System Level Power- Performance Trade-Offs in Embedded...

Preview:

Citation preview

October 2002 ISSS

1

1Center for Research on Embedded Systems and Technology (CREST), http://crest.ece.gatech.edu

2Assistant 3Professor, 1Electrical and Computer Engineering2Adjunct Assistant Professor, College of ComputingGeorgia Institute of Technology, Atlanta, GA USA

44Tallin Technical University, Tallin, Estonia

System Level Power-Performance Trade-Offs in Embedded Systems Using Voltage and Frequency Scaling of Off-Chip Buses and Memory

KiranKiran PuttaswamyPuttaswamy11, , KyuKyu--Won ChoiWon Choi11, Jun , Jun CheolCheol ParkPark11,,Vincent J. MooneyVincent J. Mooney IIIIII1,1,22, , AbhijitAbhijit ChatterjeeChatterjee1,1,33 and and Peeter Peeter EllerveeEllervee44

{{kiranpkiranp, , kwchoikwchoi, , jcparkjcpark, chat, , chat, mooneymooney}@ece.gatech.ed}@ece.gatech.eduulrvlrv@cc.@cc.ttuttu..eeee

2Hardware/Software Hardware/Software CodesignCodesign Group, http://Group, http://codesigncodesign..eceece..gatechgatech..eduedu

October 2002 ISSS

2

Overview

• Introduction• Motivation• Contribution• Framework• Methodology• Results

October 2002 ISSS

3

Introduction

• Embedded Systems –essential components of living

• Constraining Factor: Power

October 2002 ISSS

4

Motivation

• Limited Battery Capacity

• Battery Energy Supplying Characteristic10 mA , 1.5 volts = 1000 hours100 mA, 1.5 volts = 80 hours

October 2002 ISSS

5

Previous Work

• Three broad approaches to memory optimization for power/energy reduction– Cache optimizations– Memory access reduction (especially of off-chip memory)– Memory sizing/structuring and memory intensive voltage

scaling

October 2002 ISSS

6

Our Contribution

• Combination of an architectural technique (store buffer) and a circuit level technique (voltage and frequency scaling) to realize savings in both power and energy in an embedded system composed of an ARM-like processor chip plus a separate memory chip

• System savings in power from 28% to 36%• System savings in energy from 13% to 35%

October 2002 ISSS

7

Computation Part of an Embedded System

CPU

Datacache

Instruction cache

Off-ChipMemory

64

96

32

32

October 2002 ISSS

8

Power Models

• Verilog RTL model for processor (excluding caches)• Compaq Personal Server PCB Board called “Skiff”• Analytical memory model for caches and off-chip

memory

October 2002 ISSS

9

Framework

Benchmark Programs ( c )

Off-Chip Memory Power Model

Toggle Rate Generation

MARS Simulator

VHX Translation

Processor Core Power Model Off-chip Bus Power Model

System Level Power

October 2002 ISSS

10

Wither the power?• Computation in system

– MARS processor (U. Michigan, www.eecs.umich.edu/~jringenb/power)• ~30K lines Verilog

– synthesized using TSMC .25u std. cell lib. from LEDA Systems

• 4KB Icache, 4KB Dcache

– 0.5MB SRAM memory chip (L2)

• Approximately 50% of the power consumed by processor chip (excluding I/O pads and drivers)

• 50% of the power consumed to drive L2 memory: the 0.5MB memory chip + PCB bus + I/O pads/drivers

• => reduce power to drive L2 memory by 60%, system power reduced 30%

October 2002 ISSS

11

3.3 V -> 2V, SRAMdelay doubles, powerreduces up to 66%

Use TSMC 0.25 utech. param. from

MOSIS

October 2002 ISSS

12

October 2002 ISSS

13

Embedded System

CPU

Datacache

Instruction cache

Off-ChipMemory

64

96

32

32

October 2002 ISSS

14

Embedded System (with Store Buffer)

CPU

Datacache

Instruction cache

Off-ChipMemory

64

96

32

32

Store buffer

October 2002 ISSS

15

Methodology

• Voltage/frequency scaling of L2 memory accesses• Store buffer technique

October 2002 ISSS

16

Voltage/Frequency Scaling

Processor Off-chip Memory

Off-chip Buses

3.3 Volts, 100 Mhz2.75 Volts, 100 Mhz

3.3 V, 100 Mhz

2 Volts, 50 Mhz

2 Volts, 50 Mhz

October 2002 ISSS

17

October 2002 ISSS

18

October 2002 ISSS

19

October 2002 ISSS

20

Conclusion

• Reduction in both power and energy – For an ARM-like processor chip plus a separate memory chip:– System savings in power from 28% to 36%– System savings in energy from 13% to 35%– Increase in execution time from 1% to 29%

• Possible technique for power modulation by user/application

Recommended