Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Enhancing a Layout-Aware Synthesis Methodology
for Analog ICs by Embedding Statistical Knowledge
into the Evolutionary Optimization Kernel
Frederico Rocha
Instituto de Telecomunicações
Instituto Superior Técnico
IST-Torre Norte, AV. Rovisco Pais
1049-001 Lisboa, Portugal
Abstract— This paper applies to the scientific area of electronic
design automation (EDA) and addresses the automatic sizing of
analog integrated circuits (ICs). Particularly, this work presents
an approach to enhance a state-of-the-art layout-aware circuit-
level optimizer (GENOM-POF), by embedding statistical
knowledge from an automatically generated gradient model into
the multi-objective multi-constraint optimization kernel based on
the NSGA-II algorithm. The gradient model is automatically
generated by, first, using a design of experiments (DOE)
approach with two alternative strategies, the full factorial design
and the fractional factorial design, which define the samples that
will be accurately evaluated using the electrical simulator
HSPICE®, second, extracting and ranking the contributions of
each design variable to each performance measure or objective,
and, finally, building the model based on series of gradient rules.
The gradient model is embedded into the NSGA-II based multi-
objective multi-constraint optimization kernel, by acting on the
crossover and mutation operators. The approach was validated
with typical analog circuit structures, using the UMC 0.13 µm
integration technology, showing that, by enhancing the circuit
sizing optimization kernel with the gradient model, the optimal
solutions are achieved, considerably, faster and with identical or
superior accuracy. Finally, the results are Pareto Optimal Fronts
(POFs), which consist of a set of fully compliant sizing solutions,
allowing the designer to explore the different trade-offs of the
solution space, both through the achieved device sizes, or the
respective layout solutions.
Keywords-Analog Integrated Circuits Design; Automatic
Sizing; Electronic Design Automation; Evolutionary Computation;
Genetic Algorithms; Gradient Model.
I. INTRODUCTION
In the System-on-Chip (SoC) age it is common to find
devices where the whole system is integrated in a single chip,
this is done to reduce production costs and increase
performance. These complex single integrated circuit (IC)
designs are established in telecommunications, medical and
multimedia applications, where blocks of AMS, digital
processors and memory blocks appear together [1][2][3].
Presently most functions in mixed-signal ICs and SoC designs
are implemented using digital or digital signal processing
(DSP) circuitry, where analog blocks constitute only a small
part of the components, being essentially the link between
digital circuitry and the continuous-valued external world.
However, when integrating digital and analog circuits together
on the same die it becomes notorious that the development
time of analog blocks is much higher when compared with the
development time of digital blocks [1]. This difference is
because analog design in general is less systematic, more
heuristic and knowledge intensive than the digital counterpart,
and the lack of maturity of automation tools that are in fact
used by analog designers.
This paper describes a methodology to enhance the state-of-
the-art layout-aware circuit-level optimizer (GENOM-POF
[4]) by adding circuit specific knowledge that is automatically
extracted using machine learning techniques. The Gradient
Model, here introduced, is embedded in the genetic operators
of the NSGA-II [5] optimization kernel and is generated by
sampling the design space, extracting and ranking the
contributions of each design variable to each performance
measure or objective, and, finally, building the model based
on a set of gradient rules.
This paper is organized as follows. In section II, a general
overview on analog IC design automation techniques at
circuit-level sizing is given. In section III, the automatic
generation of the Gradient Model based on the Design of
Experiments (DOE) technique to sample a circuit is described.
Afterwards, in section IV, the GENOM-POF enhanced with
Gradient Model is presented. In section V, the implementation
results are presented. Finally, in section VI, the conclusions
and future work are addressed.
II. PREVIOUS WORK
In order to put analog sizing into context, before presenting state of the art approaches in automatic analog sizing, a brief introduction on analog design flow is provided.
A. Analog Design Flow
A design flow well accepted for analog-mixed signal ICs is
described in Fig. 1. This design flow was introduced by Gielen
and Rutenbar in [6], which consists of a series of top-down
design selection and specifications step by step, from system
level to the device level, and bottom up layout generation and
verification. Leveling and abstraction to describe and simulate
complex analog circuits becomes increasingly necessary. This
happens for three reasons. First, in a sizing methodology, the
detailed rules for lower levels are still unknown, it is necessary
to create models that translate high-level behavior. The second
reason is that the verification of mixed-signal ICs requires a
description of higher levels for the analog blocks, since these
ICs are highly complex and computationally heavy, so that it
can perform a full simulation of the mixed-signal project.
Finally, the use of IP macro cells in SoC, its virtual component
needs to be modeled efficiently. More
Abstract
More
Concrete
Circuit
Level
Level i
Verification
Extraction
Verification
Topology
Selection
Specification
Translation
Layout
Generation
Re
de
sig
n
Specification (level i+1) Layout (level i+1)
Specification (level i) Layout (level i)
System
Level
Device
Level
Backtracking
Redesign
Validation
Backtracking
Level i+1
...
...
Top-Down Electrical
Synthesis
Bottom-Up Physical
Synthesis
Validation
Redesign
Figure 1. From the system level to device level tasks of analog IC design [6].
In manual analog circuit design, designers are aided by
CAD frameworks comprised by many tools such as electric
circuits simulators (for example, Spectre® [8] or HSPICE®
[7]), by layout editors (e.g., layout Cadence Virtuoso [8]), or
tools of verification (e.g., Cadence [8] and Mentor Diva [9]).
Despite its fundamental aid to designers, these tools have few
automation options, and the ones available are not used by the
designers. The time required to manually implement an analog
project is usually of weeks or months, which is in opposition
to the market pressure to accelerate the release of new and
high performance ICs. To address all these difficult to solve
problems, electronic design automation (EDA) CAD is a
solution increasingly strong and solid.
The focuses of this paper is on the automation of the
specification translation at circuit-level, also known as circuit
sizing. In the next section, an overview is provided on the
major contributions to automatic sizing of analog ICs.
B. State of the Art
Many techniques for the automation of circuit sizing task
have been proposed by the scientific community over the past
20 years. Some applied to the circuit level, others at the
system level, or even both. Tools for automated circuit sizing
can be classified as knowledge-based or optimization-based,
as explained below.
1) Knowledge-based sizing
The first strategy to accomplish the task of circuit sizing
was to use a design plan derived from specialized knowledge.
In these methods, the designer’s knowledge plays an essential
role in creating the design plan, which is described by a set of
equations and strategies for sizing the circuit’s components,
that will lead to a circuit that meet the performance
requirements.
IDAC [11], uses the designer knowledge to carry out the
design plan, where all the design equations are solved during
the setup of the plan. As soon as the topology is chosen, the
plan is executed to the required specifications and a design is
achieved. The obtained design is then subject to a local
optimization loop. This tool has a varied library of circuits and
topologies with OpAmps, comparators, oscillators, DACs,
ADCs, etc. OASYS [12] addresses the issue with the same
strategy, however defines a hierarchy, and makes a design
plan for each sub-block. This tool also added backtracking
with design-reuse methodology for retrieving design errors.
BLADES[13], CAMP [14] and ISAID [15][16], present a
strategy a little different from those shown above, where the
knowledge of the designer is captured using in artificial
intelligence techniques.
The knowledge-based strategies achieved satisfactory
results both in terms of circuit and system level. The major
advantages of these methods are the short runtime and the
incorporation of the designer’s knowledge in the plan.
However, derivation of a design plan is complicated and takes
much time, another disadvantage of this approach is the
constant need to keep the design plan updated with the
evolution of technology, and finally the results are not
optimal, which makes this approach suited only for deriving
first-cut-designs.
2) Optimization-based sizing
A totally different approach emerged through optimization-
based sizing, where optimization techniques were added to the
tools. As the name suggests, the addition of such techniques
had as main objective to achieve optimal design solutions. The
optimization-based sizing can be classified into three major
subclasses based on the evaluation methods used, and they
are: equation-based, numerical-simulation-based and
numerical-model-based.
a) Equation-based
The equation-based approach consists of evaluating the
performance of the circuit through analytical design equations.
One of the most successful applications of these techniques
was GPCAD [17]. It used Geometrical Programming (GP) to
optimize a posynomial circuit model. Using this technique, the
sizing task was reduced to a few seconds while yielding
satisfactory results. Despite these promising results, the
extraction of posynomial models for new circuits proved
extremely difficult impairing the generalized use of this
technique.
The greatest advantage of the equation-based approach is
the short evaluation time. The difficulty in extracting new
models and model precision are the biggest disadvantages.
These techniques, as knowledge-based approaches, are
extremely suited to derive first-cut designs.
b) Numerical-simulation-based
The increase in computational resources open the way to the
simulation-based optimization. In this approach SPICE [18] or
spice like electrical simulators are used to evaluate the
circuits. In DELIGTH.SPICE [19], the designer introduces a
starting point from which the algorithm makes a local
optimization. In FRIDGE [20] optimization is done globally
without any starting point restriction. However, the designer
introduces a range for the variables to be optimized for a finite
search space. In GENOM-POF [4], multi-objective
evolutionary optimization is used. The output of the tool is a
set of solutions that exploit the tradeoffs between concurrent
objectives, giving the designer a wider range of sizing
solutions.
The easy modeling (through the netlist of the circuit) and
simulation accuracy are the strong points of the technique, but
the price is paid in execution time, where the circuit
simulation time makes it almost impossible to use at system
level. Furthermore, the constraints must be carefully set by the
designer, of the solutions, if found, may not be valid circuits.
c) Numerical-model-based
The basis of the numerical-model-based approach is the
evaluation of the circuit through the use of macro models, like
neural networks or support vector machines (SVM). To avoid
falling into the high execution time of the simulation-based
approaches, learning models are incorporated in the
optimization algorithm to reduce the use of circuit simulators.
In general, it is easier to generate (train) these models that to
derive the circuits’ equations, however there is still a
permanent tradeoff between model accuracy and generation
time. Barros et. al. [21][22] presents an optimization approach
that uses an SVM feasibility model to speed up the
convergence of the optimizer.
III. GRADIENT MODEL GENERATION
The automatic generation of the Gradient Model is based on
the Design of Experiments (DOE) technique to sample the
circuit behavior. DOE is a highly used technique when
studying the effects on the system outputs (or response
variables) of varying the inputs (or factors) [25]. The purpose
of using DOE is to extract the maximum amount of
information of the system with the smallest number of runs.
The gradient model is generated by sapling the influence of
the inputs on outputs (using DOE), extract and rank the
contributions of each design variable (input) to each design
performance or objective (output), and finally, building a set
of gradient rules that will be used to enhance GENOM-POF.
The simple differential amplifier, shown in Fig. 2, will be
used throughout the next section to exemplify the processes
used to compute the gradient model. The input variables and
ranges are shown in Table I and the output measures in Table
II. Notice that both the inputs and outputs are intentionally a
subset of the overall design parameters, once the main idea is
to proof that even with an extremely simple model can be used
to improve the optimization kernel by embedding design
knowledge into the automation cycle.
TABLE I. RANGE OF INPUT VARIABLES.
Inputs Minimum Range Maximum Range
W1 (m)
W2 (m)
IBias (A)
TABLE II. OBJECTIVES AND DESIGN CONSTRAINTS.
Outputs Objective
DC Gain [dB]
GBW [Hz]
M1
Ibias
Vdd
Vin
VCM
CloadM2
M3 M4
Vout
Figure 2. Differential amplifier.
A. Design of Experiments (DOE)
Two approaches of DOE will be used in this work, full
factorial design and fractional factorial design. The number of
simulations required to construct the DOE’s matrix (or just
matrix), of both strategies obeys the following equation:
(1)
where B is the number of points per variable or matrix base (B
> 1), n is the number of input variables and p the number of
non-elementary input variables. A non-elementary input
variable is a variable that is not sampled in all possible
combinations with the others variables. The value of p is zero
in the case of full fractional design, and equal or greater than 1
for the fractional factorial design.
1) Full Factorial Design
In the full fractional DOE the circuit is sampled in all the
combinations of variables’ values. For each variable (xi), B
logic levels are defined, and to each value, it is assigned a
value vi,b derived from the variable’s range according to eq. 2.
(2)
In Table III, the mapping between logic values and variable
values is illustrated for the simple differential amplifier
introduced previously, when B is set to 2.
TABLE III. ASSOCIATION BETWEEN RANGE VALUES AND DOE’S LEVEL.
Variables \ Levels 0 1
x1-W1 (m)
x2-W2 (m)
x3-1IBias (A)
Once the variables mapping is complete, the next step is to
construct the DOE’s matrix. The matrix has one line per each
possible combination of values, being the total number of lines
given by eq. 1, where p is 0. The columns are the inputs (x),
identified with the logic levels, and the outputs (y) described
by the measured value. The concatenation of the values in the
inputs is also referred as the code of the sample, as it acts as
unique identifier.
Table IV shows the 8 ( ) sample matrix, obtained for the
simple differential amplifier example. From the observation of
the matrix, it can be seen that simulation 2 does not have
values in the outputs. This situation occurs when, e.g.,
HSPICE® does not converge for that set of input parameters.
All vectors which produce an output that is not measurable are
not taken into consideration during the generations of the
model.
TABLE IV. DOE’S MATRIX RESULTANT FOR FULL FACTORIAL DESIGN.
x1-W1
(level) x2-W2
(level) x3-IBias
(level) y1-DC
Gain [dB] y2-GBW
[MHz]
1 0 0 0
2 1 0 0
3 0 1 0
4 1 1 0
5 0 0 1 0
6 1 0 1
7 0 1 1
8 1 1 1 ,00
2) Fractional Factorial Design
From eq. 1, the number of samples (and obviously
simulations) increases exponentially with the number of input
variables. To reduce this effect, the fractional factorial DOE
introduces the notion of non-elementary variable, as a variable
that is not used to generate the code of the sample, reducing
the size of the matrix. The level of the non-elementary
variables is determined from the code, i.e., from the levels of
the elementary ones. Several methods to compute the level of
non-elementary variables are available in the literature, in this
work the level Lni of the non-elementary variable i is given by
[25]:
(3)
where % is the modulo operator and L1 and L2 are the levels of
the first and second elementary variables, which ensures an
even distribution in the levels as can be seen in Table V.
Table V shows the 4 ( ) sample matrix, obtained for the
simple differential amplifier example by considering the
variable IBias as a non-elementary variable.
TABLE V. DOE’S MATRIX CONSTRUCTED WITH THE DESIGN: FRACTIONAL FACTORIAL.
x1-W1
(level)
x2-W2
(level)
x3-IBias
(level)
y1-DC
Gain [dB]
y2-GBW
[MHz]
1 0 0 0 1,70 0,34
2 1 0 1 30,56 10,24
3 0 1 1 31,13 12,18
4 1 1 0 56,46 20,34
B. Extraction and Ranking of the contributions
The next step is to perform the statistical analysis of the
experiments in order to understand which variables affect
most the outputs; this is called the main effect. The main
effect is the effect of one independent (input) variable on the
dependent (output) variable, ignoring the effects of all other
independent variables [25], where mi,j, the main effect of input
variable i in the output variable j is computed according to eq.
4.
∑
{
⁄
⁄
(4)
where k identifies the sample and y the output measure for the
sample.
When the total Main Effect of an input variable is
positive/negative, this is an indication that if the value of that
input variable is increased, the value of the output will tend to
increase/decrease.
Table VI shows the main effects of the input variables to the
outputs (DC Gain and GBW) for the fully factorial DOE
matrix in Table IV.
TABLE VI. MAIN EFFECT OBTAINED FROM THE FULL FACTORIAL DOE
MATRIX.
Input (xi) Calculation of the Main Effect to DC Gain mi,1
W1 (x1) ( – (
) 23,13
W2 (x2) ( ) – (
) 89,51
IBias (x3) ( ) – (
) 61,35
Input (xi) Calculation of the Main Effect to GBW mi,2
W1 (x1) ( ) – (
) 40,91
W2 (x2) ( ) – (
) 49,29
IBias (x3) ( – (
) 25,19
Table VII shows the main effects obtained from the
fractional factorial DOE matrix in Table V. The analysis of
Tables VI and VII shows all variables having a positive effect
in the output, due to its highest absolute value in both cases;
W2 is the input variable that gives more certainty in its effect
towards both outputs. Observing these tables of Main Effects
for both DOE strategies it can be concluded that both
strategies have concordant directions.
TABLE VII. MAIN EFFECT OBTAINED FROM THE FRACTIONAL FACTORIAL
DOE MATRIX.
Input (xi) Calculation of the Main Effect mi,2
W1 (x1) – 54,19
W2 (x2) – 55,13
IBias (x3) – ( ) 3,53
Input (xi) Calculation of the Main Effect mi,2
W1 (x1) – ( ) 18,06
W2 (x2) ) – 21,94
IBias (x3) – ( ) 1,74
C. Generation of the Gradient Rules
Once the main effects are computed, the N input variables
that have larger contributions to each output are identified as
the ones with the large absolute main effect. Then, a
refinement procedure is executed. For each output variable ,
a new DOE matrix is constructed using the fractional factorial
sampling, with the N input variables that have the larger
contributions as the only elementary variables.
The refinement DOE matrix is then converted to the set of
gradient rules for that output variable. This is done by
discarding the columns referring to non-elementary variables
and transforming the levels of the elementary variables into
input gradient symbols according to:
{
⁄
⁄
(5)
where k identifies the line of the matrix. The output gradient
symbols So are converted from the output values as:
{
( )
(6)
where and
are respectively the maximum and
minimum values of the output obtained in the DOE matrix
(not the refinement matrix), and is |
| ⁄ . The
meanings of the symbols are: (-) a decrease; (+) increase and
(U) undefined. The Table VIII illustrates the eq. 6 in Table V
as matrix of refinement.
Table IX illustrates the corresponding gradient rules. Rules
2 and 3 are unusable because they have undefined meaning,
however rules 1 and 4 can be used when trying to drive GBW
low or up, respectively.
TABLE VIII. EXTRACTION OF GRADIENT RULES FOR GBW.
(Symbol)
= 0,34
17,03 8,67
: (-)
= 10,24 : (U)
= 12,18 : (U)
= 20,34 : (+)
TABLE IX. SET OF GRADIENT RULES FOR GBW.
K Si1,2-W1 Si2,2-W2 So2-GBW
1 (-) (-) (-)
2 (+) (-) (U)
3 (-) (+) (U)
4 (+) (+) (+)
IV. GENOM-POF ENHANCED WITH THE GRADIENT MODEL
GENOM-POF’s, is part of the AIDA analog IC design
automation framework that results from the integration of two
in-house developed analog IC design automation tools, the
GENOM-POF [4] and LAYGEN-II [24]. Before moving to
the description of the how the gradient model is used to
enhance GENOM-POF, it is reviewed and contextualized.
A. AIDA Architecture
The AIDA platform, whose architecture is shown in Fig. 3,
implements a fully automatic approach from a circuit level
specification to a physical layout description. AIDA monitors
the implemented design flow allowing the designer to
intervene, e.g., by stopping the synthesis process whenever an
acceptable solution is already present or by selecting the
solution to be integrated from a Pareto set of optimally sized
circuits.
A I D A
DESIGN KIT
Module Generator
Desing Rules
LAYGEN II
Layout Generation
DESIGN
Specs.
DATABASE
TopologyTemplate
Extraction
GENOM_POF
Circuit-LevelSynthesis
VALIDATION
RE-D
ESIG
N
Figure 3. AIDA Architecture
In AIDA, GENOM-POF performs fully automated circuit-
level synthesis based on elitist multi-objective evolutionary
optimization, where the performance evaluation is done using
commercial electrical simulator HSPICE®, and LAYGEN II
implements a DRC proved fully automated layout generation
based on a sized circuit-level description and high level layout
guidelines, described in a technology independent abstract
layout template.
Besides the layout generation, LAYGEN II also implements
a lightweight placement tool, LII-lwPlacer, which uses the
same high level layout guidelines to generate efficiently and
accurately a floorplan of the circuit being evaluated. This
lightweight placer is used during GENOM-POF optimization,
enabling layout-aware circuit sizing optimization.
GENOM-POF’s architecture, the relevant tool for this work,
is briefed in the next section.
B. GENOM-POF Architecture
GENOM-POF is based on the elitist multi-objective
evolutionary optimization kernel NSGA-II, and uses the
industrial grade simulator HSPICE® and LAYGEN II to
evaluate the performance of the design. GENOM-POF targets
the design of robust circuits, by allowing the consideration of
corner cases during optimization.
GENOM-POF inputs from the designer are the circuit and
test-benches in the form of HSPICE® netlist(s)and the layout
template for LAYGEN II. The netlist(s) must have the
optimization variables as parameters, and must include means
to measure the circuit’s performance; the corner’s parameter
variations are also included in the netlist. In addition, the
designer defines ranges for the optimization variables, design
constraints, and optimization objectives.
Then, GENOM-POF, models the circuit as an optimization
problem, defined by the tuple {x - optimization variables, F(x)
-objective functions, G(x) - constraint functions} and suitable
to be optimized by the NSGA-II kernel. The output is a family
of Pareto optimal circuits that fulfill all the constraints and
represent the feasible tradeoffs between the different
optimization objectives.
The enhanced GENOM-POF architecture is shown in Fig. 4.
The gradient model is integrated in GENOM-POF by
embedding the extracted circuit knowledge into the
evolutionary kernel operators increasing their efficiency. The
next section describes in detail how the integration is done.
NSGA-II
KERNEL
Circuit Sizing
CORNERSTYPICAL
{X, F(x), G(x)}
OUTPUTS
Sized Circuits
POF
INPUTS
Targets &Constraints
Template
Circuit &Testbench
HSPICE®
Electrical Simulator
LAYGEN-II
Layout Estimator
evalu
ation
opera
tors
CROSSOVER
MUTATION
Enhanced Operators
Gradient Model
Figure 4. GENOM-POF architecture with the integrated Gradient Model.
C. Gradient Model Integration
The integration of the Gradient Model into GENOM-POF is
done in two fronts, first by embedding it in the evolutionary
operator of mutation and second by adding the Gradient
Model setup interface to the AIDA graphical interface.
1) Gradient Model applied to the Mutation Operator
The NSGA II kernel is an evolutionary optimization scheme
that simulates natural evolution. It operates over a population
composed by several chromosomes, each representing a
different candidate solution. The genetic operators, crossover
and mutation, are used to create new individuals from the
initial population (usually obtained randomly), the first, by
combining the genetic information from the parents, and the
second by introducing random changes in the individual.
The new individuals’ fitness is evaluated and they are mixed
with the parents and ranked. The fittest individuals are
selected as the new parents, and the less fit discarded. The
process is repeated until the ending criterion is reached
(usually a fixed number of iterations). The distinguishing
characteristic of NSGA-II is that the ranking is made using
Pareto dominance.
In GENOM-POF the chromosome is represented by the
vector of continuous variables representing the
design variables. In order to speed up the convergence of the
algorithm the gradient model is used to introduce design
knowledge into the mutation operator.
The mutation operator in GENOM-POF uses the continuous
valued operator introduced Deb and Goyal in [26]. In this
operator, defined as
⁄ , where
and are the mutated and original values respectively, is
the mutation perturbation applied. is a random variable,
with values in the interval of -1 to 1, and p.d.f.:
| | (7)
where is a parameter used to control the spread of the
distribution. Fig. 5 shows the p.d.f. for various values of .
η = 0
η = 1
η = 5
δ
P(δ)
-1 10
1
0.5
3
Figure 5. Probability distribution for creating a mutated value.
A factor of disturbance of can be obtained from an
uniform random number using eq. 7, which is
obtained from eq. 8 by solving ∫
.
{
]
(8)
and the mutated value, , is given by
. The gradient rules are then applied. The application of
the rules follows the expression in eq. 9:
( )
(9)
where is the variable value after the application of the rule,
is a function of the gradient symbol defined in eq. 10,
and is a uniformly distributed random number
between 0 and c, the change rate model parameter.
{
(10)
The application of the gradient rules is conditioned to
existence of a suitable rule for the optimization targets, i.e. it
is irrelevant to have a rule to decrease the gain, if the
optimization target is to increase it. The rules are selected by
searching for each optimization objective if there is a rule that
causes the desired effect in the corresponding response
variable. If this rule is found, then the variables with larger
contributions are affected as described before. Fig. 6 shows an
example of the application a gradient rule.
The mutation rate and Gradient Model rate are also
controlled by the parameters mrate and gmrate, respectively.
W1
216.3
W2 L1
0.35
L2
0.35
Ib
309
Individual: after gradient model
W1
210
W2
105
L1
0.35
L2
0.35
Ib
300
Individual: after mutation
Gradient Model
GRADIENT MODEL
W1
(+)GBW,(+)
W2
(+)
L1 L2 Ib
A0,(-) (+) (+)
Rules:
Parameters: c = 0.03
Objective: max(GBW)
111.4
Figure 6. Example of application of the Gradient Model in the mutation.
2) Gradient Model Graphical User Interface (GUI)
The AIDA platform includes a simple but efficient GUI
implemented in Java™ 1.6. This GUI (shown in Fig. 7)
provides a simple and fast way for the designer to set variables
ranges, optimization objectives and constraints, and the
optimization options, also it allows the designer to monitor the
automatic generation flow.
To ease the usage of the gradient model, a gradient model
specific configuration interface was added to the framework,
that provides the designer the option to use it or not, and the
control over the model parameters Apply Rate (gmrate) and
Change Rate (c). This is used together with the NSGA-II
optimization options, also shown in Fig. 7, to setup the
optimization.
V. RESULTS
The proposed methodology was run, on an Intel® Core™ 2
Quad CPU 2.4 GHz with 6 GB of RAM and with multi-
threads to perform the evaluation process of each population,
at each generation.
A. Single-Ended Folded Cascode Amplifier
The circuit used to compare the GENOM-POF with
GENOM-POF integrated with Gradient Model is the single
ended folded cascade amplifier, presented in Fig. 8. For the
setup of this comparison the items required were the netlist
and the testbench of the circuit. This case study was done
considering 15 input variables, 2 objectives and 19 constraints
defined in Table X.
The optimization variables are the widths and lengths of, the
cascade bias tensions vbnc and vbpc respectively, and the bias
current. This circuit is optimized in both GENOM-POF and
GENOM-POF integrated with Gradient Model in exactly the
same conditions, for a fair analysis and comparison between
them.
This case study was optimized with the following setup:
mutation rate of 30%, population size of 128 and crossover
over of 90%. The Gradient Model parameters were: apply rate
of 50% and change ratio of 3%. The values of these
parameters were determined using the analysis of previous
experimental data.
For this study all the 15 input variables are considered, the
Gradient Model was generated with a base of two (B = 2) and
considering only the design variable with larger contribution
(N = 1). The extracted gradient rules for the optimization
objective are shown in Table XI. The model was automatically
generated in less than 5 minutes, and can be reused if only the
constraints are changed.
Figure 7. AIDA GUI.
ip
In
Vbp
Vdd
Vss
Ib
Ibp=1.1Ib
Vac
Vip
Vbpc
Vbnc
Vb Vss Vbnc
out
Vbpc Vdd
W=W5L=L5M=M5
W=W5L=L5M=M5
W=W7L=L7M=M7
W=W7L=L7M=M7
W=W9L=L9M=M9
W=W9L=L9M=M9
M1
Vdd
Vbp
Vbpc
inip
Vss
out
Vbnc
vb
M2
M4M11 M12
M9 M10
M7 M8
M5 M6
W=W1L=L1M=M1
W=W1L=L1M=M1
W=W4L=L4M=M4
W=W11L=L11M=M11
W=W11L=L11M=M11
Amplifier
(b)
(a)
Figure 8. Electrical schematic and testbench of the single-ended folded
cascode amplifier.
TABLE X. RANGES, OBJECTIVES AND CONSTRAINTS.
Variables: cn, cp, l1, l4, l5, l7, l9, l11,
ib, w1, w4, w5, w7, w9, w11
Ranges: 0.18e-6 <= l* <= 5.0e-6
0.24e-6 <= w* <= 200.0e-6 -0.4 <= cn <= 0.0
0.0 <= cp <= 0.4
30.0e-6 <= ib <= 400.0e-6
Objectives: min(area) max(a0)
Constraints: gb >= 1.2e7
a0 >= 80
55 <= pm <= 90 sr >= 1e7
ov_m(*) >= 30e-3
d_m(*) >= 1.2 osp >= 0.3
osn <= -0.3 (*) the constraints apply to:
M1, M4, M5, M7, M9 and M11
TABLE XI. GRADIENT RULES.
Target Variable / Gradient
A0, (-) L9, (-)
A0, (+) L9, (+)
area,(-) W11, (-)
area,(+) W11, (+)
Fig. 9 illustrates the improvements achieved by the use of
this simple model. It shows that the Gradient Model enhanced
GENOM-POF achieves better solutions at generation 2.000
then GENOM-POF at generation 2.000 or 4.000.
Figure 9. GENOM-POF (for 60.000, 4.000 and 2.000 generations) vs
GENOM-POF integrated with Gradient Model (for 2.000 generations).
The Fig. 9 also shows that even for 60.000 generations the
GENOM-POF can’t reach the maximum DC Gain reached by
GENOM-POF integrated with Gradient Model. To confirm
that this is not an isolated case, 20 executions with different
seeds were done. The output is shown in Fig. 10, were it can
be seen that the inclusion of gradient model consistently lead
to better solutions.
Figure 10. GENOM-POF vs GENOM-POF integrated with Gradient Model
for 20 different initial populations (for 2.000 generations).
Table XII shows the average over the 20 runs of: the
number of point in the final POF, and the normalized non-
dominated area, which measures the ratio between the non-
dominated and dominated area in the performance plane. It
confirms the analysis of Fig. 10, where the GENOM-POF
enhanced with the gradient model outputs more and better
solutions.
TABLE XII. GENOM-POF VS GENOM-POF INTEGRATED WITH
GRADIENT MODEL (2.000 GENERATIONS).
Nr. Points POF Non-dominated area
GENOM-POF 51.55 0.426
GENOM-POF
integrated with
Gradient Model 81.7 0.2
VI. CONCLUSIONS AND FUTURE WORK
The work presented in this paper corresponds to an
innovative IC design automation approach by embedding a
simple but effective design knowledge model, Gradient
Model, into the evolutionary optimization kernel of a state-of-
the-art sizing tool. The new technique proved to be capable to
accelerate and reduce the execution time of the circuit-level
optimizer GENOM-POF. This integration of the Gradient
Model with GENOM-POF enhances the optimizer efficiency,
forwarding the data to the desired objectives and causing a
significant reduction in the number of electrical simulations.
The model potential has been proved through a complex case
study presented. Finally, the proposed objectives for this work
were achieved and a new optimizer was created.
In analog IC design automation, the developments of new
and better approaches are always needed and there is still a
long way to end in this domain. Based on this work, some
suggestions can be made for future research which certainly
will improve even more the efficiency of the proposed
approach. The first suggestion is the application of the
Gradient Model during Corners validation. The second
suggestion to improve the accuracy of the model is to refine
the model of the circuit after the first POF solution (feasible
region) is reached, to increase the accuracy of the model.
The integration of the model in GENOM-POF can be
performed for alternatives approaches. For example, the
application of the Gradient Model to each objective or all the
objectives with different weights defined by the user. An
example of this approach is presented in Fig. 11. The expected
result of this alternative approach is to accelerate the
optimization problem by the application of the model to all the
objectives, and at the same time to maximize/minimize even
more the objectives by the single application of the model to
an objective.
50% of application of the
Gradient Model to all
Objectives
25% of application of the
Gradient Model to Area
25% of application of the
Gradient Model to Gain
Figure 11. Alternative approach of application of Gradient Model.
ACKNOWLEDGMENT
This work was partially supported by the Instituto de Telecomunicações (Grant PEst-OE/EEI/LA2008/2011).
REFERENCES
[1] G. G. E. Gielen, “CAD tools for embedded analogue circuits in mixed-signal integrated systems on chip,” IEE Proceeding on Computers and Digital Techniques, vol. 152, no. 3, pp. 317–332, May 2005.
[2] “International Technology Roadmap for Semiconductors 2012 Edition,” http://public.itrs.net/.
[3] N. Kordas, Y. Manoli, W. Mokwa and M. Rospert, “The properties of integrated micro-electrodes for CMOS-compatible medical sensors,” Engineering in Medicine and Biology Society, 1994 Advances: New Opportunities for Biomedical Engineers. Proceedings of the 16th Annual International Conference of the IEEE, vol. 2, pp. 828-829, 1994.
[4] N. Lourenço and N. Horta, “GENOM-POF: Multi-Objective Evolutionary Synthesis of Analog ICs with Corners Validation”, GECCO’ 12: Proceedings of the fourteenth international conference on Genetic and evolutionary computation conference, July 2012.
[5] Deb, K., Pratap, A., Agarwal, S. and Meyarivan, T. 2002. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE T Evolut Comput. 6, 2 (Apr. 2002), 182-197. DOI= http://dx.doi.org/10.1109/4235.996017
[6] G. G. E. Gielen, and R. A. Rutenbar, “Computer-aided design of analog and mixed-signal integrated circuits,” Proceedings of the IEEE, vol. 88, no. 12, pp. 1825-1852, Dec. 2000.
[7] Synopsis Inc. (2012) HSPICE® simulator. http://www.synopsys.com/Tools/Verification/AMSVerification/CircuitSimulation/HSPICE/Pages/default.aspx. Accessed September 2012.
[8] Cadence Inc. (2012) Spectre simulator and other cadence products. http://www.cadence.com/products/pages/default.aspx Accessed September 2012.
[9] Mentor Graphics Inc. (2012) IC verification and other products. http://www.mentor.com/products/ Accessed September 2012.
[10] Rutenbar, R.A.; Gielen, G.G.E.; Roychowdhury, J.; “Hierarchical Modeling, Optimization, and Synthesis for System-Level Analog and RF Design”, Proceedings of the IEEE on Digital Object Identifier, vol.95, pp.640-669, March 2007.
[11] M. G. R. Degrauwe, O. Nys, E. Dijkstra et al., “IDAC an interactive design tool for analog CMOS circuits,” IEEE Journal of Solid-State Circuits, vol. 22, no. 6, pp.1106-1116, 1987.
[12] R. Harjani, R. A. Rutenbar, and L. R. Carley, “OASYS a framework for analog circuit synthesis,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 8, no. 12 pp.1247-1266, 1989.
[13] F. El-Turky, and E. E. Perry, “BLADES an artificial intelligence approach to analog circuit design,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 8, no. 6, pp. 680-692, 1989.
[14] B. J. Sheu, J. C. Lee, and A. H. Fung, “Flexible architecture approach to knowledge-based analogue IC design,” Circuits, Devices and Systems, IEE Proceedings G, vol. 137, no. 4, pp. 266-274, 1990.
[15] C. A. Makris, and C. Toumazou, “Analog IC design automation. II. Automated circuit correction by qualitative reasoning,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 14, no. 2, pp. 239-254, 1995.
[16] C. Toumazou, and C. A. Makris, “Analog IC design automation. I. Automated circuit generation: new concepts and methods,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 14, no. 2, pp. 218-238, 1995.
[17] M. del Mar Hershenson, S. P. Boyd, and T. H. Lee, “GPCAD: a tool for CMOS op-amp synthesis,” International Conference on Computer-Aided Design, Digest of Technical Papers of the IEEE/ACM pp. 296-303, 8-12 Nov 1998, 1998.
[18] L. W. Nagel, “SPICE2: A Computer Program to Simulate Semiconductor Circuits,” EECS Department, University of California, Berkeley, 1975.
[19] W. Nye, D. C. Riley, A. Sangiovanni-Vincentelli et al., “DELIGHT.SPICE: an optimization-based system for the design of integrated circuits,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 7, no. 4, pp. 501-519, 1988
[20] F. Medeiro, F. V. Fernandez, R. Dominguez-Castro et al., "A Statistical Optimization-based Approach For Automated Sizing Of Analog Cells.", Conference on Computer-Aided Design, pp. 594-597, 1994.
[21] M. F. M. Barros, J. M. C. Guilherme, and N. C. G. Horta, Analog circuits and systems optimization based on evolutionary computation techniques, Berlin: Springer, 2010
[22] M. Barros, J. Guilherme, and N. Horta, "GA-SVM Optimization Kernel applied to Analog IC Design Automation.", IEEE Internation Conference on Electronics, pp. 486-489, 2006.
[23] J. Antony, N. Capon, “Teaching Experimental Design Techniques to Industrial Engineers”, Int. J. Engineering, No. 5, pp. 335-343, Aug. 1998
[24] Martins, R., Lourenco, N. and Horta, N. 2012. LAYGEN II: Automatic Analog ICs Layout Generator based on a Template Approach. In Proceedings of the Genetic and Evolutionary Computation Conference (Philadelphia, USA, July 7 - 11, 2012).
[25] D. C. Montgomery, Design and Analysis of Experiments, 5th ed. John Wiley and Sons, New York, 2001.
[26] Deb, K., Goyal, M. A combined genetic adaptive search (geneas) for engineering design. Computer Science and Informatics, 26(4): 30-45, 1996.