David Kanter (dkanter@realworldtech.com) on 7/8/10 wrote:
---------------------------
>anon (anon@anon.anon) on 7/8/10 wrote:
>---------------------------
>>David Kanter (dkanter@realworldtech.com) on 7/8/10 wrote:
>>>Um, you realize that Westmere has ~150 GFLOP/s? Almost any workload which can
>>>make good use of a GPU is going to make excellent use of a CPU, including multiple threads and SIMD.
>>
>>Is that theory or measured?
>
>Theory...exactly like the numbers for most GPUs.
>
While it is theory, x86 processors generally delivery in the 80-95% range of peak in a variety of workload. In contrast GPUs and other exotic arrays of FPUs generally deliver much much less of peak. The Prace group out of europe has been doing a lot of evaluations of different architectures and languages over the past 2 years. An example of the result can be found in http://www.prace-project.eu/documents/02_wp8prototypes_hh.pdf .

For something like an MxM kernal, cpus get ~94% efficiency where as GPUs top out at ~81% efficiency. Once you go to things that aren't as simple as MxM the differential in efficiency goes up by a lot. It effectively breaks down to unless you are pretty much doing MxM and only MxM, the best option is x86.