POWER ESTIMATION FOR REGISTER TRANSFER LEVEL BY

GENETIC ALGORITHM

Yaseer A. Durrani, Teresa Riesgo, Felipe Machado

Universidad Politécnica de Madrid

E.T.S.I. Industriales. División de Ingeniería Electrónica

C/ José Gutiérrez Abascal, 2. 28006 Madrid (Spain)

Keywords: Power estimation, Macromodel, Signal statistics, Genetic algorithm.

Abstract: In this paper, we propose a new genetic algorithm (GA) based macromodeling technique for register

transfer level. This technique allows to estimate the power dissipation of intellectual property (IP)

components using some statistical knowledge of their primary inputs. During the macromodel construction

procedure, the sequence of an input stream is generated by a GA using input metrics. Then, a Monte Carlo

zero delay simulation is performed and the power dissipation is predicted by a macromodel function. In our

experiments with IP macro-blocks, the results are effective and highly correlated, with an average error of

1%. Our model is parameterizable and provides accurate power estimation.

1 INTRODUCTION

The importance of designing low power digital

circuits is being increased rapidly. In order to handle

the ever increasingly complexity, CAD tools have

been developed that also help in minimizing power

dissipation.

In this short paper, we focus on the problem of

power estimation at register transfer level (RTL) for

IP-based designs. Various power estimation

techniques have been introduced previously. These

techniques can be divided into two categories:

probabilistic and statistical. Probabilistic techniques

(Ghose et al, 1992), (Najm et al, 1990), (Marculescu

et al, 1994) use the probabilities of the input stream

and their propagation into the circuit to estimate the

internal switching activities. These techniques are

very efficient, but they cannot accurately capture

factors like glitch generation and propagation. On

the other hand in statistical techniques (Yacoub &

Ku, 1989), (Huizer, 1990), (Deng, 1994) the circuit

is simulated under randomly generated input

patterns and the power dissipation is monitored

using a power estimator. The Monte Carlo

simulation technique was developed and presented

in (Burch et al, 1993). This technique uses input

vectors that are randomly generated and its power

dissipation is estimated using power estimator.

A look-up table (LUT) based macromodel was

presented in (Gupta & Najm, 1997) and further

improved in (Gupta & Najm, 1999). The LUT stores

the estimates for equi-spaced discrete values of the

input signal statistics. The interpolation method was

introduced to allow the estimation for the input

statistics that not correspond to LUT. In (Chen et al,

1997), (Chen & Roy, 1998) interpolation scheme

was improved by using power sensitivity concept.

For better accuracy, numerous power

macromodeling techniques have been introduced in

(Gupta & Najm, 1999), (Liu & Papaefthymiou,

2002).

Genetic Algorithms (Davis, 1991) have proved

success in solving electronic design problems

(O’Dare & Arslan, 1994), (Arslan et al, 1996) and

have shown a high degree of flexibility in handling

power constraints (Davis, 1991). They are more

dynamic to combine power of randomness and

evolution, and to analyze large solution space.

In this paper, we present a new genetic algorithm

based power macromodeling technique for power

estimation. The input metrics of our macromodel are

the average input signal probability P

in, average

input transition density D

in, input spatial correlation

in and input temporal correlation Tin. We use

intellectual property (IP) macro-blocks for our

experiments.

527

A. Durrani Y., Riesgo T. and Machado F. (2006).

POWER ESTIMATION FOR REGISTER TRANSFER LEVEL BY GENETIC ALGORITHM.

In Proceedings of the Third International Conference on Informatics in Control, Automation and Robotics, pages 527-530

DOI: 10.5220/0001214205270530

 SciTePress

The rest of this paper is organized as follows. In

Section 2 we give the background of the input

parameters of our macromodel. In Section 3, we

explore our genetic algorithm. In Section 4, we

discuss about our macromodel construction. This

macromodel is evaluated in section 5. Section 6

summarizes our work.

2 CHARACTERIZATION

Similar power macromodeling techniques were

presented in (Chen & Roy, 1998), (Gupta & Najm,

1999), (Liu & Papaefthymiou, 2002). Our

macromodel consist of a nonlinear function based on

LUT approach:

),,,(

ininininavg

TSDPfP = (1)

The function f is obtained by a given IP macro-

block. The components are simulated under different

sample streams with, P

in, Din, Sin, Tin. This model

estimates the average power dissipation.

Given an IP macro-block with the number of

primary inputs r and the input binary stream

),...,,...,,(),,...,,{(

2222111211 rr

qqqqqqq =

)},...,,(

21 srss

qqq

of length s, these metrics

are defined in (Gupta & Najm, 1999), (Gupta &

Najm, 1999), (Barocci et al, 1999), (Bernacchia &

Papaefthymiou, 1999):

∑

== 11

(2)

)1(

−×

⊕

∑∑

−

jiij

(3)

)1(

111

−××

⊕

∑∑∑

===

rrs

ikij

(4)

⊗

∑∑

+−

−=

)(

(5)

3 EXPLORATION ALGORITHM

In this section, we analyze our genetic algorithm for

power macromodeling allowing to estimate the

power dissipation. The proposed GA for the IP

system is presented in Figure 1. The flow of our GA

is explained in next sections.

GA for sequence of input pattern ()

fitness_value = 0;

num_gen = 0;

Generation of randomly population;

While(num_gen < max_num of

Generations)

Compute fitness values in

Population;

Upgrade fitness_value;

Crossover;

Mutation;

Upgrade population;

num_gen + = 1;

end while;

Figure 1: Genetic Algorithm (GA).

3.1 Setting Chromosome

In GA process, chromosomes are exposed to genetic

operators like crossover, mutation, selection etc. The

objective of these operations is to remove poor

strings and produce healthy strings. In this step, we

makeup of a given chromosome. It includes the

number of genes, and the representation of those

genes. In our problem, the chromosomes are the

primary inputs.

3.2 Fitness Function Implementation

The fitness function is responsible for performing

evaluation and returning fitness value that reflects

how optimal the solution is: the higher the number,

the better the solution. Given fitness values are then

used in a process of natural selection to choose

which potential solutions will continue on to the

next generation, and which will die out. The natural

selection process does not choose the top n number

of solutions; the solutions are instead chosen

statistically that are more likely with a higher fitness

value. Our algorithm is used with measurements to

evolve the population of solutions toward a more

optimal set of solutions.

3.3 Setup and Creating Population

GA needs to provide three extra pieces of

information like fitness function, chromosome setup,

ICINCO 2006 - ROBOTICS AND AUTOMATION

528

and the chromosome population. A large number of

population size means more potential solutions to

choose and more generic diversity. Each potential

solution is represented by a chromosome. A

population of chromosome is called a Genotype and

that is we need to create our population. We setup a

configuration object and return a Genotype with the

correct number of chromosomes, each one of this

has its genes set to random values. In other words, it

generates a random population. It is necessary to

create initial population for our potential solutions.

We have specified an initial population as an input

of GA. The initial population contains M number of

random strings of length N, where M and N are the

input parameters used in GA.

3.4 Evolutionary Process

After setup and creation of population, our GA

evolves the population until it contains satisfying

potential solutions. We choose to evolve the

population a set number of times and then we check

what is the best solution produced by our genetic

algorithm.

4 MODEL CONSTRUCTION

Several approaches (Chen & Roy, 1998), (Gupta &

Najm, 1999), (Liu & Papaefthymiou, 2002) have

been proposed to construct power macromodel on

ISCAS-85 benchmark circuits. We have found that

the same methodology works as well for IP macro-

blocks such as array multipliers or comparators in

terms of the statistical knowledge of their primary

inputs. By the following method (Gupta & Najm,

1999), we found that a good choice of the P

avg in

(1), is quadratic for array multipliers and

comparators respectively. The sequence of input

stream is generated by GA for a desired input

metrics. Then, a Monte Carlo zero delay simulation

(Burch et al, 1993) is performed and the power

dissipation is obtained using a power estimator.

Different input stream sequences are generated for

different signal statistics and the corresponding

power consumptions are measured. Using all these

measures, the function f in (1) is calculated.

Table 1: Accuracy of Power Estimates.

Circuits Average

Error

Max Error

Mult8x8-1 0.76% 2.35%

Mult8x8-2 1.50% 3.19%

Mult4x4-1 0.67% 2.30%

Mult4x4-2 2.15% 3.00%

Comp-1 1.44% 2.95%

Comp-2 0.33% 1.43%

Comp16x16-1 0.58% 0.95%

Comp16x16-2 0.47% 0.64%

5 MACROMODEL EVALUATION

Experimental results show that our GA can produce

sequences with accurate statistics and highly

convergence. For the input metrics P

in, Din, Sin, Tin,

we specify the range between [0.1, 0.9]. We

generated 450 sequences with 8, 16 and 32 bits

wide. The sequence length is 2000 and 1000 vectors

for macro-blocks. In table (1) the first column shows

the name of the circuits. Columns two and three give

the average and maximum relative error for the

estimates obtained with our macromodel. The

reference values for the circuit’s power dissipation

are obtained using time delays from the Synopsys

PowerCompiler. In our experiments, the average

absolute error is 0.98%, and the average maximum

error is 2.1%. The maximum worst-case error is no

more than 3.19%.

In figure 2 we illustrate the combined scatter plot

of IP macro-blocks between the macromodel and the

reference simulated power. Regression analysis to fit

the model’s coefficients is performed. For different

blocks, the prediction correlation coefficient is

measured around 97%. The sequences have high

convergence and uniformity. Figure 3 plots the

variation of the power value with the trial interval

length. This figure shows that the interval length is

1000. The warm-up length is about 400 while the

vertical line represents the steady state value at 800.

6 CONCLUSIONS

We have presented a new genetic algorithm based

power macromodeling technique for high-level

power estimation. Our technique has been applied

on IP macro-blocks and has demonstrated good

accuracy. Our model shows an average error of 1%

and a prediction correlation coefficient of 97%. We

are currently evaluating our macromodel on

sequential circuits.

POWER ESTIMATION FOR REGISTER TRANSFER LEVEL BY GENETIC ALGORITHM

529

100 200 300 400 500 600 700 800 900 1000

1.05

1.1

1.15

1.2

1.25

1.3

Interval Length (Clock Cycle)

Power ( mW)

Steady State

Figure 2: Power comparison between four dimensional macromodel and reference simulated power.

9 9.5 10 10.5

9.2

9.4

9.6

9.8

10.2

10.4

10.6

Power from Macromodel

Power f rom Simul at ion

Figure 3: Power changes with respect to sequence length for different IP blocks.

REFERENCES

Ghose, A. A., Devdas, S., Keutzer, K., & White, J., June

1992. Estimation of average switching activity in

combinational and sequential circuits. In Proceedings

of the 29

Design Automation Conference, pages 253-

259.

Marculescu, R., Marculescu, D., & Pedram, M.,

November 1994. Logic level power estimation

considering spatiotemporal correlations. In

Proceedings of the IEEE International Conference on

Computer Aided Design, pages 224-228.

Yacoub, G.Y., & Ku, W. H., 1989. An accurate simulation

technique for short-circuit power dissipation. In

Proceedings of International Symposium on Circuits

and Systems, pages 1157-1161.

Deng, C., April 1994. Power analysis for CMOS/BiCMOS

circuits. In Proceedings of 1994 International

Workshop on Low Power Design, pages 3-8.

Burch, R., Najm, N. F., Yang, P., & Trick, T., March

1993. A Monte Carlo approach for power estimation.

IEEE Transactions on VLSI Systems, 1(1):63-71.

Gupta, S., & Najm, F. N., June 1997. Power

macromodeling for high level power estimation. In

Proceeding 34

Design Automation Conference.

Gupta, S., & Najm, F. N., March 1999. Analytical model

for high level power modelling of combinational and

sequential circuits. In Proceeding IEEE Alessandro

Volta Workshop on Low Power Design.

Chen, Z., Roy, K., & Chou, T. L., Nov. 1997. Power

sensitivity-a new method to estimate power dissipation

considering uncertain specifications of primary inputs.

In Proceeding IEEE International Conference on

Computer Aided Design.

Chen, Z., & Roy. K., June 1998. A power macromodeling

technique based on power sensitivity. In Proceeding

Design Automation Conference.

Gupta, S., & Najm, F. N., 1999. Power Macromodeling

for High Level Power Estimation. In Proceeding IEEE

Transactions on VLSI.

Liu, X., & Papaefthymiou, M. C., May, 2002.

Incorporation of input glitches into power

macromodeling. In Proceeding IEEE Inter. Symp. On

Circuits and Systems.

Davis, L., 1991. Handbook of Genetic Algorithm. Van

Nostrand Reinhold, New York.

O’Dare, M.J., & Arslan, T., May 1994. Generating test

patterns for VLSI circuits using a Genetic Algorithm.

In Proceeding IEE Electronics Letters, Vol. 30, No.

10, pp. 778-779.

Arslan, T., Horrocks, D. H., Ozdemir, E., 1996. Structural

cell-based VLSI circuit design using a Genetic

Algorithm. International symposium on circuits and

systems, Atlanta, USA.

ICINCO 2006 - ROBOTICS AND AUTOMATION

530