Professional Documents
Culture Documents
Abstract— Equivalent circuit synthesis of an interconnect In the convolution-based methods, the impulse responses of the
macromodel is commonly used as an intermediate step before network are computed via inverse Fourier transform, and being
conducting transient simulation of the channel. Depending on convolved with the excitation signals at the ports to obtain the
the type, for example, S, Y , and Z parameters, and form, for
example, pole–residue, state-space, of the macromodel, various corresponding output waveforms. This kind of methods is best
circuit topologies have been proposed in the literature. In this suited for channels where the impulse responses decay quickly,
article, we focus on the synthesis of multiport S parameters in since the computational overhead of convolution scales as
the pole–residue form. A topology widely used in commercial O(n 2 ), where n is the number of nonzero points in the
software is to convert the pole–residue model into a sparse discretized impulse response. In the macromodeling-based
state-space model and then realize the state equations and output
equations using capacitors, resistors, and controlled sources. methods, the channel S parameters are first fit by a rational
Another topology, based on the generalized pi-model, has also function, which could be in polynomial-ratio form [2], pole–
been proposed. In this article, we show that although these residue form [3], state-space description [4], and so on. Such
conventional methods work well for the multi-input multi-output rational representation is then imported into circuit simulators
(MIMO) case (all Si j share a common set of poles), they are in various ways, to be discussed immediately. The advantage
not optimal in terms of circuit complexity for the multi- single-
input-single-output (SISO) case (each Si j has its own distinct of macromodeling is that the rational representation enables
poles). A new topology is proposed accordingly, which has smaller the network to be simulated via the highly efficient recursive
complexity than the conventional ones. Transient simulation convolution [5]. Another favorable feature of macromodeling
on four different circuit solvers is performed to compare the is that, during or after the fitting process, there exist various
efficiencies of the various topologies. It is found that the proposed systematic algorithms to assess and enforce the stability,
topology can achieve more than 1.3 times speedup against the
conventional topologies (except the one with explicit pole/residue causality, and passivity of the rational model [6], which are
specification) on most simulators, in terms of the normalized time important properties for a physically meaningful channel.
per step. Among the various macromodeling methods, the vector fitting
Index Terms— Circuit synthesis, equivalent circuit, macromod- (VF) [3] is known for its robustness and high efficiency [7],
eling, scattering parameters, vector fitting (VF). and is one of the most prominent algorithms in rational
fitting [6]. The raw model delivered by VF is in pole–residue
I. I NTRODUCTION
form, while it can be converted to a diagonal state-space form
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
1972 IEEE TRANSACTIONS ON COMPONENTS, PACKAGING AND MANUFACTURING TECHNOLOGY, VOL. 11, NO. 11, NOVEMBER 2021
Fig. 1. Synthesis of one-port Y parameter [14], for the case of (a) one constant term, (b) one real pole, (c) a pair of complex conjugate poles, in the form of
RLCR, (d) again a pair of complex conjugate poles, but in the form of RLCG, where G represents a VCCS, and (e) N terms of pole–residue pairs, in which
the admittance blocks of each term are simply connected in parallel. In the above equations, the subscripts r and i denote the real and imaginary parts of the
quantity, respectively.
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
CHOU AND SCHUTT-AINÉ: EQUIVALENT CIRCUIT SYNTHESIS OF MULTIPORT S PARAMETERS IN POLE–RESIDUE FORM 1973
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
1974 IEEE TRANSACTIONS ON COMPONENTS, PACKAGING AND MANUFACTURING TECHNOLOGY, VOL. 11, NO. 11, NOVEMBER 2021
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
CHOU AND SCHUTT-AINÉ: EQUIVALENT CIRCUIT SYNTHESIS OF MULTIPORT S PARAMETERS IN POLE–RESIDUE FORM 1975
TABLE I
C OMPARISON OF C IRCUIT C OMPLEXITY OF D IFFERENT M ETHODS
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
1976 IEEE TRANSACTIONS ON COMPONENTS, PACKAGING AND MANUFACTURING TECHNOLOGY, VOL. 11, NO. 11, NOVEMBER 2021
Fig. 8. Technique to make all RLC values in the RLCG form of Fig. 1 (d) positive. Because an additional VCVS (E element) is used, this topology is called
RLCGE.
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
CHOU AND SCHUTT-AINÉ: EQUIVALENT CIRCUIT SYNTHESIS OF MULTIPORT S PARAMETERS IN POLE–RESIDUE FORM 1977
TABLE II
C OMPARISON OF S IMULATION T IME FOR 100-ns T RANSIENT S IMULATION
To compare the transient simulation time, we use four circuit and LTspice [27]. All simulators are run in sequential mode
solvers: Ansys AEDT [11], Keysight ADS [10], Ngspice [26], (nonparallel). Because each engine has its own specific
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
1978 IEEE TRANSACTIONS ON COMPONENTS, PACKAGING AND MANUFACTURING TECHNOLOGY, VOL. 11, NO. 11, NOVEMBER 2021
implementation of the transient algorithm, the runtimes differ. definitely result in more time steps than one another. For
Through the comparison across different solvers, we can example, for the coupled_line (linear) case, the number of time
understand the efficiency of the circuit topologies more com- steps of PI-RLCR is the second smallest in LTspice (11 337)
prehensively. The results are listed in Table II. among the six topologies, but the largest in AEDT (22 999).
For the cases denoted by “linear termination,” a periodic As another example, in LTspice, the PI-RLCG has the smallest
pulse voltage source is applied to port 1, while all the other number of steps for the coupled_line case (11 208), while the
ports are terminated by resistors. For the cases denoted by largest for fr4_channel (49 158) (among the six topologies).
“CMOS,” the same pulse source is applied to a CMOS inverter In this regard, the total simulation time, due to its dependence
first which drives port 1 of the channel, and the receiving port on the number of steps, may be less indicative of the circuit
is terminated by an inverter as well. The simulation is run to complexity than the normalized time per step.
100 ns.
Typical transient solvers use variable instead of fixed time D. RLCGE Topology
step to achieve faster speed; when the waveform is relatively The proposed-RLCGE topology, trading complexity for
smooth, the solver may try to increase the time step for positive circuit elements, is roughly 1.3 times slower than the
acceleration. Consequently, the total number of steps required proposed RLCG in terms of normalized time per step. The
to march to 100 ns differs. In Table II, we record not only total simulation time, likewise, is more difficult to quantify
the total simulation time, but also the number of time steps, due to the solver-dependent time step control.
and the normalized time per step for comparison. For AEDT,
Ngspice, and LTspice, we set the maximum time step (delmax)
E. POLE-Element
to 10 ps, while the actual time step is typically around 4–9 ps.
ADS, on the other hand, offers the option of fixed time step, The POLE-element, which directly specifies the pole/zero
and we set it as 1 ps. All simulations are performed on the values in the netlist, runs fastest on ADS, and the normalized
same computer. time per step can be ten times smaller than other circuit
We note that, primarily, the numbers in Table II should synthesis approaches. This is to be expected, however, since
be compared horizontally, because they correspond to the explicit pole/zero specification allows the solver to use the
same macromodel and same solver settings, and thus the highly efficient recursive convolution. AEDT, on the other
difference in runtime solely reflects the effects of the circuit hand, is less efficient in handling the POLE-element circuit
topologies. Different solvers, on the other hand, have different and is even slightly slower than the proposed-RLCR synthesis
transient algorithms, which affect the timing results signifi- approach in terms of normalized runtime.
cantly. We summarize the observations from Table II into the In summary, from the results in Table II, we could say
following points. that the proposed topology can indeed reduce the circuit
complexity and thus the transient simulation time. Four
A. Normalized Time Per Step different solvers were used to compare the transient time,
For AEDT, ADS, and LTspice, the normalized time per and it was observed that either the proposed-RLCR or
step of the proposed topologies (RLCR and RLCG) are both proposed-RLCG topology has the smallest runtime per step on
smaller than the conventional topologies (not including the most simulators. In particular, when compared with the widely
POLE element). The speedup factors against the PI topologies adopted state-space topology, the proposed-RLCG form can
are at least 1.5, sometimes much larger. The speedup factors achieve 1.3–2 times speedup. The only exception is Ngspice,
against the state-space topology ranges from 1.3 to 2. on which the proposed topologies are roughly the same
For Ngspice, the proposed topologies have nearly the same or slightly slower than the state-space topology. If negative
runtime as the state-space topology for the coupled_line case, circuit elements are not permitted, then the proposed-RLCGE
while for the fr4_channel case, the proposed topologies are can be used, which is roughly 1.3 times slower than the
slower than the state-space by roughly 1.2 times. proposed-RLCG topology.
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.
CHOU AND SCHUTT-AINÉ: EQUIVALENT CIRCUIT SYNTHESIS OF MULTIPORT S PARAMETERS IN POLE–RESIDUE FORM 1979
Authorized licensed use limited to: Hochschule Osnabruck. Downloaded on December 21,2023 at 22:12:58 UTC from IEEE Xplore. Restrictions apply.