lifting the energy veil a plea for data and science philip levis computer systems lab stanford...

8

Upload: miles-ross

Post on 02-Jan-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University
Page 2: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University

Lifting the Energy VeilA plea for data and science

Philip LevisComputer Systems Lab

Stanford University

Page 3: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University

Amdahl’s Law Extended

• Costs bound benefits

• Corollary: optimize the big cost

• Death of a thousand cuts?

80%

20%

Page 4: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University

Streetlights

?

Page 5: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University

Well-Lit CPUs

In practice, a 3-5% improvement is achievable.

50%

Page 6: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University

Contrast: Displays

35%

In practice, a 20% improvement is achievable.

Page 7: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University

It’s No Wonder….

The issue is that people don’t realize all of the options in the design space. You can turn sleep current arbitrarily low.

Bill Dally, Chief Scientist, NVIDIA

You could have lower power RAM, but engineering it is expensive and the market would not pay for the investment.

Krisztian Flautner, Director of Advanced Research, ARM

I think you will find that a part designed to run at 50MHz and above with 64K of SRAM will always draw more power in sleep mode.

Paul Kimelman, chief architect of ARM Cortex M3

Page 8: Lifting the Energy Veil A plea for data and science Philip Levis Computer Systems Lab Stanford University

Data is Truth

• Separate artifact from essence– (addressing either is commendable)

• “Interesting” is a bad word

• Open, public data sets