Practical Strategies for Power-Efficient Computing Technologies Karim Al-Sheraidah December 8 th 2011.

Practical Strategies forPower-Efficient ComputingTechnologies

Karim Al-Sheraidah

December 8th 2011

2

Overview

Survey of Power reduction techniques ~8x improvement in power efficiency No performance lose Voltage Scaling Optimum VDD = 0.5V

IBM Blue Gene system

3

Introduction

The Regime of interest

4

Introduction cont…

Pactive = Ceff V2 ƒ + Ileak V

Ceff is V dependent

Ceff V2 ∞ V2.5

ƒ is linearly V dependent

ƒ = α(V – V0) V0 ≈ 0.25V

Pactive = αCeff V2 (V – V0) + Ileak V ∞ V3

5

The Case for Voltage Scaling

Departing from scaling theory

6

The Case for Voltage Scaling

Optimum VDD = 0.5v

7

The Case for Voltage Scaling cont…

Optimum VDD = 0.5v

8

Enablement (1)

Operating Margin improvement

9

Enablement (2)

Low variability devices

ET-SOI Fin-FET

10

Enablement (3)

Digital Noise

Resistive: dVR/VDD = IR/VDD ∞ ( VDD – VT)1.5/VDD

Capacitive: dVC/VDD = [CaggVDD/(Cagg + Cvic)]/VDD = Cagg/(Cagg + Cvic)

Inductive: dVL/VDD = [ L ∂I/∂t ]/VDD ∞ (L I)/(VDDҭ) ∞ (VDD – V0)( VDD – VT)1.5/VDD

11

Enablement (4)

On-Chip Power System

12

Case study (IBM Blue Gene)

- Top500 HPC from 2004 to 2007

- Operating at 850MHz

- Performance of up to 13.9Tflop

- 4096 parallel processor cores

- Three chip voltage bins

13

Conclusion

- Power efficiency through voltage scaling.

- Optimum VDD = 0.5v.

- lowering of variability.

- Increasing margin.

- Massive parallelism.

- High integration.