Bill Broadley wrote: > Craig Tierney wrote: >> Where did you get the 1/12th number for NVIDIA? For each streaming >> multiprocessor (SM) >> has 1 single precision FPU per thread (8 threads per SM), but only 1 double >> precision FPU >> on the SM. So that ratio would be 1/8. > > I just used the nvidia provided information: > http://www.nvidia.com/object/product_tesla_c1060_us.html > > Click on specifications. 933/78=11.96 > >> I have demonstrated this ratio on a simple >> code that required little to no memory transfers. > > Maybe the ratios are different when the workload isn't optimal for getting the > peak rate. Peak numbers often require very special situations, something like > interleaved adds and multiplies or a fused instruction that does 2 flops. So > maybe for pure multiplications you get 1/8th, but for the perfect workload you > get 1/12th. > > >
Go figure, it does say that. I dug in and the explanation I found is that the single precision number is based on 3 operations per cycle. Two come from the FP unit (MADD operation) and the third comes from the "special function" unit. The special function unit is for transcendental function units, and several other functions. If you are smart (or lucky) you could get the 1/12 ratio. If based on just MADD it is 1/8. http://forums.nvidia.com/lofiversion/index.php?t75452.html http://www.pcper.com/article.php?aid=578 Craig -- Craig Tierney (craig.tier...@noaa.gov) _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
