GTX580 floating point performance

[Lars1]

Hi.

I have just finalized some floating point performance test of the GeForce GTX580. As expected it was the fastest single precision performance I have ever measured. The peak single precision performance was 849 GFlops, and peak double precision performance was 193 GFlops. For comparison the Tesla C2050 delivered 550 GFlops single precision peak performance, and 250 GFlops in double precision performance.

Single precision:

i7x975_GTX580_flopsSGeMMjkt_W7J160.jpg (182.74 KiB) Viewed 2420 times

Double precision:

i7x975_GTX580_flopsDGeMMjkt_W7J160.jpg (180.08 KiB) Viewed 2420 times

I hope the results may be useful. I will soon upload it to Torben's Corner - but it will likely have to wait for my Christmas holidays.

Torben

[gallagher.pryor]

Wow - that's really impressive. Is it then correct to compare these results to the charts @ http://wiki.accelereyes.com/wiki/index.php/MTIMES_Benchmarks ?

Are these still up to date? Here are the charts themselves... (wow the img tag worked! awesome!)

[Lars1]

Yes, I believe you can compare.

Below find we a couple of different GPUs for comparison.

Single precision:

specialPlot_SP.jpg (265.16 KiB) Viewed 2407 times

Double precision:

specialPlot_DP.jpg (262.36 KiB) Viewed 2407 times

Torben

[rollyng]

Hi, the above benchmarks seem verified the claim of deliberate reduction of double precision floating point performance of the GForce series...
http://forums.nvidia.com/index.php?showtopic=164417

Anyway, in terms of Cash/Performance ratio, GTX580 is still better than Tesla series

[malcolm]

Jacket 1.7 is in the works and will be using CUDA 3.2. In our experiments, we found that CUBLAS matrix multiply performance went down for certain sizes. We benchmarked cublasSgemm on a Tesla C1060 using the 260.x driver and just switching the toolkit library between cuda3.1 and cuda3.2. The new version using MAGMA is still a big improvement in consistency! It's very tough to get such stable trendlines (blue).

cublasSgemm comparison in GFLOPs

cublas_comparison.png (17.41 KiB) Viewed 2186 times