GPU benchmarks

[Morbid Angel]

This sounds a bit too slow no? I mean JamesCutler got 1m39 on one GTX 690...?

[vlado]

This sounds a bit too slow no? I mean JamesCutler got 1m39 on one GTX 690...?

It's exactly right; a 690 is a bit less than twice as fast as a 680; 4x 680 would be a bit more than twice as fast as a single 690; so 1m39s=99s which divided by two is about 50s, and 3LP's time is 45s - so all is good.

Best regards,
Vlado

[Morbid Angel]

Ah so 680 is a single gpu and 690 is a double gpu..?

[raymarcher]

Ah so 680 is a single gpu and 690 is a double gpu..?

Indeed, that is why it would be cool to have a 790 if ever released , would be equivalent of 2 780's.

[Jiri.Matys]

I do benchmark with GTX 580 again and here is results:

V-Ray 2.40 - OpenCL
GeForce GTX580 (basic without overclock)
Driver: 314.07
TIME: 1m 47s
------
V-Ray 2.40 - CUDA
TIME: 1m 20s

great!!

[Guest]

1m 7s on a GTX 580 and 1m 39s on a GTX 690? Seems off

[vlado]

Yes, it does seem a bit off, unless there was more than one gtx 580.

Best regards,
Vlado

[Jiri.Matys]

Sorry, dont know why I wrote 1m 7s. It was 1m 20s in CUDA. In OpenCL it was 1m 47s.

And I have one GeForce GTX 580 card, but newest drivers and VRay 2.40. I think all the test was done with 2.30.

Dont know if there is another problem, like vray 2.40 doesnt reduce noise the same way like 2.30 and this is why its so fast or something like that. But I done it a few hours ago and I can do it again and post a screen to believe me.

And what I know about GeForce, 6xx series is worse for GPGPU then the 5xx series.

[jpmartin]

Just did a benchmark for the Quadro 5000 since I didn't see it in here. 512 paths per pixel in 5m 33.8s. Not very impressive, considering the price of the card a year ago. Looks like I might need to augment my system!

V-Ray 2.40.03 - CUDA
Quadro 5000
Driver: 305.93
TIME: 5m 33.8s

[Jiri.Matys]

Now I know, why it was so fast. I havent textures

Its ok now, do a test again:

VRay RT 2.40.01 - CUDA
GeForce GTX 580 - driver 314.07
512 paths/pixel reached
TIME: 2m 41.1s

[Guest]

That sounds better

[Jiri.Matys]

That sounds better

Yes But still speed up! I done this test before, you can see my old result on the table: Jiri.Matys: time 3m 8s. Now its 2m 41s. Cool!

[stilgarna]

Hello,

I've done some quick test (CUDA) between GTX580 (3gb) and GTX680 (2gb), and I've got really close render times.

Any idea how to explain that ?
Any chance to see improvement with GTX680 ?? (drivers or vray build)

Best regards...

[Alan Iglesias]

Hello,

I've done some quick test (CUDA) between GTX580 (3gb) and GTX680 (2gb), and I've got really close render times.

Any idea how to explain that ?
Any chance to see improvement with GTX680 ?? (drivers or vray build)

The simple explanation is that although there are many more cores in the 680, they are completely different technology (Fermi vs Kepler) and the 680's Kepler cores for whatever reason raytrace (at least in RT/GPU) at less than a third of the speed than the Fermi cores did. That's why your 1536 (Kepler) core 680 renders with ~5% less speed as the 512 (Fermi) core GTX 580.

I was told by nVidia project manager Greg Estes at Siggraph last year that the fact that the Fermi cores raytraced so quickly was not something they had set out to do in the first place, and that the Kepler technology would be the direction they would be going in.

Whether or not there can be some optimization in RT/GPU to speed up the Kepler technology any more than it already is is question for Vlado of course, but it was great to see the ~14% speed up in RT/GPU using CUDA in the last update.

Hope this makes some sense,

-Alan

[stilgarna]

The simple explanation is that although there are many more cores in the 680, they are completely different technology (Fermi vs Kepler) and the 680's Kepler cores for whatever reason raytrace (at least in RT/GPU) at less than a third of the speed than the Fermi cores did. That's why your 1536 (Kepler) core 680 renders with ~5% less speed as the 512 (Fermi) core GTX 580.

I was told by nVidia project manager Greg Estes at Siggraph last year that the fact that the Fermi cores raytraced so quickly was not something they had set out to do in the first place, and that the Kepler technology would be the direction they would be going in.

Whether or not there can be some optimization in RT/GPU to speed up the Kepler technology any more than it already is is question for Vlado of course, but it was great to see the ~14% speed up in RT/GPU using CUDA in the last update.

Hope this makes some sense,

-Alan

Thanks !!!