A Framework for Dynamic Energy Efficiency and

Temperature Management(DEETM)

Michael Huang, Jose Renau, Seung-Moon Yoo, Josep TorrellasUniversity of Illinois at Urbana-Champaign

http://iacoma.cs.uiuc.edu/flexram

Micro’33, Dec 2000 2

Motivation

Increasingly high power consumption– High temperature– Inefficient use of energy

Limitations of existing approaches– Static optimizations – Coarse grain dynamic management (DPM)– Inefficient temperature control (sleep)– Independent targets: temp, energy

efficiency

Micro’33, Dec 2000 3

Goal

Temperature: enforce limit while minimizing slowdown

Energy efficiency: maximize energy saving while exploiting performance slack

Unified framework for Dynamic Energy Efficiency and Temperature Management (DEETM)

Micro’33, Dec 2000 4

Contribution: DTEEM Framework

Existing limitations:

– static application

– coarse grain dynamic

– inefficient techniques

– independent targets: temperature

control energy efficiency

DEETM approach:

– multiple techniques

– dynamic

– fine grain

– order techniques for maximum efficiency

– unified target

Micro’33, Dec 2000 5

DEETM Framework

Monitors temperature & execution slack

Runs decision algorithm periodically: {Thermal, Slack} components

Activates low-power techniques – dynamically– incrementally– in prioritized order

Micro’33, Dec 2000 6

Techniques

Instruction filter cache

Data cache subbanking

Voltage scaling

Voltage scaling: DRAM

only

Light sleep

Micro’33, Dec 2000 7

Instruction Filter Cache

Filter Cache

Instruction Memory

Processor

Filter Cache

L1 Cache

ProcessorHigh power mode:

Time: 1

Energy: E Filter Cache

Instruction Memory

ProcessorHigh power mode:

Time: 1

Energy: E Filter Cache

Instruction Memory

Processor

Low power mode:

Time: 1 + mr*1

Energy: e + mr*E

Micro’33, Dec 2000 8

Techniques

Instruction filter cache

Data cache subbanking

Voltage scaling

Voltage scaling: DRAM

only

Light sleep

Micro’33, Dec 2000 9

Chip Environment

Processor-in-memory

64 lean processors– 2-issue static

Optimized memory hierarchy

Integrated thermal sensors and instruction counter

Processor

D-CacheI-Cache

DRAM Bank

Instruction Memory

ProcessorDRAM Bank

Interconnect

Micro’33, Dec 2000 10

Individual Techniques-E*D

IFilterSubBankVoltFreq

0.7

Nor

mal

ized

Ene

rgy-

Del

ay P

rodu

ct1.6

1.5

1.4

1.3

1.2

1.1

1.0

0.9

0.8

Normalized Average Power1.00.90.80.70.60.5

MemVoltSleep

Micro’33, Dec 2000 11

Combinations - E*D1.0

0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

0.9

0.8

0.7

0.6

0.5

Normalized Average Power

Nor

mal

ized

Ene

rgy-

Del

ay P

rodu

ct

VoltFreq

IFilterGradual

Micro’33, Dec 2000 12

Summary

Techniques ordered by efficiency (same order apply to both targets)

System applies techniques in order, dynamically and incrementally

Micro’33, Dec 2000 13

Thermal Algorithm

Tem p > MaxTem p

Tem p < M inTem p

Yes

No

Add the nextm ost effic ient

technique

Rem ove the lasttechniqueYes

Macrocycle

Micro’33, Dec 2000 14

Temperature Control - Limit

*1

* %25%75 iii PowerPowerPower

0

10

20

30

40

50

60

70

80

90

0.6 0.8 1 1.2Normalized Power* Limit

Mac

rocy

cle

Ove

r lim

it (%

)

Uncontrolled

Gradual

Micro’33, Dec 2000 15

Temperature Control - Slowdown

0

20

40

60

80

100

0.6 0.8 1 1.2

Normalized Power* Limit

Slow

dow

n (%

)

Sleep

Gradual

Micro’33, Dec 2000 16

Slack Algorithm

MacrocycleTest baseline IPC

Add technique; Testnew effective IPC*

Slowdownenough?

No

Adjust duty cycle oftechnique if necessary

YesEffective IPC*: frequency-adjusted

Microcycle

Every Macrocycle (HW/OS)

The First Few Microcycles (HW)

Micro’33, Dec 2000 17

Slack - Energy Consumption

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 5 10 15 20 25 30 35 40 45 50Slack (%)

Nor

mal

ized

Ene

rgy

E*D*D = Constant

Gradual

Micro’33, Dec 2000 18

Slack Misprediction

0.85

0.9

0.95

1

1.05

1.1

1.15

0 5 10 15 20 25 30 35 40 45 50Slack (%)

Use

d Sl

ack/

Slac

k

GradualPerfect

Micro’33, Dec 2000 19

Related Issues

Algorithm interaction

Selecting macrocycle

Handling Thermal Crisis situation

Reducing technique activation overhead

Flexible technique ordering

Hardware vs. software implementation

Micro’33, Dec 2000 20

Conclusions

Effective & efficient temperature control– very few macrocycles still over limit – 27% longer execution vs. 98% (by

sleeping)

Efficient & accurate fine-grain exploitation of execution slack– 5% slack 27% energy saved– small slack misprediction