386
IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 12, NO. 4, NOVEMBER 1999
Constraint Transformation for IC Physical Design
Enrico Malavasi, Member, IEEE, and Edoardo Charbon, Member, IEEE
Abstract— In a top-down design methodology, design tasks
are divided into simpler subtasks across levels of a hierarchy
as an effective divide-and-conquer technique. For every task,
tolerances are defined on all performance characteristics to take
into account parasitics, mismatches, and other nondeterministic
process parameter variations. Constraint transformation is a
process used to translate performance specifications into subtask
requirements. This paper introduces the problem of constraint
transformation and describes some formal solutions for analog
circuit applications. Examples illustrate the methodology and
show the suitability of this approach in industrial-strength applications.
Index Terms— Constraints, nondeterministic transformations,
tolerance-aware models, top-down design, sensitivity.
I. INTRODUCTION
T
HE progressive consumerization of the integrated circuit (IC) market leaves less and less time to develop
reliable components and systems to implement the growing
demand for functionality. To sustain time-to-market pressure,
it is necessary to stress design methodology and to facilitate
correct-by-construction paradigms.
In a top-down design methodology [1], the design flow
proceeds through levels of abstraction. At each level, design
tasks are divided into independent subtasks for which constraints are generated. If all subtasks are executed correctly
and their constraints are met, the original task will also be
satisfactorily completed. Given a set of constraints for a design
task, optimal generation of constraints for all subtasks is a
complex problem, usually referred to as “constraint transformation.” By repeatedly solving task partitioning and constraint
transformation at each level of abstraction, the problem of
designing a complex circuit is transformed into a set of
independent simpler tasks with well defined specifications.
Such a constraint-driven design environment is robust, since
it allows to identify constraint violations as early as possible
in the design process [2], [3].
In the IC fabrication process, all design parameters are
subject to deviations from their nominal values. Both absolute
values and mismatch of parasitics play a role in the deviation
of performance functions from their nominal behavior. The
objective of constraint transformation is to generate bounds
on the variations of all physical design parameters such that
performance degradation is maintained within specifications.
Constraint transformation is solved with a two-step procedure.
Manuscript received August 31, 1998; revised July 28, 1999.
The authors are with Cadence Design Systems, Inc., San Jose, CA 95134
USA.
Publisher Item Identifier S 0894-6507(99)09247-7.
i) Model generation: a model is created relating performance measures to the design parameters.
ii) Constraint generation: actual bounds are generated on
all design parameters.
As a simple example, consider a delay constraint
on a net. During model generation, we model
, where
and
are the lumped resistance and capacitance of the
net. During the constraint generation phase, we determine max
such that
.
bounds for and
Another example is the use of performance sensitivities,
which provide a powerful model generation technique whenever a circuit behavior can be linearized around a nominal
point. Constraint generation then can be obtained through the
solution of a linear system, usually aimed at some optimization
objective [2].
In this paper, we will describe the constraint transformation
problem as it has been formulated for the specific case of
physical design for IC’s. In this field, research has focused on:
1) linearized models built using sensitivity analysis [4], [5];
2) optimization techniques for bound generation, based on
some model of the “flexibility” [2], [5] of the layout
tools, i.e., a measure of how easy it would be for a
design tool to enforce a given set of bounds.
Then, constraint-driven design tools [2], [3] enforce bounds on
low-level design parameters, ignoring the original performance
constraints used to generate those bounds [6].
In this formulation, the criticality of parasitics is quantified
based on the cumulative effect to performance, and it is used
to drastically reduce the number of specifications on each
interconnect realization. The deterministic nature of parasitics
is postulated in these approaches as a necessary condition
to yield meaningful bounds. Nondeterministic effects are due
mainly to mask misalignment, technology gradients in the
fabrication process, and substrate noise. Both deterministic and
stochastic design parameters must be taken into account in this
approach. Their models can be constructed in a similar fashion,
using appropriate statistical modeling for the nondeterministic
parameters.
The paper is organized as follows. The constraint-driven
physical design paradigm is outlined in Section II. The
generalized constraint generation problem is formalized in
Sections III and IV for deterministic and nondeterministic
design parameters, respectively. Nondeterministic substraterelated effects are covered in Section V, where constraint
generation strategies are also provided. The constraint
generation engine is described in Section VI. Several examples
illustrating the approach are presented in Section VII.
0894–6507/99$10.00 1999 IEEE
MALAVASI AND CHARBON: CONSTRAINT TRANSFORMATION FOR IC PHYSICAL DESIGN
387
Fig. 1. Constraint-driven physical design flow.
II. THE CONSTRAINT-DRIVEN DESIGN PARADIGM
In a general semiautomated constraint-driven physical design flow, three main components can be recognized: a translation engine, a set of constraint-driven physical assembly tools,
and an extraction or verification utility. Fig. 1 shows a general
constraint-driven physical design flow.
The following are the operations needed to support such a
flow.
1) Define performance specifications. They can be either
determined by the user or propagated from a higher
level of hierarchy.
2) Translate specifications onto a set of constraints for
physical design. The semantic of these constraints is
compatible with and can be successfully implemented
by the physical layout tool set.
3) Detect overconstraints.
4) Manually edit constraints, if needed.
5) Enforce constraints in every physical design phase. This
step, indicated by the bold arrow in Fig. 1, can be manually bypassed or guided when appropriate. Additional
geometric constraints can be programmed to enhance the
chances of meeting high-level specifications.
6) Verify original specifications. This step corresponds to
the thin arrow in Fig. 1. If constraints are not met,
iterate.
The constraint translation and enforcement phases may
be applied incrementally while the design unfolds or could
be obtained from previously designed systems undergoing
redesign or technology migration. Editing and local interactive
rebudgeting are also allowed.
denoted by . All design parameters are subject to variations
with respect to their nominal (simulated or expected) values
.
With no loss of generality, we can assume that all design
parameters are normalized, so that their deviations are always
. Performance measures are functions of
nonnegative:
all design parameters
where
.
The constraints on maximum performance degradation can
be upper or lower bounds. The following notation will be used:
(1)
The problem of constraint transformation can be formulated
on which
as follows. Determine a feasible set of bounds
allow all performance constraints (1) to be met. In general,
more than one solution exists for this problem. Therefore,
constraint transformation is formulated as an optimization
of how
problem, the objective function being a model
costly it would be to meet the given set of bounds
minimize:
subject to:
(2)
is denoted as flexibility function [4].
Function
Often a linearized model is used to represent the dependency
of performance measures on design parameters. This model
can be obtained by first order Taylor expansion of
(3)
III. THE CONSTRAINT GENERATION PROBLEM
In a circuit with
performance characteristics, let
be the vector of the nominal (simulated or expected) values
for all performance measures. Because of process variations
and parasitics in the actual circuit, the performance measures
. The array of all performance deviations will
are
. Performance measures are
be denoted as
functions of the design parameters and parasitics, which will be
where is the matrix of the performance sensitivities relative
to , defined as the following:
Each entry
in the matrix represents the sensitivity of the
th measure of
with respect to the th component in .
388
IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 12, NO. 4, NOVEMBER 1999
Sensitivities can be computed using standard circuit simulators
such as SPICE, using the adjoint technique [7] or perturbation
methods [8].
The entries of can be either positive or negative; hence,
cancellations may occur within the model. However, since the
exact value of the parasitics is not known a priori, the global
effect on performance may be underestimated. Deviations of
performance measures in the positive and negative directions
must be decoupled.
By substituting the linearized expression (3) in inequalities
(1), the general problem is rewritten as
representation. Constraints cannot be defined on these parameters because a nonzero probability exists that they will not
be met. Instead, constraints must be set on the moments of a
parasitic component rather than on its actual value.
Let array
of size
define a set of known nondeterbe the array of
ministic parasitic components, and let
. Assume
all statistical distributions associated with
now that the first moments exist and are bounded for all
components of . Moreover, assume that all components of
be statistically independent. Then, the first
moments
of degradation
can be computed as
(6)
(4)
denotes the th moment and
a matrix whose
entries equal the th power of the corresponding
entry of
.
Formulation (2) of the constraint generation problem is used
to derive bounds on the first moments of , replacing the
constraint formulations with the following constraints:
where
is the matrix of the worst-case positive sensitivities
where
is the matrix of the absolute values of the worst-case
and
negative sensitivities
In the remainder of this paper, the “ ” and “ ” signs will be
omitted. The linearized model (4) will be expressed as
(5)
IV. HANDLING NONDETERMINISTIC CONSTRAINTS
Let us assume that a joint probability distribution function
is known for all nondeterministic parasitics in the circuit;
then, a moment-generating function can be derived and a
multidimensional tensor can be built for all combinations
of th order moments for each pair of parasitics. Generally,
designers assume nondeterministic parasitics to be jointly
Gaussian and hence uniquely characterized in terms of mean
vector and variance-covariance matrix. For every performance
measure, a compact stochastic model can be built if such
a representation is used. Next, high-level specifications on
performance variance can be mapped onto constraints on the
maximum degradation allowed on nondeterministic parasitic
components. Moreover, given a relationship between parasitic
tolerance and layout geometry [9], physical constraints can
be generated on the relative distances and orientations of all
layout components.
The fundamental assumption of this method is that the
statistical properties of all parasitics be known a priori and that
a bounded cross correlation matrix exist. Sensitivities are used
in a constrained optimization loop to obtain constraints on the
statistical parameters of parasitics. The optimization objective
is a continuous bounded flexibility function, which represents
an estimate of how realistic the constraint will be in the actual
design. Technology-dependent models of all known parasitic
dependencies from physical realization guide the constraintdriven tools to meet all performance specifications.
Compact models are available to model interconnect lines
and integrated devices with reasonable accuracy over a wide
range of operating frequencies. However model parameters are
technology dependent and are characterized using a statistical
where
denotes a constraint on the maximum accept.
able value of the th moment of
The fact that in most designs the entries of are not statistically independent makes the method inadequate to compute
meaningful constraints. To solve this problem, (6) must be
modified, while the optimization problem remains fundamenbe the variance-covariance matrix of
tally unchanged. Let
defined as
(7)
indicates the expected value. By conwhere operator
is a symmetric positive-definite
square
struction,
matrix. Consequently, it can be decomposed as following:
(8)
nonsingular mapping and a diagonal
where is a
is
positive definite matrix. One can easily show that if
mapped onto vector
, the new parameters in are
is
uncorrelated, i.e., the variance-covariance matrix
diagonal. The original vector can be obtained as
(9)
is due to the positive definite and
where relation
.
symmetric characteristics of
are jointly Gaussian random variIf the parameters in
ables, then the entries of are statistically independent. This
technique has been used in the past in analog test generation
and yield analysis to efficiently characterize all correlations of
Gaussian random parameters. A model based on can then
be constructed and used directly in (6) where is substituted
. Consequently problem (2) can be solved in terms of
with
vector
, i.e., the bound on the intermediate parameters
. Finally, a bound on the moments of can be obtained by
MALAVASI AND CHARBON: CONSTRAINT TRANSFORMATION FOR IC PHYSICAL DESIGN
389
where vector
is expressed in terms of all sensitivities
with respect to the above parameters as
Suppose now that the exact waveform felt at is not known,
and only an estimate can be derived. Moreover, suppose that
a range can be set for
(12)
Fig. 2. The principle and modeling of local generators.
applying the mapping of (9) to the resulting bounds. Consider
; then the transformation translates onto
the case
(10)
V. NONDETERMINISTIC SUBSTRATE CONSTRAINTS
Constraint generation requires parasitics to be associated
with one or more physical structures in the layout. In the
case of switching noise, i.e., noise produced by high-frequency
digital circuits, the physical location and transmissions paths
through the substrate may not be known before the general
floorplan is performed on the chip. For this reason, the
constraint generation process cannot take place before layout
is generated.
To address this issue we introduce the concept of local noise
generators. A local noise generator is defined as a voltage
or current source producing the equivalent cumulative noise
generated by the real noise sources located in the substrate. The
generator should simulate as closely as possible the waveform
felt at a location , including distortions, attenuations, and
group delays which transformed the original noise signal. Call
the sensing node (see Fig. 2).
such a waveform, where is the time
Let us call
and
is a vector of all the parameters relevant to it.
as a local noise generator producing waveLet us define
form . Due to the diverse nature of its parameters, can be
.
split into its basic components
represents process-dependent and layout-related parameters,
is the temperature and
the local substrate potential. One
as
can also define vector
from nominal.
the variation of
, its degradation from
Consider performance measure
nominal is given by the product of the th row of the sensitivity
with vector
matrix
(11)
and
are known vectors. Assuming that
where
the sensitivity of performance
with respect to
has
been computed, bounds on all parameter variations
can be calculated using constrained optimization as shown
in Section III. Hence, constraints are generated to bound the
amount of noise at the sensing nodes a priori, without a precise
knowledge of the structure of the layout being built. During the
physical assembly of the circuit, all precomputed constraints
will be enforced separately on each component of the layout.
Let us now generalize the problem to a large number of
sensing nodes. From a theoretical standpoint, at each receptor
a different waveform could be felt. However, since the size of
the analog section of a mixed-signal circuit is small compared
to the distance to the noise sources, we can safely assume that
all substrate nodes are reached by an identical waveform at
different times.
sensing nodes exist, each of them connected to a
Suppose
, with
, where
local generator
is the propagation delay between nodes. Due to the highly
nonlinear dependence of performance on phase, an additive
linearization around a nominal value could inaccurately model
the parasitic effects of the substrate.
The problem can be addressed by deriving a set of worstthe array of
case sensitivities as described in [10]. Call
all design parameters for which
is not strongly nonlinear.
,
Hence, a conservative estimate of the total variation of
is derived as
due to node
(13)
Using the same formalism of (11) and considering all the
in the circuit, we can define the following
sensing nodes
matrices:
..
.
and
..
.
Thus, the degradation of performance
trace
(14)
is expressed as
(15)
Equation (15) models the contributions of all sensing nodes
. Bounds on the parameters associated
onto performance
390
IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 12, NO. 4, NOVEMBER 1999
sensitivity analysis is performed on the circuit by means of
standard symbolic, numerical, or mixed techniques [7], [8].
A linearized performance model is built accounting for both
deterministic and nondeterministic parasitics
(16)
Fig. 3. Constraint check.
is the mean of . Using the variance-covariance
where
information associated with all nondeterministic parasitics,
is derived by means of singular value decomposition. The new
performance model becomes
(17)
where the new sensitivity matrix
is derived as
(18)
The mean value
is computed as follows:
(19)
and
.
Without loss of generality, assume that
Using a priori estimates of upper-bounds on parasitics
,
critical parasitics is detected by eliminating
the subset of
all parasitics for which
(20)
Fig. 4. The constraint generation flow.
can be computed using
with each sensing node
constrained optimization provided that conservative upper- and
are also available for
lower-bounds on the realization of
each sensing node . The use of worst-case sensitivity matrix
has the advantage of reducing the parameter space of .
Due to the mechanism of noise modeling obtained using local generators, constraints on noise parameters can
be derived independently of a particular IC process. Hence,
the constraint generation is required only once for a given
circuit. During physical assembly, process-dependent substrate
extraction, in combination with estimates of the sources of
switching noise, are used to enforce the bounds. Once the
substrate has been extracted, a transfer function
can be computed relating each noise source
to receptor
. Assuming that approximations or exact waveforms are
known for each noise source, waveform
and the
can be easily evaluated for each
corresponding parameter
node . Thus, a simple check can be performed to verify that
, and hence the original specifications
constraints
have been met (see Fig. 3).
VI. CONSTRAINT GENERATION ENGINE
Fig. 4 shows the constraint generation process in detail.
From hardware description and performance specifications,
is chosen so as to not excessively
is satisfied. The term
reduce the range in which the critical parasitics will be
. The
allowed to vary, while at the same time maximizing
value of
does not seem to critically affect the constraint
generation process. In our experiments, we have observed that
an acceptable number of critical parasitics can be obtained if
. For deterministic parasitics, the condition for
finding
critical parasitics is the following:
(21)
The final bounds on critical parasitics and on the first
moments of the intermediate parameters are computed using
a standard quadratic programming approach. All intermediate
parameter bounds are then reconverted into bounds on the
original parameters using (9).
The generalized constraint generation process is the core of
Cadence’s new constraint-driven physical design environment,
presently under development. The translation from high-level
specifications onto parasitic constraints can be completely
automated or driven by human interactions. The extent of the
user’s control in the constraint computation process is not limited by any of the physical assembly tools. All the constraint
generation phases are fired by a constraint manager, which is
intended to serve all the existing and future physical assembly
tools within our constraint-driven design environment. The
constraint manager will also supervise the translation of constraints at higher level, namely the budgeting process used to
MALAVASI AND CHARBON: CONSTRAINT TRANSFORMATION FOR IC PHYSICAL DESIGN
391
A. Clocked Comparator
Consider the clocked comparator COMPL, depicted in
Fig. 6. This comparator has been used as a benchmark in
several works on analog CAD [3], [11], [12], due to its
high performance sensitivity to layout details. Assume a
specification of 7 ns is imposed to switching delay
and
of 1 mV to systematic offset
. Then, vectors ,
,
are defined as
and
ns
ns
mV
mV
Let us consider the interconnect capacitances and stray resistances at all the nodes represented in the schematic of Fig. 6.
The initial number of parasitics is 19 for capacitances, (19
18/2) for capacitive cross-coupling, 28 for stray resistances,
27/2) for resistive coupling. Using conservative
and (28
values for the minimum and maximum parasitic estimates, as
Fig. 5. Constraint-driven physical layout design flow.
fF
propagate global specifications through the hierarchical levels
of complex mixed-signal designs as outlined in [1].
The constraint manager and its interactions with physical
assembly tools in the constraint-driven design environment is
shown in Fig. 5.
Constraints are generated on demand in every phase of
the layout design. The set of all constraints is partitioned
into subsets with all the constraints needed by a particular
layout phase. The constraint generator operates on the subsets
incrementally, i.e., calculating all subset constraints while
using extracted values for parasitics of previous phases and
estimates for parasitics of future phases. This is equivalent to
modifying problem (2) as
(22)
maximize:
fF
and the specifications on performance and the sensitivity
matrix obtained from SPICE, the number of parasitics for
which we need to compute a bound is reduced from
to
, namely
fF
fF
fF
fF
subject to:
(23)
(24)
(25)
represent extracted parwhere the terms
asitics associated with layout structures which have already
relates to
been generated. On the contrary, term
the budgeted performance degradation due to layout structures
which still need be generated. This flow accommodates both
comprehensive and partial constraint enforcement, where the
reproducibility of a design is ensured by a constraint status
trace.
VII. RESULTS
The constraint generation techniques outlined in this paper
have been tested on a wide range of fully analog and switched
circuits, selected from a set of commonly used industrial
applications. We will discuss here a few circuits to highlight
the proposed features. For the computation of all sensitivities,
standard analysis options available in SPICE or perturbation
based methods have been used in dc, ac, and transient domains.
where symbols
,
denote capacitive cross couplings
and stray resistance mismatches, respectively. Here the relation
between sensitivity and tightness of bounds is evident. Only
a few parameters critically affect the performance of this
circuit and therefore need be bounded tightly. In practice, only
the mismatch between the source resistances in the differen)
tial pair and between the two current mirrors (
) are responsible for the offset. The entire
and (
constraint generation procedure was performed in less than
1 s, while the sensitivity analysis required 9 s on a DEC
AlphaServer 2100 5/250 . The nondeterministic constraint
analysis, omitted here, was performed in 5 s on a DEC
45 varianceAlphaServer 2100 5/250 resulting in a 45
covariance matrix.
B. Low Power Amplifier
Consider the low power amplifier MPH, depicted in Fig. 7.
This circuit is of interest because of the presence of several
feedback and forward paths, provided to ensure stability.
The feedback paths may cause an increased susceptibility to
cross coupling parasitics, which mimic the intended stabilizing
392
IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 12, NO. 4, NOVEMBER 1999
Fig. 6. Clocked comparator COMPL.
Fig. 7. Low power amplifier MPH. (Courtesy of R. G. H. Eschauzier and J. H. Huijsing, TU Delft, The Netherlands.)
TABLE I
MPH: NOMINAL PERFORMANCE AND SPECIFICATIONS
TABLE II
MPH: CONSTRAINTS ON CRITICAL DETERMINISTIC PARASITICS
behavior of shunt capacitances. Another point of interest is
the low supply required by the circuit, making it particularly
sensitive to all device threshold voltages. Table I lists all the
performance specifications for this circuit.
A summary of all calculated constraints on deterministic
parasitics is shown in Table II.
153 constraint
Table III shows a section of the 153
variance-covariance matrix.
MALAVASI AND CHARBON: CONSTRAINT TRANSFORMATION FOR IC PHYSICAL DESIGN
MPH: BOUNDS ( 1006 )
2
ON THE
393
TABLE III
ACCEPTABLE VARIANCE–COVARIANCE MATRIX
TABLE IV
MPH: CPU TIMES FOR THE COMPUTATION OF ALL CONSTRAINTS
(a)
Fig. 8. PLL schematic.
TABLE V
PLL SPECIFICATIONS
(b)
The CPU times required for all constraint derivations is
reported in Table IV for a DEC AlphaServer 2100 5/250.
C. Phase Locked Loop
The architecture of the phase locked loop (PLL), similar to
the one proposed in [13], is shown in Fig. 8.
Schematic design and sizing were performed using topdown, constraint-driven design [14]. The specifications for the
PLL are summarized in Table V. The jitter is defined as the
ratio between the variation of an oscillation period and the
period. Due to its time variance, it is measured in terms of
peak-to-peak or RMS value.
The PLL is composed of a digital section, i.e., three divideby- modules and a phase-frequency detector (PFD), and
several purely analog circuits, namely a low-pass filter (LPF)
and a charge pump (CP). The voltage-controlled oscillator
(VCO) serves as the interface between analog and digital
sections.
The VCO consists of a ring of eight basic delay cells and
two replica bias blocks, depicted in Fig. 9(a)–(c), respectively.
The dependency of the PLL’s frequency-to-voltage signal path
from process, temperature variations, and layout parasitics was
contained by careful sizing, optimized to keep the sensitivity
(c)
Fig. 9. (a) VCO block diagram. (b) Bias. (c) Delay cell.
of the design with respect to these nonidealities within userdefined constraints [14].
The VCO, on the contrary, is sensitive to both interconnect
parasitics and substrate noise. To account for these effects, we
proceeded as follows. First, a worst-case sensitivity analysis
was performed on the delay element of Fig. 9 and the entire
ring oscillator, modified to account for resistive/capacitive
parasitics as shown in Fig. 10.
A sensitivity-based model of the VCO was constructed,
relating the jitter performance to all the known deterministic
interconnect-induced parasitics. Then, using the techniques
outlined in Section III, a set of constraints on critical interconnect lines was derived. The results are listed in Table VI, along
with the values extracted at the completion of the physical
assembly phase.
Second, using the concept of local noise oscillators, constraints on the noise amplitude sensed at each critical node
in the circuit were derived in such a way that the jitter be
within specifications. The results are presented in Table VI.
The nondeterministic parasitic effects are represented, in this
case, by the noise currents injected through the substrate into
394
Fig. 10.
IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 12, NO. 4, NOVEMBER 1999
VCO architecture with interconnect models.
TABLE VI
CONSTRAINTS OBTAINED BY THE SENSITIVITY ANALYSIS
NOISE INJECTOR AND
TABLE VII
RECEPTOR STATISTICS IN THE COMPONENTS OF THE PLL
the VCO. Using (6), a relatively accurate model of the jitter
degradation from nominal could be built as a function of the
substrate injection currents at the sources. All the potential
sources of switching noise are localized in the dividers.
Injection occurs by impact ionization through the active areas
of NMOS devices (in a N-well process) and by capacitive
coupling through junctions and interconnect. Substrate noise is
received in all the active areas of both doping types. Table VII
lists the sources and the receivers of substrate noise in the
various components of the design.
These constraints were enforced directly by module generator VCOGEN [14] and by Simulated Annealing based
placer PUPPY-A [15], via an internal cost function. The entire
circuit was subsequently verified using the substrate extraction
package SUBRES [16] and the digital noise injection analyzer
SUBWAVE [17]. The CPU times for sensitivity calculation,
constraint generation, and VCO layout synthesis are listed in
Table VIII.
Table IX shows the estimated jitter levels for the PLL after
completion of the physical assembly.
VIII. CONCLUSIONS
A comprehensive methodology has been presented for the
generation of constraints in physical design flows. The central
TABLE VIII
CPU TIMES ON A DEC ALPHASERVER 2100 5/250
TABLE IX
PLACEMENT STATISTICS OBTAINED ON A DEC ALPHASERVER 2100 5/250
idea of the methodology consists in translating a set of
specifications on high-level performance onto a set of constraints which can be handled by a constraint-driven physical
assembly environment. The translation techniques, applied to
a wide range of design parameters, are based on constrained
optimization and parasitic modeling. The method requires a
reasonable degree of circuit linearity, a well-defined set of
parasitic components, and known statistical properties for all
nondeterministic parasitics.
REFERENCES
[1] H. Chang, E. Charbon, U. Choudhury, A. Demir, E. Felt, E. Liu, E.
Malavasi, A. Sangiovanni-Vincentelli, and I. Vassiliou, A Top-Down,
Constraint-Driven Design Methodology for Analog Integrated Circuits.
Boston, MA: Kluwer, 1996.
[2] U. Choudhury and A. Sangiovanni-Vincentelli, “Constraint generation
for routing analog circuits,” in Proc. IEEE/ACM DAC, June 1990, pp.
561–566.
[3] E. Malavasi, E. Charbon, E. Felt, and A. Sangiovanni-Vincentelli, “Automation of IC layout with analog constraints,” IEEE Trans. ComputerAided Design, vol. 15, pp. 923–942, Aug. 1996.
[4] U. Choudhury and A. Sangiovanni-Vincentelli, “Use of performance
sensitivities in routing of analog circuits,” in Proc. IEEE Int. Symp.
Circuits and Systems, May 1990, pp. 348–351.
[5] G. Gad-El-Karim and R. S. Gyurcsik, “Use of performance sensitivities
in analog cell layout,” in Proc. IEEE Int. Symp. Circuits and Systems,
June 1991, pp. 2008–2011.
[6] E. Charbon, E. Malavasi, and A. Sangiovanni-Vincentelli, “Generalized
constraint generation for analog circuit design,” in Proc. IEEE ICCAD,
Nov. 1993, pp. 408–414.
[7] S. W. Director and R. A. Rohrer, “The generalized adjoint network and
network sensitivities,” IEEE Trans. Circuit Theory, vol. 16, pp. 318–323,
Aug. 1969.
[8] D. A. Hocevar, P. Yang, T. N. Trick, and B. D. Epler, “Transient
sensitivity computation for MOSFET circuits,” IEEE Trans. ComputerAided Design vol. 4, pp. 609–620, Oct. 1985.
MALAVASI AND CHARBON: CONSTRAINT TRANSFORMATION FOR IC PHYSICAL DESIGN
[9] M. J. M. Pelgrom, A. C. J. Duinmaijer, and A. P. G. Welbers, “Matching
properties of MOS transistors,” IEEE J. Solid-State Circuits, vol. 24, pp.
1433–1440, Oct. 1989.
[10] E. Charbon, E. Malavasi, P. Miliozzi, and A. Sangiovanni-Vincentelli,
“Nondeterministic constraint generation for analog and mixed-signal
layout,” IEICE Trans. Inform. Systems, vol. E80-D, Oct. 1997.
[11] E. Felt, E. Malavasi, E. Charbon, R. Totaro, and A. SangiovanniVincentelli, “Performance-driven compaction for analog integrated circuits,” in Proc. IEEE CICC, May 1993, pp. 1731–1735.
[12] B. Basaran, R. A. Rutenbar, and L. R. Carley, “Latchup-aware placement
and parasitic-bounded routing of custom analog cells,” in Proc. IEEE
ICCAD, Nov. 1993, pp. 415–421.
[13] I. A. Young, J. K. Greason, and K. L. Wong, “A PLL clock generator
with 5–110 MHz of lock range for microprocessors,” IEEE J. Solid-State
Circuits, vol. 27, pp. 1599–1607, Nov. 1992.
[14] I. Vassiliou, H. Chang, A. Demir, E. Charbon, P. Miliozzi, and A.
L. Sangiovanni-Vincentelli, “A video driver system designed using a
top-down, constraint-driven methodology,” in Proc. IEEE ICCAD, Nov.
1996, pp. 463–468.
[15] E. Charbon, E. Malavasi, U. Choudhury, A. Casotto, and A.
Sangiovanni-Vincentelli, “A constraint-driven placement methodology
for analog integrated circuits,” in Proc. IEEE CICC, May 1992, pp.
2821–2824.
[16] R. Gharpurey and R. G. Meyer, “Analysis and simulation of substrate
coupling in integrated circuits,” Int. J. Circuit Theory Applicat., vol. 23,
pp. 381–394, July–Aug. 1995.
[17] E. Charbon, P. Miliozzi, L. P. Carloni, A. Ferrari, and A. L.
Sangiovanni-Vincentelli, “Modeling digital substrate noise injection
in mixed-signal IC’s,” IEEE Trans. Computer-Aided Design, vol. 18,
pp. 301–310, Mar. 1999.
Enrico Malavasi (M’97) received the B.S. degree
from the University of Bologna, Bologna, Italy,
in 1984 and the M.S. degree from the University
of California, Berkeley, in 1993, both in electrical
engineering.
Between 1986 and 1989, he was with the Department of Electrical Engineering and Computer
Science (DEIS), University of Bologna, on research
topics related to CAD for analog circuits. In 1989,
he joined the Dipartimento di Elettronica ed Informatica at the University of Padova, Italy, as
Assistant Professor. Between 1990 and 1995, he collaborated with the CAD
group of the Department of Electrical Engineering and Computer Science at
the University of California, Berkeley, where he has carried out research on
performance-driven CAD methodologies for analog design. In July 1995, he
joined Cadence Design Systems, San Jose, CA, as an Architect for physical
design, and he is currently responsible for the development of advanced
tools for the automation and acceleration of constraint-driven mixed-signal
IC layout design. He has authored or coauthored more than 40 articles for
international journals, conferences, and books, and he holds one U.S. patent.
His research interests include several areas of design automation: layout,
design methodologies, optimization, extraction, and circuit analysis.
395
Edoardo Charbon (S’90–M’92) received the
Diploma in electrical engineering from the Swiss
Federal Institute of Technology (ETH), Zurich, in
1988, the M.S. degree in electrical and computer
engineering from the University of California, San
Diego, in 1991, and the Ph.D. degree in electrical
engineering and computer sciences at the University
of California, Berkeley, in 1995.
Between 1988 and 1989, he worked at the
Department of Electrical Engineering, ETH, where
he designed CMOS A/D converters for integrated
sensor applications. In 1989, he visited the
Department of Electrical Engineering of the University of Waterloo, Canada
where he was involved in the design and fabrication of ultra low-noise
nano-Tesla magnetic sensors. At the University of California, Berkeley, he
worked on performance-directed constraint-based analog and mixed-signal
physical design automation and accelerated substrate extraction techniques.
Since 1995, he has been with Cadence Design Systems, San Jose, CA, where
he is leading the development effort on constraint management in the physical
design group. He is also the Project Leader of Cadence’s first methodology for
intellectual property protection. He has published over 40 articles in technical
journals and conference proceedings, a book, and he has been consulting with
Texas Instruments and Hewlett-Packard. His research interests include CAD
for radio-frequency IC’s, methodologies for intellectual property protection,
substrate modeling and characterization, superconducting parasitic analysis,
and micromachined sensor design.
Dr. Charbon is a Guest Editor of the IEEE TRANSACTIONS ON COMPUTERAIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS.