Power Reduction in FIR

8/12/2019 Power Reduction in FIR

1/16

Hindawi Publishing CorporationVLSI DesignVolume 2012, Article ID 870546,15pagesdoi:10.1155/2012/870546

Research ArticlePowerConsumptionModels for Decimation FIR Filters inMultistandard Receivers

KhaledGrati,1 Nadia Khouja,1 Bertrand LeGal,2 and Adel Ghazel1

1 Cirtacom Laboratory, Ecole Superieure des Communications de Tunis, 2083 Ariana, Tunisia2 IMS Laboratory, Universite de Bordeaux I, Bordeaux, 33405 Talence, France

Correspondence should be addressed to Nadia Khouja,[email protected]

Received 3 December 2011; Revised 9 March 2012; Accepted 15 March 2012

Academic Editor: Frank Kienle

Copyright 2012 Khaled Grati et al. This is an open access article distributed under the Creative Commons Attribution License,which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Decimation filters are widely used in communication-embedded systems. In fact, decimation filters are useful for implementingchannel filtering or selection with low-computation complexity requirements. Many multistandard receiver designs that arerequired in ubiquitous embedded systems are based on a cascade of decimation filter processing. Filter number and implementa-tion architectures have a significant impact on systemperformances, such as computation complexity, area, throughput, andpowerconsumption. In this work, we present filter power consumption estimation models for FIR filters. Power consumption modelswere obtained from a large number of FIRfilter syntheses using a direct form. Several curves that estimate power consumption wereextracted from these synthesis results. Then, we have evaluated the impact of polyphase decomposition on power consumption ofFIR filter and compared it with the direct form results. Some tips regarding power consumption were deduced for the polyphase

implementation form. The aim of this work is to help a system designer to select an efficient implementation for FIR in terms ofpower consumption without having to implement and synthesize the different possible solutions. The proposed method is appliedfor STMicroelectronics libraries 90 nm and 65 nm low power then validated with a use case of multistandard receiver designing.

1. Introduction

Currently, wireless technologies are widespread because oftheir flexibility of use. However, many different standards areused, and a new challenge for the communication-embeddedsystem designer is to implement multiple communicationstandard devices in order to provide easy access to in-

formation everywhere with low hardware complexity. Ingeneral, such devices include communication chains withdecimation and selector filters. Typically, multiple stan-dards communication chains are composed of an RF frontend, an over-sampler analogue to digital converter, anda cascade of decimation filters (Figure 1) [1]. The powerconsumption is an important constraint during embeddedsystem design because the design of decimation filtershas a substantial impact on the power consumption inmultistandard receivers. This work focuses on FIR filterpower consumption estimation in direct form (Figure 2).The polyphase form of FIR filters is widely recommendedfor reducing power consumption in comparison with all

possible implementation forms of these filters. To the bestof our knowledge, there has not been a clear study basedon experimental results showing how much power FIR filtersconsume. In fact, the work by Dumonteix et al. [2] is widelymentioned, and it deals only with the power consumption,area and critical path of a particular implementation of acomb filter. It was shown in this work that appropriate filter

decomposition in association with polyphase decompositioncould lead to an important significant consumption of theCIC filter.

The main objective of this work is to provide models thatevaluate FIR decimator filter power consumption in directform. These models consider the main filter parameters,which are the filter order, the input wordlength, and the co-efficient wordlength. This model was given for STM90 nmand STM60 nm low power. The operation conditions arespecified in paragraph 2.

The second objective of this paper is to study the impactof the polyphase form on the power consumption of FIRfilters. Some tips regarding the best decomposition to


2/16

2 VLSI Design

Table1: Dynamic and static power contribution using a 90 nm library of STMicroelectronics.

Filter order Dynamic power (uW) Dynamic power (%) Static power (uW) Static power (%)

13 300,9546 48,84 315,2013 51,16

22 481,3064 48,26 515,8453 51,74

31 632 46,71 721 53,29

40 722 43,54 936 56,4649 840,64 42,63 1131,1 57,37

58 973,3182 41,87 1351,3 58,13

64 1054,6 41,62 1478,9 58,38

RF filterAnalog to

digitalconverter

FIRdecimation

filter

Figure 1: General communication chain in a multistandardreceiver.

perform to save power are extracted from the synthesisresults in this work.

The ultimate aim of this study is to help filter designersdecide the best way to decompose the filter processing intostages to guarantee optimal power consumption.

This paper is organized as follows. Section 2 presentsthe specifications of the symmetric FIR filter that is used toestimate the power consumption of a decimation FIR filter.In Section 3, we present the power consumption modelsthat are obtained for the direct implementation form. InSection 4, the models are validated by using synthesis results.InSection 5, we present the implementation results for thepolyphase form usage. A use case based on established

models is presented in Section 6. Finally, conclusions aregiven inSection 7.

2. CharacterizationofDecimatorFIRFilterParameters

When designing FIR channel selection and a decimationfilter, the filter designer faces a trade-off between channelselection efficiency and filter complexity. Indeed, the designmust guarantee channel selection with the minimum com-plexity in terms of occupied area and power consumption.Depending on the communication standard, filter designerschoose the optimal filter order and coefficient length to

guarantee the required signal-to-noise ratio. The input datawordlength is typically deduced from the analogue todigital converter input signal dynamic range. Hence, for thehardware performance evaluation, the designer should con-sider the three following parameters in FIR filter powerconsumption: input wordlength, filter order, and coefficientwordlength.

To evaluate the impact of these parameters on the powerconsumption of an FIR filter, we performed a large numberof FIR filter architecture syntheses. We considered filtersorders from 7 to 64 (with a step of three). We chose toimplement both direct and polyphase forms. To reduce thehardware implementation complexity of the filter, we used

a Wallace adder tree [3, 4] to perform the addition of allmultiplication results (Figure 2).

For the polyphase decomposition, each subfilter wasimplemented as a FIR filter in the direct form. For simplicity,we chose to run the syntheses with the input wordlengthequal to 4 bits, 8 bits, 16 bits, and 32 bits. In the next section,we will show that, for intermediate values, it is possible to

interpolate the power consumption of the filter.In the same way, we choose three possible values of filtercoefficients: 4 bits, 8 bits, and 16 bits. In fact, both the filterorder and coefficient size depend on the filters mask. Weestimate that 16 bits offer enough accuracy for quantizationprocess.

The performance estimation of FIR filters was done onSTM90 nm process technology. In this 90nm process library,the static power has almost the same proportions as thedynamic power consumption (seeTable 1). For this reasonwe proposed two separate models for the dynamic and staticpower consumption. The performance estimation of FIRfilters was also done on ASIC 65 nm process technology using

a STMicroelectronics low-power library. Using this library,the power consumption of the FIR filters is reduced to thedynamic contribution because the static power is very low(seeTable 2). Using the STM 65 nm low-power technologypermits to deduce the dynamic power consumption modelthat can be verified later with the STM90 nm library process.

Design Vision of Synopsys was used to extract the perfor-mances on ASIC technology. From the experimental results,we were able to build a power consumption model of FIRfilters depending on the three main filter parameters: inputwordlength, coefficient wordlength, and filter order. Becausethe dynamic power consumption depends on frequency, themodel for dynamic power consumption obtained is also

frequency dependent.

3. PowerConsumption EstimationModels

In this section, we introduce the power consumption esti-mation models for STM65 nm and STM90 nm process tech-nologies. The STM65 nm library is low-power and operatesat 0.9 V and used in nominal case with junction temperatureof 25C. The STM90 nm library operates at 1.26 V and isused in the best case with a junction temperature of 40C.

3.1. Dynamic Power Consumption Model. For the dynamicpower consumption model we used STM 65 nm low-power


3/16

VLSI Design 3

Reg Reg Reg Reg Reg

RegRegRegRegReg

clk

Wallace tree adder

h0 h1 h2 hN

xn

yn

(a) Direct form

Reg

Reg

Reg

Wallaceaddertree

xn Fs/M

Fs/M

Fs/M

Fs/M

Fs

Fs

Fs

Fs

H0

H1

H2

HM1

yn

(b) Polyphase form

Figure2: Implementation architecture of the direct form and polyphase form of FIR filter.

Table2: Dynamic and static power contribution using a 65-nm low power library.

Filter order Dynamic Power (uW) Dynamic power (%) Static Power (uW) Static power (%)

13 169,6857 98,9 1,9 1,1

22 245,1208 98,8 3,1 1,2

31 356,7201 98,9 4,2 1,1

40 424,5542 98,7 5,7 1,3

49 497,6389 98,6 7,2 1,4

58 578,6831 98,7 8 1,3

64 628,4149 98,6 9,1 1,4

technology. In this technology, the dynamic power consump-tion is assumed to be the total power because the static poweris very low (seeTable 2). The Dynamic power consumptionis assumed to be proportional to frequency. To verify thisassumption, we performed several experiments to evaluatethe impact of the frequency constraint, provided by thelogical synthesis tool, on the occupied area and the powerconsumption of the generated design for a fixed filter order.Figure 3 confirmsthat the resources that are used to build thearchitecture for a given filter are the same regardless of the

specified frequency constraint. Hence, the logical synthesistool does not introduce any area or power optimizationregardless of the working frequency. As a consequence, thedynamic power consumption is considered linear given theconstrained frequency (Figure 3).

Following this observation, we concentrated all synthesesefforts at a fixed frequency equal to 80 MHz, which issufficient for the requirements of both GSM and UMTSstandards [5,6].

Figure 4gives the evolution of the power consumptionversus the filter order for a direct form FIR filter for differentvalues of coefficients and different inputs wordlength. Thisfigure shows that power consumption is quite linear to the

order of the filter for fixed input (4, 8, and 16 bits) andcoefficient wordlength (4, 8, and 16 bits).

Figure 5 illustrates the relationship between the powerconsumption and input wordlength for four chosen ordersand for a fixed coefficient value. For all other orders andfor the different coefficient wordlengths, the evolution of thepower consumption has the same trends. According to thesecurves, the power consumption evolution versus the inputwordlength is not linear.

However,Figure 6shows that the natural logarithm of

the power consumption is almost linear as compared withthe logarithm of the input wordlength. Hence, (1) givesthe expression of the natural logarithm of the dynamicpower consumption versus the natural logarithm of the inputwordlength. This calculation leads to relation (2), which givesthe power consumption general expression.

LN(P) = LN()+ LN(I), (1)

P= (I), (2)

whereIis the input wordlength, is the slope of the curvesin Figure 6, and LN() represents the origin value of the


4/16

4 VLSI Design

0

5000

10000

15000

20000

25000

30000

20 40 80 100 133 200 266 400

Frequency (MHz)

O13

O40

O22

O58

O31

O64

Areaoccupation(m2)

(a) Area evolution

0 50 100 150 200 250 300 350 4000

500

1000

1500

2000

2500

3000

Frequency (MHz)

Powerconsumption(W)

O13

O40O22O58O31

O64

(b) Power consumption

Figure3: Area occupation and power consumption of FIR filters in the direct form depending on frequency.

0100200300400500600700800900

1000110012001300140015001600

7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64Filter order

Powerconsumption(W)

Coef16 I4Coef16 I8

Coef16 I16Coef8 I4Coef8 I8

Coef8 I16

Coef4 I4Coef4 I8

Coef4 I16

Figure 4: Power consumption of FIR filters in the direct formdepending on input wordlength and coefficient wordlength.

same curves. Evaluating the and expressions dependingon the filter parameters and on the basis of the different

experiments and curves will lead to the establishment ofthe dynamic power consumption model. The followingsection demonstrates how each parameter of expression (2)is obtained.

We evaluate first whether parameter is independentof the filter versus expression of the parameter . Hence,according to Figure 6, all given curves LN(P) versus LN(I) arealmost parallels for a fixed coefficient value. This propertywas verified for all other orders and for all coefficientwordlength considered in this work. Consequently, the slope() of all these curves is the same regardless of the filter order.Thus, this slope () does not depend on the filter order andis only dependent on the coefficient wordlength.

0 5 10 15 20 25 30 350

200

400

600

800

1000

1200

1400

Input wordlength (bits)

Order 13Order 16

Order 19Order 22

Powerconsumption(W)

Figure 5: Power consumption of FIR filters in the direct formdepending on input wordlength forthe 65 nm technology fora fixedcoef size of 4 bits.

To evaluate the expression of as a function of coef(filter coefficient wordlength), we plot the curves given

depending on coef for fixed filter orders inFigure 7. Thesecurves are obtained from data illustrated inFigure 6.

Figure 7confirms the filter order (N). Moreover, thecurves show no linear evolution of versus coefficientwordlength (coef). The expression of as a function of coefis given in (3).

=g(coef) = a exp(b coef)+c, (3)

where a, b, and c are technology-dependent terms. Theexponential term is explained by the curves in Figure 7. Infact, the curves converge around (a+c) when the coefficientwordlength is close to 0 and then decreases very fast when thecoefficient wordlength increases to converge to the cvalue.


5/16

VLSI Design 5

1 1.5 2 2.5 3 3.54.5

5

5.5

6

6.5

7

Order 13

Order 16

Order 19

Order 22

L

N(P)

LN(I)

Figure 6: Natural logarithm of the power consumption of FIRfilters in the direct form depending on the natural logarithm of theinput wordlength for the 65 nm technology for a fixed coef size of

4 bits.

4 6 8 10 12 14 160.7

0.75

0.8

0.85

0.9

0.95

Coefficient wordlength (bits)

Order 13Order 16

Order 19

Figure 7: Evolution of depending on coefficient wordlength fora fixed input size of 4 bits and fixed filter orders for the 65 nmtechnology.

The line of equation y = c is a horizontal asymptote to thecurves inFigure 7, which explain the additive c term in (3).To set numerical value ofa,b,cfor this given technology weused Matlab curve fitting toolbox [7] and we found that,in this technology, parameters a, b, and care 0.6, 0.3, and0.7, respectively. Of course, these values are useful onlyfor the given technology; however, if we are using anothertechnology the method presented above should be repeatedat least one time as we have performed with the STM90 nmlater.

To evaluate parameter of (2), we used the same ap-proach used for expression extraction. Hence, first, thedependency ofversus filter order (N)was evaluated. Then,

13 14 15 16 17 18 19 20 21 22

30

40

50

60

70

80

90

100

110

Filter order

Coef = 4

Coef = 8

Coef = 16

20

Figure8: Evolution ofdepending on filter order fixed coefficient

wordlength for the 65 nm technology.

Figure 9: Evolution of slope depending on the coefficientwordlength for fixed filter order for the 65 nm technology.

the dependency versus coefficient wordlength (coef) wasfound.

Curves inFigure 8illustrate the dependency of param-eter on the filter orders and for three fixed coefficient

wordlengths. To obtain these curves, we extracted theorigin values fromFigure 7. These origin values representthe normal logarithm of the parameter for each fixedcoefficient size. Then, we plotted the exponential of thesevalues. The curves inFigure 8show that is almost linearversus filter order for any coefficient size. Hence, we nextevaluated the dependency of the slope of depending oncoefficient wordlength, as shown inFigure 9. According tothis figure, the slope of is also linear versus coef. Asa consequence, the relation (4) is deduced to model theevolution ofversusNand coef.

= N coef +, (4)


6/16

6 VLSI Design

0 5 10 15 20 25 30 350

500

1000

1500

2000

2500

3000

3500

4000

4500

5000


Order 13

Order 16

Order 19

Order 22

Powerconsu

mption(W)

Figure10: Relationship between the dynamic power consumptionand input wordlength for the 90nm technology for a fixed

coeffi

cient wordlength.

where and are constants and depend on the technologyconsidered. Experiments show that in the 65 nm technology,the values of,are 0.2 and 1, respectively.

Hence, using (3) and (4), equation (5) gives the dynamicpower consumption evolution versus filter parameters andnormalized frequency for a direct form (Figure 2).

PN,coef, I,f

=

N

coef +

I(aexp(bcoef)+c)

f

f0,

(5)

where f0 is the frequency used during the syntheses processand is equal to 80 MHz. When the filter order is zero, thepower consumption should be zero as well. This conditionis guaranteed because power is directly proportional to filterorder (N).

On the other hand, when the coefficient length (coef) iszero, the FIR filter is composed ofNregisters andN/2 adders.In this case, the power consumption becomes a constantmultiplying the filter order (N).

3.1.1. Verification of the Dynamic Power Consumption forSTM90-nm. To verify whether the model of (4) is compliant

with the amount of dynamic power consumption in theSTM90 nm technology, we repeated all of the synthesesusing the new library. The same parameter values wereused for the experiments. The same conclusions regardingpower consumption versus input wordlength were noticed.As shown in Figure 10, we also verified that the naturallogarithm of the power is proportional to the naturallogarithm of the input wordlength regardless of the filterorder and coefficient size. Hence, (1) and (2) are still truefor the 90 nm technology.

According toFigure 11, the independency of the param-eter in (2) regarding filter orderNis still verified. Figure12 shows the relationship between and the coefficient

1 1.5 2 2.5 3 3.54.5

5

5.5

6

6.5

7

Order 13

Order 16

Order 19

Order 22

LN(

P)

LN(I)

Figure 11: Relationship between the logarithm of dynamic power

consumption and the logarithm of the input wordlength for the90 nm technology for a fixed coefficient wordlength.

4 6 8 10 12 14 160.7

0.72

0.74

0.76

0.78

0.8

0.82

0.84

0.86

0.88

Coef (bits)

Figure12: Evolution ofdepending on the coefficient wordlengthfor fixed filter orders.

wordlength. The same trends observed inFigure 7are ob-served inFigure 13. Hence, the expression ofgiven in (3) is

verified in the 90 nm technology.Figures 13 and 14 give the evolution of parameter

(given in (3)) regarding filter order and the slope ofversuscoefficient wordlength, respectively. According to the twofigures, (4), which gives expression ofregarding filter orderand coefficient wordlength, is verified.

As a consequence, the model of (5) is applicable for theevaluation of the dynamic power consumption in the 90 nmtechnology. With the same manner, to set numerical valueofa, b, c for this given technology we used Matlab curvefitting toolbox [7] and we found that for the 90 nm process,parameters , , a, b, and cwere equal to 0.5, 4.5, 0.5, 0.2,and 0.6, respectively.


7/16

VLSI Design 7

10 20 30 40 50 60 700

100

200

300

400

500

600

700

Filter order

Coef= 4

Coef= 16

Figure 13: Evolution of depending on filter orders for a fixedcoefficient wordlength.

Figure14: Evolution of the slope ofdepending on the coefficientwordlength.

3.2. Static Power Consumption Model. TheTable 1illustratesthe dynamic and static power contribution of the power

consumption for the 90-nm process technology. It is clearfrom the table that static power cannot be neglected in thistechnology.

To establish a static power consumption model, we evalu-ated the evolution of the static power regarding filter param-eters. In particular, we noticed a nonlinear relationshipbetween the static power and input wordlength (see Figure15). The second observation concerns the linearity of thenatural logarithm of the static power versus the naturallogarithm of the input wordlength (Figure 16). Hence, forthe dynamic power consumption contribution, (1) and (2)could be used as general equation forms for the static powerconsumption of symmetric FIR filters.

0 5 10 15 20 25 30 350

1000

2000

3000

4000

5000

6000


Order 13

Order 22

Order 31

Order 40

Order 49

Staticpowercon

sumption(W)

Figure 15: Relationship between the static power consumption andinput wordlength for the 90 nm technology for a fixed coefficientwordlength.

1 1.5 2 2.5 3 3.56

6.5

7

7.5

8

8.5

9

Order 13

Order 22

Order 31

Order 40

Order 49

Log (I)

Log(P)

Figure 16: Relationship between the logarithm of the static powerconsumption and the logarithm of the input wordlength for the 90-nm technology for a fixed coefficient wordlength.

To establish expressions of the and terms (in (2)), thedynamic power consumption modeling was followed. Hence,starting from the fact thatis filter order independent (sincethe curves are parallel inFigure 16), we plot the evolution ofversus coefficient wordlength inFigure 17.

After analyzing the curves inFigure 17, (6) fits the bestcurve evolution. Indeed, the exponential term is explainedby the rapid decrease when small coefficient values areconsidered. The linear term is added because of the very slowincrease when the coefficient wordlength increases.

= exp(d coef)+e coef +g, (6)


8/16

8 VLSI Design

4 6 8 10 12 14 160.68

0.7

0.72

0.74

0.76

0.78

0.8

Coefficient wordlength

Figure17: Evolution of depending on the coefficient wordlengthfor fixed filter orders.

10 15 20 25 30 35 40

50

100

150

200

250

300

350

400

450

500

550

Filter order

Coef= 4Coef= 8

Coef= 16

Figure 18: Evolution of depending on filter orders for a fixedcoefficient wordlength.

where d, e, and g are technology dependent parameters.Using Matlab, we shown that, when using a STMicroelec-

tronics library, the parameters are equal to 0.11, 0.03, and0.04.

Parameter was then evaluated depending first on thefilter order and then on the coefficient wordlength. Figure18 shows the relationship between and N. Hence, wefound an almost linear relationship of this parameter versusfilter order. In the second step, the slope of was analyzedaccording to coefficient wordlength (seeFigure 19). Matlabverified that (7) fits the evolution of the slope ofillustratedinFigure 18.

slope =

a exp(b coef)+c, (7)

4 6 8 10 12 14 16

7

8

9

10

11

12

13

Coef wordlength

3

4

5

6

0 2

s

lope

Figure19: Evolution of theslope depending on filter orders for afixed coefficient wordlength.

where a, b, c, and are parameters depending on thetechnology. According to Matlab, when using the 90 nmSTMicroelectronics library, the parameters are equal to 0.29,0.4, 0.077, and 1, respectively. Finally,is expressed in (8):

= +

a exp(b coef)+cN. (8)

As a consequence, the general expression of the static powerconsumption for a direct form FIR filter (Figure 2) in the90 nm technology could be written as shown in (9).

Pstat(N, coef, I) = N

a exp(b coef)+c+

I(exp(dcoef)+ecoef+g),(9)

where technology dependent parametersa, b, c, d, e, and gare equal to 0.29, 0.4, 0.077, 0.11, 0.03, and 0.04, respectively.Parameters and depend also on the technology and areequal to 1.2 and 1.2, respectively, for the current model.

4. Validation of thePowerConsumptionModels

This part presents the validation of the models obtained.Hence, we present different figures comparing the synthesesvalues of power consumption and the results deduced

from established models. The aim of the comparison is todemonstrate first that the power consumption trends arerespected by the models. In fact, the power estimation valuefor a given design is not really important and the objective isto prove that decisions concerning the choice of the suitableparameters, which reduces the power consumption, are notmodified using the models. For this purpose, a parametercalled the deviation is calculated. This parameter givesthe difference for two fixed filter parameters (which are thefilter order and coefficient wordlength or input wordlength)between the power consumption value when varying thethird filter parameter (which is the coefficient wordlengthor input wordlength). Indeed, having a similar deviation


9/16


10/16

10 VLSI Design

0

500

1000

1500

2000

2500

3000

3500

4000

4500

13 22 31 40 49 58 64

Filter order

d c8e4d c8e16

model (d c8e8)

d c8e8

model (d c8e4)

model (d c8e16)

Dynamicpower(W)

(a) coef= 8

0

500

1000

1500

2000

2500

3000

3500

4000

4500

13 22 31 40 49 58 64

Filter order

d c16e4

d c16e8d c16e16

model (d c16e4)

model (d c16e8)

Dynamicp

ower(W)

(b) coef= 16

Figure22: Comparison of model dynamic power results to experimental values for the 90 nm technology for a fixed coefficient wordlength.

0

500

1000

1500

2000

2500

3000

13 22 31 40 49 58 64

Filter order

Deviation(W)

d1

d2

d3

m d1 m d2

m d3

Figure 23: Deviation parameter for model and estimation toolpower results for coef= 8.

almost equal to the deviation measured for the experimentaldynamic power estimation (Figure 23).

4.3. Validation of the Static Power Consumption Model for the90 nm Process Technology. For the dynamic contribution, wevalidated the static power model as given in (9). Figure 24shows the comparison between the results extracted fromthe model and the experimental values for static powerconsumption and for fixed input wordlength.

In the same way, we confirm that the proposed modelfor static power consumption gives static power values veryclose to the values given by the estimation tool. The modelalso does not modify trends of power and does not modifydecisions depending on the filter parameters because thedeviations are almost equal (seeFigure 25).

5. PowerConsumption EstimationTips forPolyphase ImplementationForm of FIR Filters

5.1. Results of Polyphase Implementation Form in the 65 nmTechnology. Polyphase implementations of FIR filters wereperformed with different input wordlengths, coefficient sizes,and filter orders, which lead to the same observations

illustrated inFigure 26. This figure compares the direct formimplementation of different FIR filter orders versus theirpolyphase decomposition.

The comparison is made for an input size equal to 4 bitsand includes decimation values of 2, 4, 8, and 16. AccordingtoFigure 26, it is clear that polyphase decomposition reducesthe power consumption for any decimation factor.

In fact, the following observations are clear.

(i) Decomposition into two stages allows power reduc-tion rates that increase with filter order. The reduc-tion rates are between 20% and 40% in comparisonwith the direct form implementation.

(ii) Decomposition into four stages allows reductionrates between 45% and 60% compared with the directform implementation.

(iii) Decomposition into 8, 16, and 32 stages allows reduc-tion rates between 60% and 70% for 8 stages basedarchitecture, 60% and 75% for the 16 stages, and 55%and 75% for the 32 stages.

As a consequence, polyphase decomposition is always benefi-cial in terms of power consumption reduction. However, thereduction average depends on the decimation factor that isused for a given filter order. Equation (10) gives the relationbetween the decimation factor and filter order up to 128 to


11/16

VLSI Design 11

Staticpow

er(W)

s c4e4 s c8e4

s c16e4 Model (s c8e4)

Model (s c4e4) Model (s c16e4)

0

500

1000

1500

2000

2500

3000

13 22 31 40 49 58 64

Filter order

(a) I= 4

s c4E8 s c8e8

s c16e8 Model (s c4e8)

Model (s c8e8) Model (s c16e8)

Staticpow

er(W)

0

500

1000

1500

2000

2500

3000

3500

4000

4500

13 22 31 40 49 58 64

Filter order

(b) I= 8

Figure24: Comparison of the model static power results with experimental values for the 90 nm technology for a fixed Input wordlength.

0

100

200

300

400

500

600

700

800

900

1000

13 22 31 40 49 58 64

Deviation(W)

Filter order

d1 d2 d3

m d1 m d2 m d3

Figure 25: Deviation parameter for model and estimation toolpower results for an input size of 8 bits.

offer the best power consumption reduction. It is estimatedthat order 128 is sufficiently large for channel selection filters.

Mopt =

8 si 8


12/16

12 VLSI Design

0

200

400

600

800

1000

1200

13 22 34 43 49 59 61

Filter order

Dynamicpowerconsumption(W)

d m2

d m4

d m8

d m16 Direct form

Figure 27: Dynamic power evolution of poly phase form imple-mentations of a symmetric FIR filter for the 90 nm technology.

0

500

1000

1500

2000

2500

13 22 34 43 49 59 61

Filter order

Staticpowerconsumption(W)

s m2

s m8

Direct form

s m4

s m16

Figure28: Static power evolution of poly phase form implementa-tions of a symmetric FIR filter for the 90 nm technology.

InFigure 29, the total power consumption results for theFIR filter considered are plotted and compared to the powerconsumption results of the direct form. According to thisfigure and considering the total power consumption criteria,the polyphase implementation could lead in some cases to apower consumption increase in comparison with the direct

form power consumption results.However, decimation by a factor of 2 and 4 always re-duces the total power consumption in comparison with thedirect form, with advantageous reduction for a decimationfactor of 4.

6. Comparison of PowerConsumption ofFilteringArchitectures forGSMandUMTSStandardsUsingEstablishedModels.

In this section, we examine the power consumption estima-tion of different decimation filter solutions for multistandardreceiver supporting UMTS and GSM standards. The aim

0

500

1000

1500

2000

2500

3000

13 22 34 43 49 59 61

Filter order

Totalpowerc

onsumption(W)

t m2 t m4t m8 t m16

Pt direct

Figure 29: Total power evolution of poly phase form implementa-tions of a symmetric FIR filter for the 90 nm technology.

of this section is to demonstrate how the models couldhelp filter designer make the correct choice regarding filterarchitecture to reduce the power consumption.

For this purpose, three different architectures suitable formultistandard receivers are analyzed (Figure 30).

The parameters of the different filters building each ar-chitecture are obtained using the specification methodologydescribed in [8]. The multistandard receiver architecture andparameters considered in this work are calculated in [9].Because filtering performances are out of the scope of thispaper, only the description of the filter parameters is given.

6.1. Filtering Solution Description for UMTS and GSMStandards. This paragraph gives the details of the differentarchitectures for both standards. Tables3 and 4 summarizethe parameters of the different stages in the case of UMTSand GSM standards, respectively.

6.1.1. 2-Stage Architecture (Arch0). For both UMTS andGSM standards, the first filter is composed of a cascade of 5-stage CIC filters. It performs decimation by a factor of 12 or4 for GSM and UMTS standards, respectively. The recursivearchitecture of the CIC filter allows the programmabilityof the filter depending on the selected standard. The input

wordlength of the filter is equal to 6 bits at its output; thewordlength is 16 bits for the UMTS standard and is equal to26 bits in the case of the GSM standard. The second filter ofthe 2-stage architecture is a symmetric FIR filter of order 45for UMTS, where 12 bits are necessary for the quantificationof the filter coefficients for this standard. In the case of GSM,the required filter order is 83, and 12 bits are sufficient for thequantization of coefficients.

6.1.2. 3-Stage Architecture (Arch1). In this architecture, thelast filter of the 2-stage architecture is split into a half-bandfilter and an FIR filter performing, each one a decimationby a factor of 2. The half-band filter has an order equal


13/16

VLSI Design 13

CIC filter M1 FIR filter M2

(a)

CIC filter M3Halfband

filter 2 FIR filter 2

(b)

CIC filter 2 CIC filter 2 Halfband

filter 2 FIR filter 2

(c)

Figure30: Decimation filter architectures for GSM and UMTS standards.

Table3: Specification of the filters in proposed architectures for the GSM standard.

Architecture Specification CIC filter HB1filter HB2 filter FIR filter

2 stages

Order 5 N/A N/A 83

Input wordlength 6 N/A N/A 17

Coefficients wordlength N/A N/A N/A 12

Decimation factor 12 N/A N/A 4

3 stages

Order 5 20 N/A 45

Input word-length 6 17 N/A 17

Coefficients wordlength N/A 11 N/A 11

Decimation factor 12 2 N/A 2

4 stages

Order 6 8 10 34

Input wordlength 6 17 17 17

Coefficients wordlength N/A 10 11 10

Decimation factor 6 2 2 2

Table4: Specification of the filters in proposed architectures for the UMTS standard.

Architecture Specification CIC filter CIC filter HB filter FIR filter

2 stages

Order 5 N/A N/A 57

Input wordlength 6 N/A N/A 15

Coefficients wordlength N/A N/A N/A 12

Decimati on factor 4 N/A N/A 4

3 stages

Order 5 N/A 14 20

Input wordlength 6 N/A 15 15

Coefficients wordlength N/A N/A 10 12

Decimation factor 4 N/A 2 24 stages

Order 5 4 10 20

Input wordlength 6 11 15 15

Coefficients wordlength N/A N/A 9 12

Decimation factor 2 2 2 2

to 14 for UMTS and 20 for GSM. The filters coefficientslength is fixed to 10 bits and 11 bits for UMTS and GSMstandards, respectively. The last filter has an order of 20 forUMTS and an order equal to 45 for GSM. For both standards,coefficients are quantized with 12 bits.

6.1.3. 4-Stage Architecture (Arch2). In the case of the GSMstandard, the 4-stage architecture is based on a CIC filter oforder 6 in its recursive form, followed by a cascade of 2 half-band filters, each one performing a decimation by a factorof 2. The last stage is composed of a symmetric FIR filter,


14/16

14 VLSI Design

Table5: Power consumption evaluation according to models in the 65 nm low-power technology.

2-stage Archpower (W)



GSM standard

Filter 1 (cic filter) 226 226 264,3

Filter 2(halfband)

N/A 42,5 32,6

Filter 3(halfband)

N/A N/A 20,4

Filter 4 (FIR) 208,6 47,8 34,7

Total GSM 434,6 336,3 352

UMTS standard

Filter 1 (cic filter) N/A N/A N/A


Filter 3(halfband)

N/A 54 52

Filter 4 (FIR) 303 56 56

Total UMTS 303 110 108

Table6: Power consumption evaluation according to models for the 90 nm technology.

2-stageArchpower (W)



GSM standard

Filter 1 (cic filter) 967,8 967,8 2796,8

Filter 2(halfband)

N/A 119,5 + 1818,8 93,7 + 720

Filter 3(halfband)

N/A N/A 59,7 + 909,4

Filter 4 (FIR) 575,3 + 8660 134,4 + 4092,4 99,5 + 3060,75

Total GSM 10203,1 7132 7739,8

UMTS standard



Filter 3(halfband)

N/A 145 + 827 142 + 818

Filter 4 (FIR) 823 + 4000 147 + 1600 147 + 1600

Total UMTS 4823 2719 2707

completing the decimation by 2 and the channel selection.The two half-band filters have orders equal to 6 and 10, andtheir coefficients are quantized on 10 and 11 bits, respectively.The last FIR filter has an order equal to 34, and we found that10 bits are required for the quantization of coefficients.

For the UMTS standard, the CIC filter used in the 3-stage architecture is split into a cascade of 2 CIC filters. The

number of required stages is 5 for the first CIC filter and 4for the second. Each filter performs decimation by a factor of2. The output size of the first CIC filter is equal to 11 bits.The third filter is a half-band filter of order 10 and has acoefficient size equal to 9 bits. Finally, a symmetric FIR filterof order 20 with coefficients quantized on 12 bits completesthe selection. As explained before, the input wordlength isconsidered equal to 15 bits, unless for the first or secondstage, for which the input wordlength is equal to 6 bits or11 bits, respectively.

6.2. Comparison to Implementation Results. On the basis ofthe filter parameters and the established models given in

(5) and (10), we estimated the power consumption of eachsub-block in the filtering architectures for both the 65 nmand 90 nm technologies. Table 5 gives power consumptionestimation results for GSM and UMTS standards for the65 nm technology. The results for both standards for the90 nm technology are given inTable 6.

For the particular case of CIC filters, the power consump-

tion comparison is performed following the work in [2].Indeed, the authors in [2] studied the power consumptionof fixed order CIC filters depending on their implementationarchitecture. In the case of the GSM standard and because thepower consumption of CIC filters of different orders is notincluded in the work in [2], the power consumption resultsconcern the FIR implementation form of the CIC filters.

According to Tables3and4, it is clear that, for both tech-nologies (65 nm and 90 nm), the architecture that presentsthe optimal power consumption in the case of the GSMstandard is the 3-stage architecture.

For the UMTS standard, the power consumption valuesobtained from models are very similar in the case of 3-stage


15/16

VLSI Design 15

Table7: Power consumption evaluation according to the prime-power tool for the 65-nm technology for UMTS and GSMstandards.

2 stages 3 stages 4 stages

GSM standard 280,6 231 260

UMTS standard 171 157 149

Table8: Power consumption evaluation according to the prime-power tool for the 90nm technology for UMTS and GSMstandards.

2 stages 3 stages 4 stages

90nm GSM 751 640 731

UMTS 5353 4231 4441

65nm GSM 280,6 231 260

UMTS 171 157 149

and 4-stage architectures. If the estimation of the power is

considered, which is done in [2], for the comparison of thepower consumption of the CIC filters, the power values in-crease, but the trends are not modified.

It is, however, important to notice that the power due tothe clock tree and stage connection is not considered in theevaluation. Thus, the power consumption for 4 stages shouldbe considered because it is composed of more stages andshould increase compared to the 3-stage architectures, in par-ticular for the 90 nm technology. Hence, it can be concludedthat the 3-stage architecture is also more advantageous interms of power consumption for the UMTS standard.

To validate these results experimentally, VHDL code of allarchitectures considered was built. The results of the power

estimated after logic syntheses are given in Tables7and8forthe 65 nm and 90 nm technologies, respectively. These tablesconfirm the conclusions obtained from established models.

7. Conclusion

This paper presented power consumption evaluation modelsof direct form FIR filters (Figure 2) used for decimation.Both models for dynamic and static power contributionsare proposed in this work in STM65 nm low-power andSTM90 nm ASIC technology. The proposed models are highlevel models, which estimate the dynamic and static powerconsumption of the FIR filter depending on three filter

parameters, which are the input wordlength, coefficientwordlength, and filter order. The aim of these models isto help the system designer compare, at the system level,different filter architectures in terms of power consumptionwithout having to implement the different filters and per-form syntheses. However, this method should be verified forany new library different form the used ones in this work.In the second step, the effect of polyphase decompositionfor FIR decimator filters was evaluated in two differenttechnologies. We found that polyphase decomposition allowsgood dynamic power reduction regarding the direct formimplementation. We observed that the dynamic powerreduction could reach 75% in some cases. To help the

system designer choose the best decimation factor in termsof power consumption, relation (3) was given between thedecimation factor and the filter order. However, the staticpower consumption increases when performing polyphasedecomposition. When the static contribution is important(as with the 90 nm technology library considered), the total

power consumption can increase in some cases in compari-son with the direct form power consumption.Finally, a case study concerning the UMTS and GSM

standards was presented. We performed a comparison be-tween three filter architectures. The power estimation basedon proposed models helped choose the suitable architecturefor power consumption optimization, and the result wasconfirmed by filter implementations.

References

[1] Y. S. Poberezhskiy and G. Y. Poberezhskiy, Flexible analog frontends of reconfigurable radios based on sampling and recon-struction with internal filtering, Eurasip Journal on WirelessCommunications and Networking, vol. 2005, no. 3, pp. 364381,2005.

[2] Y. Dumonteix, H. Aboushady, H. Mehrez, and M. Louerat,Low power comb decimation filter using polyphase decompo-sition for mono-bit analog-to-digital converters, IEEE Trans-actions on Circuits and Systems II: Analog and Digital SignalProcessing, vol. 48, no. 10, pp. 898903, 2001.

[3] M. O. Lakshmanan and M. Ali, High performance parallelmultiplier using Wallace-Booth algorithm, in Proceedings of theIEEE International Conference on Semiconductor Electronics, pp.433436, 2002.

[4] H. Lee, A power-aware scalable pipelined booth multiplier, inProceedings of the IEEE International Systems-On-Chip Confer-ence, pp. 123126, September 2004.

[5] Universal Mobile Telecommunication System (UMTS), UERadio Transmission and Reception (FDD), 3GPP TS.101 ver-sion 5.2.0 Release 5. ETSI, 2002.

[6] Radio Transmission and Reception GSM 05.05. ETSI, 1996.[7] Matlab curve fitting toolbox data sheet, http://www.mathworks

.com/products/curvefitting/.[8] R. Barrak, A. Ghazel, and F. Ghannouchi, Design of sampling-

based downconversion stage for multistandard RF subsamplingreceiver, in Proceedings of the 13th IEEE International Confer-ence on Electronics, Circuits and Systems (ICECS 06) , pp. 577580, December 2006.

[9] A. Ghazel, L. Naviner, and K. Grati, On design and imple-mentation of a decimation filter for multistandard wirelesstransceivers, IEEE Transactions on Wireless Communications,

vol. 1, no. 4, pp. 558562, 2002.


16/16

Submit your manuscripts at

http://www.hindawi.com

Documents

Power Reduction in FIR