RHS of (46) and proper ap- or (115)
plication of Leibnitz’s theorem. However, this requires the
derivatives of the exponential stamp represented by (45). A Convergence of (115), in practice requires 20–30 terms.
brief review of computation of these derivatives [72] is given It is to be noted that the convergence of the series repre-
below. sented by (110) can suffer, if for the first few terms grows
Transmission Line Moments: Consider the exponential quicker than . In order to control this problem, note that
stamp represented by (45). We wish to expand the exponen- the growth of depends on its eigenvalues. If all the eigen-
tial matrix in Taylor series, as follows: values of are within the unit circle in the complex plane,
then will decay with increasing , leading to fast con-
(109) vergence. From (110) one can see that the eigenvalues of

710 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Fig. 17. Summary of the steps involved in the MMT algorithm.

can be controlled by varying the length . By restricting A summary of the steps involved in the Padé-based circuit
to be small enough, such that the eigenvalues of reduction is given in Fig. 17.
will also be small (over a given frequency range), so as not
to cause truncation errors or slow convergence. This can be C. Limitations of Single Expansion MMT Algorithms
achieved efficiently, by noting that Obtaining a lower order approximation of the network
transfer function using a single Padé expansion is commonly
(116) referred as asymptotic waveform evaluation (AWE) in the
In other words, moments of a line can be generated by literature. However, due to the inherent limitations of Padé
squaring half-line moments. Let represent the half-line approximants, MMTs based on single expansion often give
moments, then inaccurate results. The following is a list of those properties
that have the most impact on MMTs.
• The matrix in (98) (known as Toeplitz matrix) becomes
increasingly ill-conditioned as its size increases. This
implies that one can only expect to detect six to eight
(117)
accurate poles from a single expansion.
which will give • Padé often produces unstable poles on the right-hand
side of the complex plane.
• Padé accuracy deteriorates as we move away from the
(118) expansion point.
• Padé provides no estimates for error bounds.
The line can be subdivided by power of 2 (i.e., two sec- In addition, there is no guarantee that the reduced-model ob-
tions, four sections, eight sections ) and the moments of tained as above is passive. Passivity implies that a network
the smallest section that meets the convergence requirements cannot generate more energy than it absorbs, and no pas-
are calculated. From these, the moments of the entire line can sive termination of the network will cause the system to go
be recursively calculated with the help of (118). unstable [83]–[92]. The loss of passivity can be a serious

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 711

Fig. 18. Illustration of CFH.

problem because transient simulations of reduced networks

may encounter artificial oscillations.
In systems containing distributed elements the number of
dominant poles will be significantly higher, and it is very dif-
ficult to capture all of them with a single Padé expansion.
This lead to the development of multipoint expansion tech-
niques such as complex frequency hopping (CFH), which are
summarized in the next section.

D. Complex Frequency Hopping

CFH extends the process of moment matching to multiple
expansion points (hops) in the complex plane near or on the
imaginary axis using a binary search algorithm [72]. With Fig. 19. Graphical illustration of transfer-function-based search
a minimized number of frequency point expansions, enough algorithm.
information is obtained to enable the generation of an ap-
proximate transfer function that matches the original func- 1) Transfer-Function-Based Approach: In this approach
tion up to a predefined highest frequency of interest. Using the transfer functions obtained at various hops (expansions)
the information from all the expansion points, CFH extracts are used to ensure the accuracy of the reduced-order model
a dominant pole-set as illustrated in Fig. 18(b). In addition, up to the highest frequency of interest. Steps involved in the
CFH provides an error criterion for the selection of accurate algorithm are given in Figs. 19 and 20. It is to be noted that
poles and transfer functions. the computational effort needed for a comparison as required
Selection and Minimization of Hops in CFH: A Padé ap- by Step 5 is trivial as the responses can be computed in a
proximation is accurate only near the point of expansion and closed-form using the transfer functions generated in Steps
its accuracy decreases as we move away from the point of 2 and 3. Here, is a predefined threshold for the relative
expansion (hop). In order to validate the accuracy of such error in the transfer functions. At the completion of the bi-
an approximation, at least two expansion points are neces- nary search algorithm, a set of transfer functions are gener-
sary. Accuracies of these two expansions can be verified by ated. When evaluating the frequency response at a frequency
matching the poles generated at these two hops [72] (re- point , only the transfer function that is valid in the region
ferred as pole-matching-based approach). Alternatively, the containing is used. This is repeated for all other frequency
two hops can be verified for their accuracy by comparing points to obtain the frequency response of the system.
the value of the transfer functions due to both these hops 2) Pole-Matching-Based Approach: In this approach,
at a point intermediate to them (referred as transfer-func- poles of the transfer function are explicitly evaluated at
tion-based approach) [73]. CFH relies on a binary search al- each hop and the hops are verified for their accuracy by
gorithm to determine the expansion points and to minimize comparing the poles from two adjacent hops using a binary
the number of expansions. The steps involved in the binary search algorithm. If a matching pole is found between two
search algorithm for both the above approaches are summa- adjacent expansions, then the binary search is stopped. The
rized as follows. distance between the matching pole and the expansion point

712 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Fig. 20. Transfer-function-based binary search algorithm.

Fig. 21. Pole-matching-based binary search algorithm.

under consideration defines the radius of accuracy for the E. Reduced-Order Models of Multiport Linear Networks
corresponding expansion. All the poles that are within the So far we looked into the model-reduction of single
radius of accuracy are treated as accurate poles and are input-single output systems. In this section, a discussion
retained in the final pole-set. The poles that are outside the pertaining to the multiport networks is given. Consider the
radius of accuracy are considered as inaccurate poles and general multiport interconnect linear subnetwork shown
are discarded. Fig. 21 illustrates the idea of the pole search. in Fig. 22 [this corresponds to the linear subnetwork
Steps involved in this approach are identical to the one given that is described during the general formulation (56) in
in Fig. 20, except for the matching criterion, which is based Section V]. Such a network can be characterized in terms of
on poles here. Once a set of dominant poles are obtained, (admittance), (impedance), (hybrid) or (scattering)
residues of the system are obtained using (102). Further parameters. For the purpose of simplicity of presentation,
details of CFH and its search algorithms can be found in only the discussion with respect to parameters is consid-
[72] and [73]. ered in this paper (the idea presented here can be easily

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 713

 2) CFH accurately computes the dominant poles of a
system. Generally in a system with a large number
of dominant poles (20–40), pole-sets belonging to
different driving point admittances ( ) obtained
using CFH contain mostly identical poles and with
only a very small number of poles differing among
these sets.
Residue Computation Algorithm: Residue computa-
tion algorithm [76], which combines the merits of two
different approaches of CFH, namely, transfer-func-
tion-based approach and pole-matching-based approach.
Fig. 22. Multiport linear subnetwork  . The transfer-function-based approach collects a set of
transfer functions that accurately matches the frequency
extended to other types of characterization). The network response up to the highest frequency of interest. On the
can be characterized using -parameters as other hand, pole-matching-based approach collects all the
dominant poles accurately up to the highest frequency of in-
terest. In this algorithm, the relationship between frequency
response and pole–residue model of the system is used to
compute the residues (this idea is illustrated in Fig. 23) as
(119)

Using model-reduction techniques, each entry in (119) (121)

can be approximated by a -pole lower order model having
a general form as follows:
where is the total number of dominant poles ( ) extracted
using CFH and ( ) are the corresponding unknown residues,
(120) is the direct coupling constant, and represents the fre-
quency response obtained using the transfer-function-based
where and are the th dominant pole–residue pair at approach.
a port due to an input excitation at port , is the direct In order to compute the unknown residues, a set of linear
coupling constant, and is the number of dominant poles equations can be formulated using (121), spanning many fre-
used for approximating . quency points in the region of interest as shown below:
There are two main approaches available in the literature
to obtain the matrix-transfer function represented by (119) (122)
and (120). In the first approach, a common set of poles ob-
tained for any one of the entries are used for the compu-
tation of residues of all other entries [64], [66]. However, this
can lead to inaccurate results, since the residue computation
(123)
using (102) is very sensitive to the location of poles of each
. In the second approach, separate set of poles are used
for the computation of residues of each entry [68]. This ap- where represents the th frequency point. In the case of
proach can lead to a macromodel with a very large number complex poles, both the pole as well as its conjugate should
of poles. However, it is to be noted that the nonlinear simu- be used while formulating (123). Let be the total number
lation time using general purpose simulators is superlinearly of frequency points matched. Next, equating both the real and
proportional to the number of states involved (the number imaginary parts of (123) separately, we obtain a new set of
of states is generally given by Total number of states linear equations as
Total number of poles Total number of ports ). Hence, (124)
the second approach can become inefficient in the presence
of a large number of ports. Both these difficulties can be ad- where is a vector containing the unknowns cor-
dressed using the Block CFH technique [76], which provides responding to the real and imaginary parts of residues,
schemes to minimize the number of poles, as well as to im- is a vector containing real and imaginary parts of
prove the accuracy of residues. the frequency responses represented by , and
Selection of Dominant Pole-Set: In order to minimize the consists of entries contributed by the LHS of (123).
number of poles in the matrix-transfer function, the following In case the direct coupling constant is computed by adding it
two propositions can be used. as one more variable in (124), then the order of matrix/vec-
1) In general, the pole-set corresponding to any indi- tors indicated above will change from to .
vidual transfer impedances ( ) is a subset of the Next, the solution of (124) is needed to evaluate unknown
union of all driving point impedances ( ) [158]. residues . In order to improve the accuracy of residues, it

714 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Fig. 23. Illustration of the new residue computation algorithm.

is necessary to match as many frequency points as possible F. Interface to Circuit Simulators

using (123). However, in the case of choosing a number In this section, a review of techniques to link re-
of frequency points higher than the number of unknown duced-order models to nonlinear SPICE-like simulators is
residues, matrix will not be square. To overcome this described. The reduced-order model for the linear subnet-
difficulty, (124) is solved using the least mean square error work is obtained as shown in (127) at the bottom of the
algorithm and the solution is given by a new set of normal page, where is the number of poles in the common pole-set
equations . and represent residue and the direct coupling
constant for the parameter , respectively. Derivation
of differential equations from reduced-order interconnect
(125) models is referred as macromodel synthesis. The differential
equations can be easily linked to nonlinear simulators as they
where are described in time-domain. This process is illustrated in
Fig. 24. In general, a set of first-order differential equations
in the state–space domain can be described as
(126)

is a square matrix of order . represents the trans-

pose of . Formulating the residue equations in the form (128)
given by (125) leads to additional computational savings as
the number of equations to be solved remains equal to the where is a state-matrix, is a matrix
number of unknown residues despite the increase in matched that relates the inputs to state-variables, is a ma-
frequency points. trix relating state variables to the outputs ( ),
In the multiport CFH algorithm, an accurate frequency-re- is a matrix relating inputs directly to the output, is the
sponse for each of the entries in the matrix is obtained state vector of length , and is the input vector of length
using complex frequency hopping based on transfer-function (where is the number of ports).
approach. Next, an accurate pole-set is obtained as the union Given a matrix-transfer function described by (127),
of driving point impedances. Using these accurate poles several forms of time-domain realization can be obtained.
and frequency responses, residues for each is obtained For the purpose of illustration, macromodel synthesis using
through the residue computation algorithm. Jordan-canonical [152], [153] form of realization is given

(127)

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 715

 where is a diagonal matrix constructed using poles and
is complex conjugate of . Next, introducing an equiv-
alence transformation defined by

(130)

(129) can now be easily transformed into

Re Im Re
Im Re
Fig. 24. Illustration of macromodel synthesis for nonlinear
simulation. Re Im (131)

below. Consider a two-port network containing two common An illustrative example of the above steps is given below.
poles, whose transfer function is described by Consider a two-port network containing one pair of complex
poles . Let the corresponding residues at dif-
ferent ports be . The original Gilbert’s
realization will yield

A Jordan-canonical [152], [153] form of realization for this

case would need four state variables, and it can be repre-
sented as

Define

If the matrix-transfer function contains complex

poles, then they need to be treated differently as they do not
have a direct meaning in the time-domain. However, since Next, the realization represented by (131) can be obtained
all the coefficients of the denominator polynomials of as
(which are obtained through Padé approximation) are real,
if a complex number is a root of then its complex
conjugate is also a root of . Hence, a Jordan-form
of state–space realization for a complex pole-pair would be
[153]

(129)

716 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Fig. 25. Illustration of equivalent subcircuit generation from macromodels.

Unified Transient Simulation: Once a matrix-transfer purpose of illustration, consider a simple case of two-port
function describing the multiport interconnect network is ob- network with two states represented in the form of (132)
tained, a time-domain realization in the form of state–space
equations can be obtained as

(134)
(132)

where and are the vector of terminal currents and volt- Next, (134) can be rearranged as
ages of the linear subnetwork [described by (56)]. The dif-
ferential equations represented by the macromodel (132) can
be combined with (55) using the relation as (135)
(136)
(137)
(138)

In the above equations, the port voltages and currents are

(133) represented by and , respectively. An equivalent
network representing (135)–(138) can be constructed as
shown in Fig. 25. Each state in the macromodel requires a
Using standard nonlinear solvers or any of the general-pur- separate node in the equivalent circuit and are represented by
pose circuit simulators, the unified set of differential equa- nodes . The state variables can be represented
tions represented by (133) can be solved to yield transient by the capacitor voltages. These capacitors are denoted by
solution for the entire nonlinear circuit containing intercon- and the corresponding voltages by . Next,
nect subnetworks. For those simulators (such as HSPICE) the terms such as in (135)–(138) can be represented
that do not directly accept the differential equations as input, by voltage controlled current sources. Equations (135) and
the macromodel represented by (133) can be converted to an (136) are fully represented by Fig. 25(c) and (d). Output
equivalent subcircuit and is described in the next section. equations represented by (137) and (138) are realized
Conversion of Macromodels to Equivalent Subcir- through equivalent circuits shown in Fig. 25(a) and (b).
cuits: Conversion of differential equations to equivalent Generalization of the above discussion in the presence of
subcircuits can be accomplished in several ways. For the more number of states or ports is straightforward.

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 717

VIII. MODEL-REDUCTION BASED ON KRYLOV-SUBSPACE system, as represented by (142). Expanding it in terms of
TECHNIQUES Taylor series, we have
Direct MMTs such as AWE, discussed in Section VII, have
some disadvantages associated with them. First one among
them is the ill-conditioning associated with the moment-ma-
trix. Due to this difficulty, the number of good poles that
could be extracted from any expansion point is generally
fewer than ten poles. The second major difficulty is that they
where (143)
do not guarantee the passivity of reduced models. In order
to address these difficulties, a parallel class of algorithms,
which can be classified as indirect moment-matching tech- Ideally, increasing the order of the Padé approximation
niques were developed [77]–[90]. (which is equivalent to matching more number of moments)
These algorithms are based on what is known as Krylov- should have given us better approximation results. How-
subspace formulation and Congruent transformation. One of ever, in practice, this is true only up to very limited order,
the main features of these algorithms is that they construct beyond which Padé approximation will not yield any better
the reduced-model based on the extraction of leading eigen- results [77], [160]. This can be explained by examining
values (those with the largest magnitude) of a given system the nature of higher order moments, which are given by
(on the contrary, the reduced models from the CFH technique . As can be seen, when successive moments
discussed in Section VII-D is based on extracting the dom- are explicitly calculated, they are obtained as powers of
inant poles of a given system). In the rest of this section, . With the increasing values of “ ”, this process quickly
we will describe the concept and important features of these converges to an eigenvector corresponding to an eigenvalue
algorithms. of with the largest magnitude. As a result, for relatively
large values of “ ”, the explicitly calculated moments
A. Preliminaries , will not add any extra information
Recall from Section IV, the time-domain MNA and the to the moment-matrix, as all of them contain information
corresponding output equations can be represented in the only about the largest eigenvalue. In other words, the rows
form beyond “ ” of moment-matrix are almost identical (or
parallel to each other), making the matrix ill-conditioned.
2) Relationship Between Eigenvalues and Poles of the
System: In this section we will show the correspondence
between the leading eigenvalues and poles of the system. It
(139) is important to understand this concept as the Krylov-sub-
space-based techniques obtain the reduced-models by
extracting the leading eigenvalues of a given system.
where represents the total number of MNA variables. Pre-
Consider (140), and assume that the matrix can be
multiplying both sides of (139) by , we can write
diagonalized in the form

(144)
(140)
where diag is a diagonal matrix whose
diagonal elements represent the eigenvalues of matrix . The
Taking the Laplace transform of (140), we can write
matrix contains the eigenvectors of matrix . Using (144),
the transfer-function represented by (142) can be rewritten as

(141)

Rearranging (141), we can write the transfer function

of the given system as (145)

(142)
which can be simplified as

where is an identity matrix.

1) Why Direct Padé-Based Approximation (Moment-Ma-
trix) is Ill-Conditioned: Consider the transfer-function of a (146)

718 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

where is a function of the eigenvectors of matrix , and Next, consider the circuit equations (140) and a simple
represent the residues. From (146), we can draw following similarity transformation as follows:
inferences. 1) Poles are the reciprocal of eigenvalues of
(150)
matrix ; the leading eigenvalues (those with largest mag-
nitudes) correspond to the poles closer to the origin. 2) The where the transformation matrix is defined as
transfer function of can be easily obtained in terms of
poles and residues, once the eigenvalues and eigenvectors of (151)
are available. and has the upper-Hessenberg companion form dis-
However, for large interconnect circuits, it would be im- cussed above. Obviously, since is related to the matrix
practical to compute all the eigenvalues and eigenvectors. through a similarity transformation, its eigenvalues are
Hence, in the following sections, we will review some of the the same as that of . Although it looks straightforward,
efficient techniques to extract leading eigenvalues. this approach has the following limitations.
Computation of Eigenvalues of Matrix “ ”: In general, Computation of using the relation (150)
the numerical computation of all the eigenvalues and eigen- ( ) requires the inverse of the matrix
vectors of a given matrix becomes exceedingly expensive . However, is a dense matrix and, hence, computation
as its size gets above a few hundred. The general approach of its inverse will be expensive. Also, is likely to be
in such cases is to approximate with a smaller matrix , ill-conditioned since the columns of are formed based
such that the eigenvalues of are reasonable approxima- on the sequence which quickly converges to the eigen-
tion of the leading eigenvalues of . Due to the relatively vector corresponding to the largest eigenvalue. In the next
small size of , finding its eigenvalues will be a much sim- section, we will describe general techniques to overcome
pler problem than finding the eigenvalues of . Next, we will these problems. These algorithms belong to a class of
review some of the basic matrix forms [160], which would methods known as Krylov-subspace techniques.
be helpful in understanding the eigenvalue computation al-
gorithms presented in this section. B. Krylov-Subspace Methods for Iterative Computation of
Orthogonal Matrices: A real matrix is orthogonal if Eigenvalues
(147) We will start by replacing the matrix in (150) with
an orthogonal matrix such that for all , the leading
All columns, of orthogonal matrices have unit two norms
columns of and span the same space. This space is
or (which implies that ) and are orthog-
called Krylov-subspace and is denoted by . In
onal to one another (which means that ). For the
other words, any vector that is a linear combination of the
special case where is a square matrix, the above definition
leading columns of can be expressed also as a linear
implies .
combination of the leading columns of . Mathematically,
QR Decomposition: Let be a matrix with
we will express this as
. Suppose that has full column rank. Then there exists a
unique orthogonal matrix and a unique upper-tri-
angular matrix with positive diagonals ( ) such
that . There are several techniques (such as mod- (152)
ified Gram–Schmidt orthogonalization process) available in
the literature for performing this decomposition [160]. In contrast to matrix , the matrix has the following ad-
Upper-Hessenberg Matrix: A matrix is called vantages.
Upper-Hessenberg if for . For example, • is well conditioned.
consider an upper Hessenberg matrix of order , having the • It is easy to invert since .
following form (which is known as companion form): • Most importantly, we can compute only as many
leading columns of as needed to get accurate solu-
tion (more details about this are covered later in this
.. .. section).
. .
.. .. .. (148) The next question is, how do we get the matrix ? We can
. . . achieve this as follows. Expressing the matrix using QR
.. .. .. decomposition as (where is an orthogonal ma-
. . .
.. .. trix and is an upper-triangular matrix), we can modify
. . (150) as
One of the important advantages of the above companion
form is that its characteristic polynomial can be ana-
lytically computed and is given by

(149) (153)
or
The roots of give the eigenvalues of . (154)

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 719

Since and are both upper triangular and is upper
Hessenberg, it is easy to prove that the new matrix,
, is also upper Hessenberg.
Next, let us assume that we will use only the first leading (a) Computation of q
columns ( ) of . In this case, the matrices under con-
sideration will have the following dimensions: ,
and . The implications of (154) is that we can re-
duce the matrix of dimension to a smaller upper
Hessenberg matrix of dimension using orthogonal
transformation. In addition, the eigenvalues of the smaller
system are approximations of the first leading eigen-
values of the larger system represented by . (b) Computation of q ;h ; andh
Next, we will show that the columns of can be com- Fig. 26. Illustration of steps in Arnoldi algorithm.
puted one at a time giving us the advantage of computing
only as many leading columns of as needed. One of the Premultiplying both sides by , we have
popular approaches used for partial reduction of a large ma-
trix to a smaller upper Hessenberg matrix using , is known
as Arnoldi’s algorithm [80]–[87], [160]. More details about
this are given in the next section. (158)

C. Arnoldi Algorithm for (Partial) Reduction Knowing the value of and using the fact that ,
Assume , where represents the th we can compute from (157) as
column of matrix . From (154), we have , which
can be written as (159)

The direction for can be obtained using (157) as [illus-

trated in Fig. 26(b)]

(160)

Similarly, the rest of the columns of and matrices can

be obtained by generalizing the steps as shown in Fig. 27.
(155) Note that we did not need to explicitly compute the product
.. .. .. .. ..
. . . . . . As a result, we were able to avoid the ill-conditioning
.. .. .. .. .. problem arising due to the quick convergence of the sequence
. . . . .
to the eigenvector of the largest
eigenvalue.
Recall that all columns of orthogonal matrices have The columns computed by Arnoldi algorithm are called
(which implies that ) and are orthogonal Arnoldi vectors. The loop over updating corresponds
to one another (which means that ). Using this to the modified Gram–Schmidt algorithm [160], which
information, the first few steps in obtaining the and subtracts the components in the directions to away
matrices are outlined below. from , leaving them orthogonal to . Computing a total of
Since the , an easy way to compute it is to divide “ ” Arnoldi vectors costs matrix–vector multiplications
the vector [from (139), assuming the single-input, single- involving , plus related cost.
output case, will be a vector] by its magnitude (we There are several alternative methods available in the liter-
get a unit vector in the direction of ). This step is illustrated ature for finding the basis for Krylov-subspace [85]. For ex-
in Fig. 26(a) ample, one can use multiple passes of orthogonalization to
increase the robustness of the modified Gram–Schmidt or-
(156) thogonalization process.
To recap, we started with the circuit equations
To determine and the first column of , we multiply and . We formed the product
by the first column of . This gives us , which is the first . Using orthogonal transformation, we were able
column on LHS of (155). Equating it with the first column to determine the leading eigenvalues of that correspond to
of RHS, we have the dominant poles of the transfer-function. In the following
section, we will show how to use this information to perform
(157) circuit reduction.

720 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Fig. 27. Pseudocode for Arnoldi algorithm.

D. Circuit Reduction Using Arnoldi Algorithm An important criterion during the above reduction is the
Finding the reduced-order circuit equations can be ex- accuracy of the response of the reduced system given by
plained by a change of variables in (139) by mapping the (164). The frequency response of the reduced system (164) is
vector of dimension into a smaller vector of dimension also a good approximation of the frequency-response of the
( ) using the orthogonal matrix [80]–[92] original transfer function (142). An indicator for the accuracy
of the response of the reduced system is the total number of
(161) moments it can preserve (match), for a given order of reduc-
tion ( ). It can be proved that the reduced system (164) of
Using (161), we can rewrite Laplace-domain circuit equa- order preserves the first moments of the original network
tions in (140) as [84].
In essence, we are able to implicitly match the moments
and obtain a reduced-model without the need to directly use
(162) the moments as in the AWE algorithm. Hence, we will not
suffer from the same numerical ill-conditioning that is as-
Premultiplying both sides of (162) by and using the sociated with direct moment-matching algorithms. The ac-
relation , we have curacy of the Arnoldi approximation gradually increases as
the order is increased since more moments of the original
transfer function will be matched.
A question that may possibly arise here: how are the
accuracies of Arnoldi-based approximation and direct
Padé-based approximation are compared? It was shown in
(163)
Section VII that a Padé approximation of order matches the
Hence, the transfer-function of the reduced system can be first moments. However, an Arnoldi-based reduction of
written as order matches only first moments [84]. Essentially, this
means that, for a comparable accuracy, the reduced-model
from Arnoldi will have double the size of the reduced
model from direct Padé-based approximation (in other
(164) words, direct Padé-based models are more optimal). On the
other hand, due to the ill-conditioning, direct Padé-based
Comparing the original transfer-function repre- approximation cannot achieve higher order approximation,
sented by (142) with the transfer-function of the whereas Arnoldi-based approximation can.
reduced system represented by (164), we can draw the In an alternative approach, the reduction based on Lanczos
following conclusions. The eigenvalues of are given algorithm can preserve the first moments [77]–[79] (like
by the eigenvalues of . However, since the eigenvalues of direct Padé-based approximation). The difference between
are good approximation of the leading eigenvalues of , Arnoldi and Lanczos algorithms is that in Lanczos algorithm
we can conclude that the eigenvalues of the transfer function we transform the matrix to a tridiagonal matrix . In addi-
of the reduced system are good approximation of the poles tion, Lanczos algorithm uses two biorthogonal Krylov space,
of the original transfer function. to recursively compute [77]. However, the macromodels

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 721

using the above Lanczos-based reduction scheme does not
guarantee the passivity of the macromodel.

E. Multiport Reduction using Arnoldi Algorithm

Consider an -port characterization of a system based on
the admittance parameters. The multiport circuit equations
for this can be written as [80], [84]

(165)

where the input and the output

Fig. 28. Transient response of a nonpassive macromodel with
are vectors of port voltages and passive terminations.
currents, respectively, instead of scalars. In this case,
becomes a matrix, where . In this case, it can be
instead of a vector. represents the th column of (an proved that the reduced system (168) of order preserves
example of multiport formulation is given in Section VIII-F). the first block moments of the original network [84].
In the case of multiport characterization, block moments This implies that for a desired predefined accuracy, the order
are used, instead of moments of a particular output. Block of the reduced system should be increased with the increase
moments are defined as in the number of ports.
In the next section, the review of an important property of
macromodels, passivity, and extension of Krylov-subspace
(166) techniques for passivity preservation are provided.

where an entry in the th row and th column of repre- F. Passivity Preservation

sents the th moment of the current (output) that flows into Passivity implies that a network cannot generate more en-
port due to a voltage source (only nonzero source) at port ergy than it absorbs, and no passive termination of the net-
. work will make the system unstable. Passivity is an impor-
Block Krylov-subspace spanning columns can be con- tant property, because stable but not passive macromodels
structed using the block moments (166) as can lead to unstable systems when connected to other pas-
sive systems. The loss of passivity can be a serious problem
because transient simulations of nonpassive networks may
encounter artificial oscillations. This is illustrated in Fig. 28,
which represents the transient response of a reduced-order
macromodel of a large linear circuit, when connected
to an external load of 50 k . In this section, we will review
(167)
important algorithms that are available in the literature, for
where the operator represents the truncation to the nearest preservation of passivity during the reduction of interconnect
integer, toward zero. networks [83]–[92].
For example, assume a three-port characterization in- 1) Review of Passivity Properties: A passive network
volving , and let . In this case, we have cannot generate power on its own. It is essential that the
; and (167) can be reduced-order model must be passive. A network with
written as admittance matrix represented by is passive iff [83],
[84], [154]–[157]:
a) where “ ” is the complex conjugate
operator;
b) is a positive real (PR) matrix, that is the product
Using the block Arnoldi algorithm [80], matrix can be re- for all complex values of
duced to a small block upper Hessenberg matrix, . with and any arbitrary vector .
The multiport admittance matrix of the reduced-system is 2) Passive Reduction: The algorithm discussed here
given by is based on the PRIMA technique [84]. Recall from Sec-
tion VIII-D, where circuit reduction was achieved by
applying Arnoldi on the set of equations represented by
.. ..
. . (168) (162) (we refer this approach as classical Arnoldi). However,
this approach does not preserve the passivity of the reduced

722 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

Fig. 29. Example circuit for MNA formulation compatible with passive reduction algorithms.

system. Instead, if the Arnoldi is applied on the original 3) MNA Formulation Compatible with Passive Reduction
set of circuit equations represented by (165), reducing the Algorithms: Consider the lumped circuit equations (162).
and directly, passivity can be preserved (with certain The matrices , can be formulated, such that
conditions, which is discussed later in this section). A brief
description of this process is given below.
Using (161), we can rewrite (165) in the Laplace-domain
as
(174)

where represent the MNA variables, represent the

(169)
conductance and susceptance matrices, with an important
note that the rows corresponding to current variables are
Premultiplying both sides MNA equations of (169) by , negated (such that the diagonal entries of matrix con-
we have tributed by inductor elements remain positive).
are the matrices containing the stamps of resistors, capac-
itors and inductors, respectively. Matrix
corresponds to the current variables in a KCL formulation.
Provided that the original network is composed of passive
(170) elements only, lead to symmetric nonnegative
definite matrices. With this MNA formulation, it can be
where proved that the resulting is also symmetric nonnegative
definite [84]. An illustration of the above formulation is
given at the bottom of the next page in (175) for the circuit
in Fig. 29.
(171)
IX. RELATED TOPICS AND FURTHER READING
The above type of transformations are known as congruence
In addition to the interconnection simulation algorithms
transformations. Using (170) and (171), the output equation
discussed in the previous sections, there are several related
in (169) can be rewritten as
topics that may be of interest to the readers of this paper.
Full-Wave Models: At further subnanosecond rise times,
(172) the line cross section dimensions become a significant frac-
tion of the wavelength and field components in the direction
From (172), it can be easily noted that the admittance matrix of propagation can no longer be neglected. Consequently,
of the reduced system is given by full-wave models that take into account all possible field
components and satisfy all boundary conditions are required
to accurately estimate high-frequency effects. However, cir-
(173) cuit simulation of full-wave models is highly involved. The
information that is obtained through a full-wave analysis is in
It can be proved that the reduced model, given by (173) terms of electromagnetic field parameters such as propaga-
is passive [84]. In addition, like the block Arnoldi algorithm tion constant, characteristic impedance, etc. However, a cir-
also preserves the first block moments of the original cuit simulator requires the information in terms of currents,
network). One of the conditions required to prove the pas- voltages, and circuit impedances. This demands a general-
sivity of the reduced model is that the original matrix must ized method to combine modal results into circuit simula-
be a symmetric and nonnegative definite matrix. This condi- tors in terms of a full-wave stamps. References [26]–[29],
tion can be satisfied with a slight change during the MNA [56], [74] provide solution techniques and moment genera-
formulation step, as follows. tion schemes for such cases.

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 723

 Measured Data: In practice, it may not be possible to quire high order to capture high-frequency effects such as
obtain accurate analytical models for interconnects because skin effect. This can make the transient simulation expen-
of the geometric inhomogeneity and associated disconti- sive as the reduced model becomes expensive to evaluate and
nuities. To handle such situations, modeling techniques post-processing of macromodels to further reduce the order
based on measured data have been proposed in the liter- may become necessary [94], [95].
ature [102]–[113]. In general, the behavior of high-speed
interconnects can easily be represented by measured fre- ACKNOWLEDGMENT
quency-dependent scattering parameters or time-domain The authors wish to thank and acknowledge the help and
terminal measurements. However, handling measured data contributions provided by the current and past graduate
in circuit simulation is a tedious and a computationally students of the Computer-Aided Engineering Group at
expensive process. References [102]–[113] address such Carleton University. Relevant information regarding this
cases. Proceedings paper can be accessed at http://www.doe.
EMI Subnetworks: Electrically long interconnects func- carleton.ca/~achar/ieee.html.
tion as spurious antennas to pick up emissions from other
nearby electronic systems. This makes susceptibility to REFERENCES
emissions a major concern to current system designers of
[1] H. B. Bakoglu, Circuits, Interconnections and Packaging for
high-frequency product. Hence, the availability of intercon- VLSI. Reading, MA: Addison-Wesley, 1990.
nect simulation tools including the effect of incident fields [2] H. W. Jhonson and M. Grahaml, High-Speed Digital De-
is becoming an important design requirement. References sign. Englewood Cliffs, NJ: Prentice-Hall, 1993.
[3] M. Nakhla and R. Achar, High-Speed Circuit and Interconnect Anal-
[119]–[139] provide analysis techniques for interconnects ysis. Multimedia Learning Course, OT: Omniz Global Knowledge
subjected to external EM interferences and also for radiation Corporation (www.omniz.com), 2001.
analysis of interconnects. [4] R. K. Poon, Computer Circuits Electrical Design. Englewood
Cliffs, NJ: Prentice-Hall, 1995.
Sensitivity Analysis: Sensitivity analysis involving large [5] W. W. M. Dai, Guest Editor, “Special issue on simulation, modeling,
interconnect subnetworks can be highly CPU intensive. and electrical design of high-speed and high-density interconnects,”
Model-reduction-based approaches provide an efficient IEEE Trans. Circuits Syst., vol. 39, pp. 857–982, Nov. 1992.
[6] M. Nakhla and A. Ushida, Guest Editors, “Special issue on modeling
means for this purpose [114]–[118]. and simulation of high-speed interconnects,” IEEE Trans. Circuits
Minimum Realization of Reduced-Order Models: It is ev- Syst., vol. 47, pp. 239–305, Apr. 2000.
ident from the discussions in Sections VII and VIII that the [7] M. Nakhla and R. Achar, “Interconnect modeling and simulation,” in
The VLSI Handbook. Boca Raton, FL: CRC Press, 2000, ch. XVII,
size of the reduced-order model increases with the increase in pp. 17.1–17.29.
the number of ports. In such cases, it may become essential to [8] E. Chiprout and M. Nakhla, Asymptotic Waveform Evaluation
realize the macromodels with a minimum possible number of and Moment Matching for Interconnect Analysis. Boston, MA:
Kluwer, 1993.
states, so as to achieve fast transient simulations [93]. Also, it [9] A. Deustsch, “Electrical characteristics of interconnections for high-
is observed that Krylov-subspace-based reduced models re- performance systems,” Proc. IEEE, vol. 86, pp. 315–355, Feb. 1998.

(175)
Port Currents

724 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

 [10] R. Goyal, “Managing signal integrity,” IEEE Spectr., pp. 54–62, Mar. [38] C. Moler and C. Van Loan, “Nineteen dubious ways to compute the
1994. exponential of a matrix,” SIAM Rev., pp. 801–836, Oct. 1978.
[11] J. B. Faria, Multiconductor Transmission Line Structures. New [39] A. Dounavis, X. Li, M. Nakhla, and R. Achar, “Passive closed-loop
York: Wiley, 1993. transmission line model for general purpose circuit simulators,”
[12] C. Paul, Analysis of Multiconductor Transmission Lines. New IEEE Trans. Microwave Theory Tech., vol. 47, pp. 2450–2459, Dec.
York: Wiley, 1994. 1999.
[13] , Introduction to Electromagnetic Compatibility. New York: [40] A. Dounavis, R. Achar, and M. Nakhla, “Efficient passive circuit
Wiley, 1992. models for distributed networks with frequency-dependent parame-
[14] K. C. Gupta and R. Grag, Microstrip Lines and Slotlines. Boston, ters,” IEEE Trans. Adv. Packag., vol. 23, pp. 382–392, Aug. 2000.
MA: Artech House, 1996. [41] F. H. Branin Jr., “Transient analysis of lossless transmission lines,”
[15] A. E. Ruehli, “Equivalent circuit models for three dimensional mul- Proc. IEEE, vol. 55, pp. 2012–2013, 1967.
ticonductor systems,” IEEE Trans. Microwave Theory Tech., vol. 22, [42] M. Cases and D. M. Quinn, “Transient response of uniform dis-
pp. 216–224, Mar. 1974. tributed RLC transmission lines,” IEEE Trans. Circuits Syst., vol.
[16] , “Inductance calculations in a complex integrated circuit envi- 27, pp. 200–980, 1980.
ronment,” IBM J. Res. Develop., pp. 470–481, Sept. 1972. [43] F. Y. Chang, “Transient analysis of lossless coupled transmission
[17] P. K. Wolff and A. E. Ruehli, “Inductance computations for com- lines in a nonhomogeneous medium,” IEEE Trans. Microwave
plex three dimensional geometries,” IEEE Trans. Circuits Syst., pp. Theory Tech., vol. MTT–18, pp. 616–626, Sept. 1970.
16–19, 1981. [44] , “The generalized method of characteristics for waveform re-
[18] A. E. Ruehli and P. A. Brennan, “Efficient capacitance calculations laxation analysis of lossy coupled transmission lines,” IEEE Trans.
for three dimensional multiconductor systems,” IEEE Trans. Mi- Microwave Theory Tech., vol. 37, pp. 2028–2038, Dec. 1989.
crowave Theory Tech., pp. 76–82, Feb. 1973. [45] , “Waveform relaxation analysis of nonuniform lossy trans-
[19] A. E. Ruehli, “Survey of computer-aided analysis of integrated mission lines characterized with frequency-dependent parameters,”
circuit interconnections,” IBM J. Res. Develop., pp. 626–639, Nov. IEEE Trans. Circuits Syst., vol. 38, pp. 1484–1500, Dec. 1991.
1979. [46] Q. Chu, Y. Lau, and F. Y. Chang, “Transient analysis of microwave
[20] A. E. Ruehli and H. Heeb, “Circuit models for three dimensional active circuits based on time-domain characteristic models,” IEEE
geometries including dielectrics,” IEEE Trans. Microwave Theory Trans. Microwave Theory Tech., pp. 1097–1104, Aug. 1998.
Tech., pp. 1507–1516, Mar. 1992. [47] A. R. Djordjevié, T. K. Sarkar, and R. F. Harrington, “Analysis of
[21] D. D. Ling and A. E. Ruehli, “Interconnection modeling,” in Ad- lossy transmission lines with arbitrary nonlinear terminal networks,”
vances in CAD for VLSI, 3—Part II, Circuit Analysis, Simulation IEEE Trans. Microwave Theory Tech., vol. 34, pp. 660–666, June
and Design. Amsterdam: Noth-Holland, 1987, pp. 211–291. 1986.
[22] J. Cullum, A. Ruehli, and T. Zhang, “A method of reduced-order [48] , “Time-domain response of multiconductor transmission
modeling and simulation of large interconnect circuits and its appli- lines,” Proc. IEEE, vol. 75, pp. 743–764, June 1987.
cation to PEEC models with retardation,” IEEE Trans. Circuits Syst., [49] H. Grabinski, “An algorithm for computing the signal propagation
pp. 261–273, Apr. 2000. on lossy VLSI interconnect systems in the time-domain,” Integr.
VLSI J., pp. 35–48, Oct. 1989.
[23] T. L. Quarles, “The SPICE3 implementation guide,” Univ. Cali-
[50] M. Nakhla, “Analysis of pulse propagation on high-speed VLSI
fornia, Berkeley, Tech. Rep., ERL-M89/44, 1989.
chips,” IEEE J. Solid-State Circuits, vol. 25, pp. 490–494, Apr.
[24] D. Gao, A. Yang, and S. Kang, “Modeling and simulation of inter-
connection delays and cross talks in high-speed integrated circuits,” 1990.
IEEE Trans. Circuits Syst., pp. 1–9, Jan. 1990. [51] R. Griffith and M. Nakhla, “Time-domain analysis of lossy mul-
[25] H. Hasegawa and S. Seki, “Analysis of interconnection delay on very ticonductor transmission lines,” IEEE Trans. Microwave Theory
high-speed LSI/VLSI chips using a microstrip line model,” IEEE Tech., vol. 38, Oct. 1990.
Trans. Electron Devices, pp. 1954–1960, Dec. 1984. [52] , “Mixed frequency/time domain analysis on nonlinear
[26] T. Itoh and R. Mittra, “Spectral domain approach for calculating circuits,” IEEE Trans. Computer-Aided Design, vol. 10, pp.
the dispersion characteristics of microstrip lines,” IEEE Trans. Mi- 1032–1043, Aug. 1992.
crowave Theory Tech., pp. 496–499, Feb. 1973. [53] E. C. Chang and S. M. Kang, “Computationally efficient simulation
[27] D. Mirshekar-Syahkal, Spectral Domain Method for Microwave In- of a lossy transmission line with skin effect by using numerical in-
tegrated Circuits: Joinery, 1990. version of Laplace transform,” IEEE Trans. Circuits Syst., vol. 39,
[28] R. H. Jansen, “Spectral domain approach for microwave integrated pp. 861–868, July 1992.
circuits,” IEEE Trans. Microwave Theory Tech., vol. MTT-33, pp. [54] D. Kuznetsov and J. E. Schutt-Aine, “Optimal transient simulation of
1043–1056, Feb. 1985. transmission lines,” IEEE Trans. Circuits Syst., vol. 43, pp. 110–121,
[29] R. Wang and O. Wing, “A circuit model of a system of VLSI inter- Feb. 1996.
connects for time response computation,” IEEE Trans. Microwave [55] W. T. Beyene and J. E. Schutt-Aine, “Accurate frequency-domain
Theory Tech., vol. 39, pp. 688–693, April 1991. modeling and efficient simulation of high-speed packaging intercon-
[30] A. R. Djordjevié and T. K. Sarkar, “Closed-form formulas for nects,” IEEE Trans. Microwave Theory Tech., pp. 1941–1947, Oct.
frequency-dependent resistance and inductance per unit length of 1997.
microstrip and strip transmission lines,” IEEE Trans. Microwave [56] R. Wang and O. Wing, “Transient analysis of dispersive VLSI in-
Theory Tech., vol. 42, pp. 241–248, Feb. 1994. terconnects terminated in nonlinear loads,” IEEE Trans. Computer-
[31] A. R. Djordjevié, R. F. Harrington, T. K. Sarkar, and M. Bazdar, Ma- Aided Design, vol. 11, pp. 1258–1277, Oct. 1992.
trix Parameters for Multiconductor Transmission Lines: Software [57] W. C. Elmore, “The transient response of damped linear networks
and Users Manual. Boston, MA: Reteach House, 1989. with particular regard to wide-band amplifiers,” J. Appl. Phys., pp.
[32] A. Deutch et al., “Modeling and characterization of long on-chip in- 55–63, Jan. 1948.
terconnections for high performance microprocessors,” IBM J. Res. [58] J. Rubinstein, P. Penfield, and M. Horowitz, “Signal delay in RC
Develop., vol. 39, pp. 547–567, Sept. 1995. trees,” IEEE Trans. Computer-Aided Design, pp. 202–211, July
[33] A. Deutch et al., “High-speed signal propagation on lossy transmis- 1983.
sion lines,” IBM J. Res. Develop., vol. 39, pp. 601–615, July 1990. [59] L. T. Pillage and R. A. Rohrer, “Asymptotic waveform evaluation
[34] A. Deutch et al., “When are transmission-line effects important for for timing analysis,” IEEE Trans. Computer-Aided Design, vol. 9,
on-chip interconnections,” IEEE Trans. Microwave Theory Tech., pp. 352–366, Apr. 1990.
Oct. 1997. [60] G. A. Baker Jr., Essential of Padé Approximants. New York: Aca-
[35] J. Poltz, “Optimizing VLSI interconnect model for SPICE simula- demic, 1975.
tion,” J. Analog Integr. Circuits Signal Process., vol. 5, Jan. 1994. [61] J. H. McCabe, “A formal extension of the Padé table to include two
[36] T. Dhane and D. D. Zutter, “Selection of lumped element models point Padé quotients,” J. Inst. Math. Applicat., vol. 15, pp. 363–372,
for coupled lossy transmission lines,” IEEE Trans. Computer-Aided 1975.
Design, vol. 11, pp. 959–967, July 1992. [62] S. Kumashiro, R. A. Rohrer, and A. J. Strojwas, “Asymptotic
[37] M. Celik, A. C. Cangellaris, and A. Yaghmour, “An all purpose trans- waveform evaluation for transient analysis of 3-D interconnect
mission line model for interconnect simulation in SPICE,” IEEE structures,” IEEE Trans. Computer-Aided Design, vol. 12, no. 7,
Trans. Microwave Theory Tech., pp. 127–138, Oct. 1997. pp. 988–996, 1993.

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 725

 [63] T. Tang and M. Nakhla, “Analysis of high-speed VLSI intercon- [87] A. Dounavis, E. Gad, R. Achar, and M. Nakhla, “Passive
nect using asymptotic waveform evaluation technique,” IEEE Trans. model-reduction of multiport distributed networks including fre-
Computer-Aided Design, pp. 2107–2116, Mar. 1992. quency-dependent parameters,” IEEE Trans. Microwave Theory
[64] D. Xie and M. Nakhla, “Delay and crosstalk simulation of high speed Tech., pp. 2325–2334, Dec. 2000.
VLSI interconnects with nonlinear terminations,” IEEE Trans. Com- [88] M. Raugi, “Wavelet transform solution of MTL transients,” IEEE
puter-Aided Design, pp. 1798–1811, Nov. 1993. Trans. Magn., pt. I, pp. 1554–1557, May 1999.
[65] S. Lin and E. S. Kuh, “Transient simulation of lossy interconnects [89] S. Barmada and M. Raugi, “Transient numerical Solutions of
based on the recursive convolution formulation,” IEEE Trans. Cir- nonuniform MTL equations with nonlinear loads with wavelet
cuits Syst., vol. 39, pp. 879–892, Nov. 1992. expansion in time or space-domain,” IEEE Trans. Circuits Syst. I,
[66] S. Y. Kim, N. Gopal, and L. T. Pillage, “Time-domain macromodels pp. 1178–1190, Aug. 2000.
for VLSI interconnect analysis,” IEEE Trans. Computer-Aided De- [90] A. C. Cangellaris, S. Pasha, J. L. Prince, and M. Celik, “A new
sign, vol. 13, pp. 1257–1270, Oct. 1994. discrete time-domain model for passive model order reduction
[67] Q. Yu and E. S. Kuh, “Exact moment-matching model of transmis- and macromodeling of high-speed interconnections,” IEEE Trans.
sion lines and application to interconnect delay estimation,” IEEE Comp., Packag., Manufact. Technol., pp. 356–364, Aug. 1999.
Trans. VLSI Syst., pp. 311–322, Jun. 1995. [91] M. Celik and A. C. Cangellaris, “Simulation of dispersive multi-
[68] V. Raghavan, J. E. Bracken, and R. A. Rohrer, “AWESpice: A general conductor transmission lines by Padé approximation via Lanczos
tool for accurate and efficient simulation of interconnect problems,” process,” IEEE Trans. Microwave Theory Tech., pp. 2525–2535,
in Proc. ACM/IEEE Design Automation Conf., June 1992, pp. 87–92. Dec. 1996.
[69] J. E. Bracken, V. Raghavan, and R. A. Rohrer, “Interconnect simu- [92] P. Gunupudi, M. Nakhla, and R. Achar, “Simulation of high-speed
lation with asymptotic waveform evaluation (AWE),” IEEE Trans. distributed interconnects using Krylov-subspace techniques,” IEEE
Circuits Syst., pp. 869–878, Nov. 1992. Trans. Computer-Aided Design, vol. 19, pp. 799–808, July 2000.
[70] R. Griffith, E. Chiprout, Q. J. Zhang, and M. Nakhla, “A CAD frame-
[93] R. Achar and M. Nakhla, Minimum Realization of Reduced-Order
work for simulation and optimization of high-speed VLSI intercon-
Models of High-Speed Interconnect Macromodels. Boston, MA:
nections,” IEEE Trans. Circuits Syst., vol. 39, pp. 893–906, Nov. 1992.
Kluwer, 1998, ch. Signal Propagation on Interconnects, pp. 23–44.
[71] T. Tang, M. Nakhla, and R. Griffith, “Analysis of lossy multicon-
[94] M. Kamon, F. Wang, and J. White, “Generating nearly optimally
ductor transmission lines using the asymptotic waveform evaluation
technique,” IEEE Trans. Microwave Theory Tech., vol. 39, Dec. 1991. compact models from Krylov-subspace based reduced-order
[72] E. Chiprout and M. Nakhla, “Analysis of interconnect networks models,” IEEE Trans. Circuits Syst., pp. 239–248, Apr. 2000.
using complex frequency hopping,” IEEE Trans. Computer-Aided [95] R. Khazaka and M. Nakhla, “Multilevel passive order reduction of
Design, vol. 14, pp. 186–199, Feb. 1995. interconnects,” in IMS Dig., May 2001, to be published.
[73] R. Sanaie, E. Chiprout, M. Nakhla, and Q. J. Zhang, “A fast method [96] O. Palusinski and A. Lee, “Analysis of transients in nonuniform
for frequency and time domain simulation of high-speed VLSI and uniform multiconductor transmission lines,” IEEE Trans. Mi-
interconnects,” IEEE Trans. Microwave Theory Tech., vol. 42, pp. crowave Theory Tech., pp. 127–138, Jan. 1989.
2562–2571, Dec. 1994. [97] S. Manney, M. S. Nakhla, and Q. J. Zhang, “Time domain analysis
[74] R. Achar, M. Nakhla, and Q. J. Zhang, “Full-wave analysis of of nonuniform frequency dependent high-speed interconnects,” in
high-speed interconnects using complex frequency hopping,” IEEE Proc. IEEE Int. Conf. Computer-Aided Design, 1992, pp. 449–453.
Trans. Computer-Aided Design, pp. 997–1016, Oct. 1998. [98] N. Boulejfen, A. Kouki, and F. Ghannouchi, “Frequency and time-
[75] M. A. Kolbehdari, M. Srinivasan, M. Nakhla, Q. J. Zhang, and R. domain analysis nonuniform lossy coupled transmission lines with
Achar, “Simultaneous time and frequency domain solution of EM linear and nonlinear terminations,” IEEE Trans. Microwave Theory
problems using finite element and CFH techniques,” IEEE Trans. Tech., pp. 367–379, Mar. 2000.
Microwave Theory Tech., vol. 44, pp. 1526–1534, Sept. 1996. [99] C. Yen, Z. Fazarinc, and R. L. Wheeler, “Time-domain skin-effect
[76] R. Achar, M. Nakhla, P. Gunupudi, and E. Chiprout, “Passive in- model for transient analysis of lossy transmission lines,” Proc. IEEE,
terconnect reduction algorithm for distributed/measured networks,” pp. 750–757, 1982.
IEEE Trans. Circuits Syst. II, pp. 287–301, Apr. 2000. [100] R. Khazaka, E. Chiprout, M. Nakhla, and Q. J. Zhang, “Analysis of
[77] P. Feldmann and R. W. Freund, “Efficient linear circuit analysis by high-speed interconnects with frequency dependent parameters,” in
Padé via Lanczos process,” IEEE Trans. Computer-Aided Design, Proc. Int. Symp. EMC, Zurich, Switzerland, Mar. 1995, pp. 203–208.
vol. 14, pp. 639–649, May 1995. [101] W. Sui, D. A. Christensen, and C. H. Durney, “Extending the two-di-
[78] R. W. Freund, “Reduced-order modeling techniques based on Krylov mensional FDTD method to hybrid electromagnetic systems with ac-
subspace and their use in circuit simulation,” in Tech. Memo.: Lucent tive and passive lumped elements,” IEEE Trans. Microwave Theory
Technologies, 1998. Tech., vol. 40, pp. 724–730, Apr. 1992.
[79] P. Feldmann and R. W. Freund, “Reduced order modeling of large [102] Y. Tsuei, A. C. Cangellaris, and J. L. Prince, “Rigorous electromag-
linear subcircuits via a block Lanczos algorithm,” in Proc. Design netic modeling of chip-to-package (first-level) interconnections,”
Automation Conf., June 1995, pp. 474–479. IEEE Trans. Comp. Hybrids Manufact. Technol., vol. 16, pp.
[80] I. M. Elfadel and D. D. Ling, “A block rational Arnoldi algorithm 876–883, Aug. 1993.
for multiport passive model-order reduction of multiport RLC net- [103] M. Picket-May, A. Taflove, and J. Baron, “FD-TD modeling of digital
works,” in Proc. ICCAD-97, Nov. 1997, pp. 66–71. signal propagation in 3-D circuits with passive and active loads,” IEEE
[81] L. M. Silviera, M. Kamen, I. Elfadel, and J. White, “A coordinate Trans. Microwave Theory Tech., vol. 42, pp. 1514–1523, Aug. 1994.
transformed Arnoldi algorithm for generating guaranteed stable re- [104] P. C. Cherry and M. F. Iskander, “FDTD analysis of high frequency
duced-order models for RLC circuits,” in Tech. Dig. ICCAD, Nov. electronic interconnection effects,” IEEE Trans. Microwave Theory
1996, pp. 2288–2294. Tech., vol. 43, pp. 2445–2451, Oct. 1995.
[82] L. M. Silviera, M. Kamen, and J. White, “Efficient reduced-order [105] S. D. Corey and A. T. Yang, “Interconnect characterization using
modeling of frequency-dependent coupling inductances associated time-domain reflectometry,” IEEE Trans. Microwave Theory Tech.,
with 3-D interconnect structures,” IEEE Trans. Comp., Packag., vol. 43, pp. 2151–2156, Sept. 1995.
Manufact. Technol. A, pp. 283–288, May 1996. [106] B. J. Cooke, J. L. Prince, and A. C. Cangellaris, “S -parameter
[83] K. J. Kerns and A. T. Yang, “Preservation of passivity during RLC analysis of multiconductor integrated circuit interconnect systems,”
network reduction via split congruence transformations,” IEEE IEEE Trans. Computer-Aided Design, pp. 353–360, Mar. 1992.
Trans. Computer-Aided Design, pp. 582–591, July 1998. [107] L. Vakanas, A. C. Cangellaris, and O. Palusinski, “Scattering param-
[84] A. Odabasioglu, M. Celik, and L. T. Pillage, “PRIMA: Passive re- eter based simulation of transients in lossy, nonlinearly terminated
duced-order interconnect macromodeling algorithm,” IEEE Trans. packaging interconnects,” IEEE Trans. Comp., Packag., Manufact.
Computer-Aided Design, pp. 645–654, Aug. 1998. Technol. B, pp. 472–479, Feb. 1994.
[85] A. Odabasioglu, M. Celik, and L. T. Pillage, “Practical considera- [108] M. Celik, A. C. Cangellaris, and A. Deutsch, “A new moment gener-
tions for passive reduction of RLC circuits,” in Proc. DAC, June ation technique for interconnects characterized by measured or cal-
1999, pp. 214–219. culated S-parameters,” in IEEE Int. Microwave Symp. Dig., June
[86] Q. Yu, J. M. L. Wang, and E. S. Kuh, “Passive multipoint moment- 1996, pp. 196–201.
matching model order reduction algorithm on multiport distributed [109] J. E. Schutt-Aine and R. Mittra, “Scattering parameter transient anal-
interconnect networks,” IEEE Trans. Circuits Syst. I, pp. 140–160, ysis of transmission lines loaded with nonlinear terminations,” IEEE
Jan. 1999. Trans. Microwave Theory Tech., pp. 529–536, 1988.

726 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

[110] , “Nonlinear analysis of coupled transmission lines,” IEEE [132] N. Ari and W. Blumer, “Analytic formulation of the response of a
Trans. Circuits Syst., vol. 36, pp. 959–967, 1989. two-wire transmission line excited by a plane wave,” IEEE Trans.
[111] W. T. Beyene and J. E. Schutt-Aine, “Efficient transient simulation Electromagn. Compat., vol. 30, pp. 437–448, Nov. 1988.
of high-speed interconnects characterized by sampled data,” IEEE [133] C. R. Paul, “Literal solutions for the time-domain response of a two-
Trans. Comp., Packag., Manufact. Technol. B, vol. 21, pp. 105–113, conductor transmission line excited by an incident electromagnetic
Feb. 1998. field,” IEEE Trans. Electromagn. Compat., vol. 37, May 1995.
[112] G. Zheng, Q. J. Zhang, M. Nakhla, and R. Achar, “An efficient ap- [134] P. Bernardi, R. Cicchetti, and C. Pirone, “Transient response of
proach for simulation of measured subnetworks with complex fre- a microstrip line circuit excited by an external electromagnetic
quency hopping,” in Proc. IEEE/ACM Int. Conf. Computer Aided source,” IEEE Trans. Electromagn. Compat., vol. 34, pp. 100–108,
Design, San Jose, CA, Nov. 1996, pp. 23–26. May 1992.
[113] R. Achar and M. Nakhla, “Efficient transient simulation of em- [135] G. J. Burke, E. K. Miller, and S. Chakrabarti, “Using model based
bedded subnetworks characterized by S -parameters in the presence parameter estimation to increase the efficiency of computing elec-
of nonlinear elements,” IEEE Trans. Microwave Theory Tech., vol. tromagnetic transfer functions,” IEEE Trans. Magn., vol. 25, pp.
46, pp. 2356–2363, Dec. 1998. 2087–2089, July 1989.
[114] N. Liu, M. Nakhla, and Q. J. Zhang, “Time domain sensitivity of [136] F. M. Tesche, M. V. Ianoz, and T. Karlsson, EMC Analysis Methods
high-speed VLSI interconnects,” Int. J. Circuit Theory Applicat., and Computational Models. New York: Wiley, 1997.
vol. 22, pp. 479–511, Nov. 1994. [137] A. K. Agrawal, H. J. Price, and S. H. Gurbaxani, “Transient response
[115] Q. J. Zhang, S. Lum, and M. Nakhla, “Minimization of delay of multiconductor transmission lines excited by a nonuniform elec-
and crosstalk in high-speed VLSI interconnects,” IEEE Trans. tromagnetic field,” IEEE Trans. Electromagn. Compat., vol. 22, pp.
Microwave Theory Tech., vol. 42, pp. 1555–1563, July 1992. 119–129, May 1980.
[116] S. Lum, M. Nakhla, and Q. J. Zhang, “Sensitivity analysis of lossy [138] Y. Kami and R. Sato, “Circuit-concept approach to externally excited
coupled transmission lines with nonlinear terminations,” IEEE transmission lines,” IEEE Trans. Electromagn. Compat., vol. 27, pp.
Trans. Microwave Theory Tech., vol. 42, 1994. 177–183, Nov. 1985.
[117] R. W. Freund and P. Feldmann, “Small signal circuit analysis and [139] R. Raut, W. J. Steenart, and G. Costache, “A note on the optimum
sensitivity computations with PVL algorithm,” IEEE Trans. Circuits layout of electronic circuits to minimize electromagnetic field
Syst. II, pp. 577–585, 1996. strength,” IEEE Trans. Electromagn. Compat., pp. 88–89, Feb. 1988.
[118] C. Jiao, A. C. Cangellaris, A. Yaghmour, and J. L. Prince, “Sensi-
[140] C. W. Ho, A. E. Ruehli, and P. A. Brennan, “The modified nodal
tivity analysis of multiconductor transmission lines and optimization
approach to network analysis,” IEEE Trans. Circuits Syst., vol.
for high-speed interconnect design,” IEEE Trans. Comp., Packag.,
CAS-22, pp. 504–509, June 1975.
Manufact. Technol., pp. 132–141, May 2000.
[141] A. E. Ruehli, Circuit Analysis, Simulation and Design. New York:
[119] R. Khazaka and M. Nakhla, “Analysis of high-speed interconnects
Noth-Holland, 1988.
in the presence of electromagnetic interference,” IEEE Trans. Mi-
[142] J. K. White and A. S. Vincentelli, Relaxation Techniques for the Sim-
crowave Theory Tech., vol. 46, pp. 940–947, July 1998.
ulation of VLSI Circuits. Boston, MA: Kluwer, 1990.
[120] I. Erdin, R. Khazaka, and M. Nakhla, “Simulation of high-speed
[143] D. O. Pederson, “A historical review of circuit simulation,” IEEE
interconnects in the presence of incident field,” IEEE Trans. Mi-
crowave Theory Tech., vol. 46, pp. 2251–2257, Dec. 1998. Trans. Circuits Syst., vol. CAS-31, no. 1, Jan. 1984.
[121] I. Ierdin, M. Nakhla, and R. Achar, “Circuit analysis of electro- [144] A. S. Vincentelli, “Circuit Simulation,” in Computer Design Aids for
magnetic radiations and field coupling effects for networks with VLSI Circuits, P. Antognetti, D. O. Pederson, and H. De Man, Eds:
embedded full-wave modules,” IEEE Trans. Electromagn. Compat., Martinus Nijhoff, 1986, pp. 19–122.
vol. 42, pp. 449–460, Nov. 2000. [145] J. K. Ousterhout, “CRYSTAL: A timing analyzer for NMOS VLSI
[122] F. Olyslager, D. D. Zutter, and A. T. de Hoop, “New reciprocal circuit circuits,” in Proc. 3rd Caltech. Conf. VLSI, Mar. 1983, pp. 57–69.
model for lossy waveguide structures based on the orthogonality of [146] N. P. Jouppi, “Timing analysis and performance improvement of
the eigenmodes,” IEEE Trans. Microwave Theory Tech., vol. 42, no. MOS VLSI design,” IEEE Trans. Computer-Aided Design, vol. 6,
12, pp. 2261–2269, Dec. 1994. pp. 650–665, July 1987.
[123] F. Olyslager, D. D. Zutter, and K. Blomme, “Rigorous analysis of [147] S. Lin, M. M. Sadowska, and E. S. Kuh, “SWEC: A step wise equiv-
the propagation characteristics of general lossless and lossy multi- alent conductance timing simulator for CMOS VLSI circuits,” in
conductor transmission lines in multi-layered media,” IEEE Trans. Proc. Electron. Design Automation Conf., 1991, pp. 142–148.
Microwave Theory Tech., vol. 41, pp. 79–88, Jan. 1993. [148] A. S. Vincetelli, E. Lelarasmee, and A. Ruehli, “The waveform relax-
[124] C. D. Taylor, R. S. Satterwhite, and C. W. Harrison, “The response ation method for the time-domain analysis of large scale integrated
of a terminated two-wire transmission line excited by a nonuniform circuits,” IEEE Trans. Computer-Aided Design, vol. 1, pp. 131–145,
electromagnetic field,” IEEE Trans. Antennas Propagat., vol. 13, pp. Aug. 1982.
987–989, Nov. 1965. [149] A. Devgan and R. A. Roher, “Adaptively controlled explicit simula-
[125] A. A. Smith, “A more convenient form of the equations for the re- tion,” IEEE Trans. Computer-Aided Design, vol. 13, Jun. 1994.
sponse of a transmission line excited by nonuniform fields,” IEEE [150] S. Lele, “Compact finite difference schemes with spectral-like reso-
Trans. Electromagn. Compat., vol. 15, pp. 151–152, Aug. 1973. lution,” J. Comput. Phys., vol. 103, pp. 16–42, 1992.
[126] C. R. Paul, “Frequency response of multiconductor transmission [151] J. Vlach and K. Singhal, Computer Methods for Circuit Analysis and
lines illuminated by an incident electromagnetic field,” IEEE Trans. Design. New York: Van Nostrand, 1983.
Microwave Theory Tech., vol. 22, pp. 454–457, Apr. 1976. [152] T. Kailath, Linear Systems. Toronto: Prentice-Hall, 1980.
[127] , “A comparison of the contributions of common-mode and dif- [153] C. T. Chen, Linear System Theory and Design. New York: Holt,
ferential-mode currents in radiated emissions,” IEEE Trans. Electro- Rinehart and Winston, 1984.
magn. Compat., pp. 189–193, May 1989. [154] L. Weinberg, Network Analysis and Synthesis. New York: Mc-
[128] , “A SPICE model for multiconductor transmission lines ex- Graw-Hill, 1962.
cited by an incident electromagnetic field,” IEEE Trans. Electro- [155] W. Louis, Network Analysis and Synthesis. New York: McGraw-
magn. Compat., pp. 342–354, Nov. 1994. Hill, 1962.
[129] I. Wuyts and D. De Zutter, “Circuit model for plane-wave incidence [156] E. Kuh and R. Rohrer, Theory of Active Linear Networks. San
on multiconductor transmission lines,” IEEE Trans. Electromagn. Francisco, CA: Holden-Day, 1967.
Compat., vol. 36, pp. 206–212, Aug. 1994. [157] E. A. Guillemin, Synthesis of Passive Networks. New York: Wiley,
[130] E. S. M. Mok and G. I. Costache, “Skin-effect considerations on tran- 1957.
sient response of a transmission line excited by an electromagnetic [158] M. E. V. Valkenburg, Introduction to Modern Network Syn-
wave,” IEEE Trans. Electromagn. Compat., vol. 34, pp. 320–329, thesis. New York: Wiley, 1960.
Aug. 1992. [159] U. S. Pillai, Spectrum Estimation and System Identification. New
[131] Y. Kami and R. Sato, “Circuit-concept approach to externally excited York: Springer-Verlag, 1993.
transmission lines,” IEEE Trans. Electromagn. Compat., vol. 27, pp. [160] J. W. Demmel, Applied Numerical Linear Algebra. Philadelphia,
177–183, Nov. 1985. PA: SIAM, 1997.

ACHAR AND NAKHLA: SIMULATION OF HIGH-SPEED INTERCONNECTS 727

 Ramachandra Achar (Member, IEEE) received Michel S. Nakhla (Fellow, IEEE) received the
the B.Eng. degree in electronics engineering M.A.Sc. and Ph.D. degrees in electrical engi-
from Bangalore University, India, in 1990, the neering from University of Waterloo, Ontario,
M.Eng. degree in microelectronics from Birla In- Canada in 1973 and 1975, respectively.
stitute of Technology and Science, Pilani, India, He is Professor of Electrical Engineering
in 1992, and the Ph.D. degree from Carleton at Carleton University, Ottawa, ON, Canada.
University, Ottawa, ON, Canada, in 1998, where From 1976 to 1988, he was with Bell-Northern
he received the University Medal for his doctoral Research, Ottawa, as the Senior Manager of the
work on high-speed VLSI interconnect analysis. Computer-Aided Engineering Group. In 1988,
He currently serves as an Assistant Professor he joined Carleton University as a Professor and
in the Department of Electronics, Carleton Uni- the holder of the Computer-Aided Engineering
versity. He spent the summer of 1995 working on high-speed interconnect Senior Industrial Chair established by Nortel Networks and the Natural
analysis at T. J. Watson Research Center, IBM, New York. He was a Grad- Sciences and Engineering Research Council of Canada. He is the founder
uate Trainee at Central Electronics Engineering Research Institute, Pilani, of the High-Speed CAD Research Group at Carleton University and is
India, during 1992 and was also previously employed at Larsen and Toubro a frequent invited speaker on the topic of high-speed interconnects. His
Engineers Ltd., Mysore, India, and at Indian Institute of Science, Bangalore, research interests include CAD of VLSI and microwave circuits, modeling
India, as an R&D Engineer. During 1998–2000, he served as a Research En- and simulation of high-speed interconnects, nonlinear circuits, multidisci-
gineer in the CAE Group at Carleton University. His research interests in- plinary optimization, thermal and electromagnetic (EM) emission analysis,
clude modeling and simulation of high-speed interconnects, numerical algo- MEMS, and neural networks. He serves as Technical Consultant for several
rithms, and development of computer-aided design tools for high-frequency industrial organizations and is the principal investigator for several major
circuit analysis. sponsored research projects.
Dr. Achar received several prestigious awards, including NSERC
(Natural Science and Engineering Research Council) Doctoral Award
(2000), Strategic Microelectronics Corporation (SMC) Award (1997),
Canadian Microelectronics Corporation (CMC) Award (1996), and the Best
Student Paper Award in the 1998 Micronet (a Canadian network of centers
of excellence on Microelectronics) annual workshop.

728 PROCEEDINGS OF THE IEEE, VOL. 89, NO. 5, MAY 2001

View publication stats