# school of eece newcastle university ... 1 energy-modulated computing alex yakovlev school of eece...

Post on 14-Mar-2020

0 views

Embed Size (px)

TRANSCRIPT

1

Energy-Modulated Computing

Alex Yakovlev

School of EECE

Newcastle University

holistic e n e r g y h a r v e s t i n g

EACO 19 October 2011

2

Outline

• Introduction: better resource awareness

• Energy-uncertainty-QoS interplay

• Power-adaptive computing

• System design aspects

–Power proportional and power efficient systems

–Experiment in power-proportionality

– System Design for energy-harvester power supply

• Circuits for power-adaptive systems

–SRAM

–Voltage sensors

–Novel power electronics

• Conclusion

3

Introduction

• Systems that are performant and resourceful

• Green, energy-frugal, power-proportional ….

• Energy supply perspective: battery, harvester, power control, regulation

• Energy consumption spectrum: Mwatts (data plants), 100s Watts (many core chips), uWatts (implantable devices)

4

Introduction

Challenge: does energy aspect modulate computations?

• both in every system design case,

• as well as an evolution process (e.g. can we come up with some “Mooreish” law of system performance driven by energy levels?)

Before the mountain assault:

• Measure your terrain – identify important constraints and criteria

• Prepare your gear – work from design examples

5

Introduction

Energy-constrained systems

• Solar energy, e-beam power supply, small batteries, …

Unreliable power supply

• Voltage fluctuations, low battery, …

Hostile environments

• High/low temperatures, noise, …

Energy

QoS Uncertainty

Working conditions:

6

The “Holistic” Project

EPSRC project: Next Generation Energy-Harvesting Electronics: Holistic Approach,

Universities of Southampton, Bristol, Newcastle and Imperial College London

"Energy Harvesting Systems: A Block Diagram (2010, July 16). Holistic Energy Harvesting

[Online]. Available: http://www.holistic.ecs.soton.ac.uk/res/eh-system.php"

http://www.holistic.ecs.soton.ac.uk/res/eh-system.php http://www.holistic.ecs.soton.ac.uk/res/eh-system.php http://www.holistic.ecs.soton.ac.uk/res/eh-system.php

7

Energy-proportional computing

Energy

per

action

Activity level

Ideal

proportionality

Real consumption

Design should push it down!

“Systems tend to be designed and optimized for

peak performance. In reality, most computation

nodes, networks and storage devices typically

operate at a fraction of the maximum load, and

do this with surprisingly low energy efficiency.

If we could design systems that do nothing well

(as phrased by David Culler), major energy

savings would be enabled. Accomplishing

energy-proportional computing requires a full-

fledged top-down and bottom-up approach to

the design of IT systems.” (from Jan Rabaey’s

lecture The Art of Green Design: Doing

Nothing Well – March 2010)

8

Power-proportional vs Power-Efficient

Power level

QoS

Level

Design 1 power

proportional and

efficient for low

power

Design 2 less

power proportional

and more efficient

for high power

9

Power-proportional vs Power-Efficient

Power level

QoS

Level

Design 1 power

proportional and

efficient for low

power

Design 2 less

power proportional

and more efficient

for high power

Ideally, we need

a hybrid design

that can adapt to

power levels!

10

Relationship with timing variability

Timing robustness

Source of variability

analysis:

Yu Cao, Clark, L.T.,

2007

Technology node:

90nm

11

Power proportionality experiments

• Experiment setup

• Synthesised for Faraday library

• Based on UMC 90nm technology process

• 8-bit Booth’s multiplier – speed as the QoS

• SPECTRE analogue simulation

• Runs at 1V, 0.9V,…, 0.1V source voltage

• Iterative FFT – precision as the QoS

• VCS and PrimeTime-PX digital simulation

• Runs as nominal 1V source voltage

12

Benchmarks: 8-bit Booth’s Multiplier

• Synchronous

• Rigid 1GHz clock

• Frequency scaling

• Tuned for 1GHz, 500MHz and 250MHz

• Asynchronous, bundled data

• Extra control logic and delay lines

• Asynchronous, dual-rail

• Double comb. logic and FF size (more leakage)

• Extra completion detection and single-rail to dual-rail converters

• Double switching activity (spacer/code-word)

13

Benchmark Architectures

Adaptive frequency scaling Bundled data

Dual-rail

14

Multiplier: Simulation Results

Area (cells)

Energy consumption (pJ) @ Computation time (ns)

1.0V 0.9V 0.8V 0.7V 0.6V 0.5V 0.4V 0.3V 0.2V

A 2495 22.5

@18.0 17.9

@18.0 13.8

@18.0 - - - - - -

B 2495 22.5

@18.0 17.9

@18.0 13.8

@18.0 10.4

@36.0 7.5

@36.0 5.2

@72.0 - - -

C 2931 39.6

@17.2 30.8

@19.2 23.3

@22.2 17.2

@26.9 12.3

@35.2 8.4

@52.1 5.4

@96.1 3.3

@263.9 -

D 7683 279.0

@38.1 217.3

@41.6 166.1

@49.5 123.8

@61.5 87.1

@106.1 60.8

@142.6 39.1

@283.6 24.3

@831.4 19.1

@4140

A - synchronous, fixed frequency @1GHz B - synchronous, frequency scaling @1GHz, 500MHz, 250MHz C - asynchronous, bundled data D - asynchronous, dual-rail

15

Multiplier: Quality of Service

16

Benchmarks: Iterative FFT • Non-reconfigurable FFT

• Fixed transform size (1024 points)

• Reconfigurable sample size

• Changeable transform size (1024/512/256 points)

• Reconfigurable data precision

• Variable data representation (16bit / 12 bit / 8 bit)

• Reconfigurable sample size and data precision

• Asynchronous reconfigurable FFT implementation

• Bundled data with adjustable delay

– load / calculate / unload modes

–16 bit / 12 bit / 8 bit data precision

17

Iterative FFT: Non-reconfigurable

18

Iterative FFT: Reconfigurable

19

Iterative FFT: Asynchronous

20

Iterative FFT: Clock Generator

21

Iterative FFT: Simulation Results

Area (cells)

Energy consumption (µJ)

1024 points 512 points 256 points

16bit 12bit 8bit 16bit 12bit 8bit 16bit 12bit 8bit

A 914K 5.51 - - - - - - - -

B 920K 6.59 - - 3.01 - - 1.36 - -

C 925K 5.83 4.71 3.40 - - - - - -

D 936K 6.66 5.39 3.90 2.98 2.41 1.76 1.32 1.08 0.79

E 942K 6.09 4.92 3.55 2.73 2.20 1.60 1.21 0.98 0.72

A - non-reconfigurable B - reconfigurable transform size C - reconfigurable data precision D - reconfigurable size and precision E - asynchronous reconfigurable

22

Iterative FFT: Quality of Service

23

Towards Power-Adaptive Systems

• Truly energy-modulated design must be power- adaptive

• Systems that are power adaptive are more resourceful and more resilient

• Power-adaptive systems can work in a broad range of power levels

• How to design such systems?

• Let’s first consider the difference between battery- powered and harvester-powered system designs …

24

Portable Power Supplies

For mobile computing applications the choices of power supply are

either batteries or emerging energy-harvester supplies.

Battery:

• Can supply finite energy (E)

– depends on the battery

capacity.

• The available power (P) can

be very large.

Energy-Harvester:

• Can supply infinite energy

(E).

• The rate of energy

production (dE/dt = P) is

variable and can be small.

S P Beeby et al., 2007, “A micro electromagnetic generator for vibration

energy harvesting”, J. Micromech. Microeng. 17 (2007) 1257–1265.

Typical Battery Discharge Curve

A Micro-Electromagnetic

vibration harvester output voltage

25

Circuit Designer Choices (1)

• Determine from T0 the required power

consumption P0.

• Design the circuit for constant P0 consumption

→ constant V0 supply → constant f0 performa