Michael S (already5chosen@yahoo.com) on 7/10/10 wrote:
---------------------------
>Ian Ameline (ian.ameline@gmail.com) on 7/9/10 wrote:
>---------------------------
>>David Kanter (dkanter@realworldtech.com) on 7/8/10 wrote:
>>
>>>That's simply not true. I've already made it clear that if we compare a high-end
>>>GPU to a high-end CPU, the GPU has at best a 8X advantage in FLOP/s and 5.5X in memory bandwidth. 5.5 < 8 < 10.
>>>
>>
>>There are some other things the GPU can do better than the CPU. If you need fast
>>but not perfectly accurate sqrt or trig fcns, the GPU apparently can do those much
>>faster than the CPU. Similarly, the texture samplers can be very helpful -- if you
>>need to do 1d, 2d, or 3d table lookups with interpolation, the texture samplers
>>with their caches can be very fast. (Just remember that those caches are not coherent
>>with any other part of the caches on the GPU.)
>>
>>So it is possible (but not easy) to get greater speedup on the GPU than the 5.5x
>>bandwidth or 8x flops would lead you to expect.
>>
>>-- Ian.
>>
>
>Very good point, IMHO, esp about utilizing texture units.
>However I don't quite understand what you are saying about approximation of sqrt.
>On Intel CPUs RSQRTPS has half the throughput of ADDPS and the same latency. On
>the new AMD processors it is faster yet - same throughput as ADDPS.
>Do GPUs have even better ratio between calculation of approximate sqrt and regular arithmetics?

No similar throughput and latency on GPUs, but rsqrtps produces 12 good bits of precision in the fraction. On GPUs, you can get much more (being deliberately vague, as I recall being told the exact precision by one GPU vendor a number of years ago, but I'm reasonably sure that my recollection also includes the fact that the whole meeting/conversation was covered by an NDA.)

--Ian