GPU Render Benchmark Test for GTX470 + GTX680

[prorender]

Dear V-Ray users,

today I tested the new GTX680 with GK104 Kepler GPU and GTX470 with GF100 Fermi GPU.

The renders finished with reaching of 256 paths/pixel.

Here are the results:

GTX680

rendered in 2min32s

GTX470

rendered in 3min28s

Comparison:
GTX680 Performance 137%
GTX470 Performance 100%

Regards
Daniel Krüger

www.prorender.de
renderfarm & render service

[vlado]

Nice; I just got my 680 too - can't wait to test it.

Best regards,
Vlado

[Alan Iglesias]

The 470 has 448 cores next to the 680's 1536. So we're seeing a 37% performance increase with well over 200% more cores.

Well, let's see where we go from here...

-Alan

[vlado]

It has been pointed out that they are not the same cores... The CUDA cores in the 680 cards are much simpler and "dumber" than the ones in the Fermi cards. For the Kepler cards, the OpenCL/CUDA compiler must reorder the GPU code to achieve maximum performance - apparently it doesn't do that optimally just yet. For Fermi cards the reordering is apparently done by the hardware on the card itself. That's my understanding anyways...

Best regards,
Vlado

[ivaylo.ivanov]

The new cores are different from the old ones. They are not directly comparable. Also, there isn't a Kepler based Tesla yet, and probably GK104 was never meant for computing. If you remember, GTX480 was presented along with the Tesla cards. It was a high-end chip. GTX 680 is not a high-end chip. Read more carefully the specifications.

[Alan Iglesias]

It has been pointed out that they are not the same cores... The CUDA cores in the 680 cards are much simpler and "dumber" than the ones in the Fermi cards. For the Kepler cards, the OpenCL/CUDA compiler must reorder the GPU code to achieve maximum performance - apparently it doesn't do that optimally just yet. For Fermi cards the reordering is apparently done by the hardware on the card itself. That's my understanding anyways...

Best regards,
Vlado

Ah, most interesting. Thanks for that bit of insight, Vlado, it is much appreciated. I wasn't expecting the same performance as the fermi. As I said, let's see where we go from here...

-Alan