1

What are the best "performance per watt" (measured in MFLOPS/W) for current CPUs and GPU's?

I hear it is near 500 MFLOPS/W for entire computer ( www.green500.org ), but What is the current record for bare CPU or GPU chip?

Improve this question
2
  • 2
    Conveniently, www.green500.org is announcing their new list for the most energy-efficient HPC systems (MFLOPS/W) tomorrow at the SC10 (Super Computing 2010) conference. Some of those CPUs probably offer the best performance per watt.
    – Stefan Lasiewski
    Nov 18, 2010 at 1:21
  • Green 500 count entire Big Iron systems. What about usual GPU card which I can buy on ebay?
    – osgx
    Nov 18, 2010 at 1:28

1 Answer 1

Reset to default
2

The November 2011 results from green500.org lists top 1 as 2026.48 MFLOPS/W. Considering that the top-of-the-line ATI Radeon (Radeon HD 7970) does 16473 MFLOPS/W according to http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units I'd say that any HPC system that does not use high end GPUs is wasting a lot of electricity. I'd guess that the best method would be to use the slowest CPU possible to still feed the GPU (that is, try to do no work on the CPU, only use it as a fancy IO-processor for the GPU).

Of course, this is the maximum computing power you can get per watt. For a generic HPC system running a program not specifically made for the ATI Radeon GPU, the results will be much worse.

Improve this answer
3
  • "16473 MFLOPS" are this flops for 32 or for 64 bit? As I know, green500 is for 64-bit flops.
    – osgx
    Jun 25, 2012 at 10:11
  • @osgx: it would be interesting to know. In worst case, it's 32 bit and the performance ratio of HD 7970 for single/double precision is 947.2/3788.8 giving 4118 MFLOPS/W for double precision. Still quite bit better than top 1 as of November 2011...
    – Mikko Rantalainen
    Jun 25, 2012 at 11:42
  • 1
    Interesting numbers from GPUS AND THE FUTURE OF PARALLEL COMPUTING, Keckler 2011 (www.davidglasco.com/Papers/ieee-micro.pdf): there are 50 picoJoules (pJ) needed for one double-precision fused multiply-add (DFMA) in 2010 with 40 nm node, and the number will be 6-9 pJ in 2017 on 10 nm node. There is also some energy to store/read data in SRAM: 14 pJ in 2010 and 2 pJ in 2017.
    – osgx
    Feb 16, 2014 at 21:57

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .