GPU benchmarks

[3LP]

I was wondering if we shouldn't do a bench that would be stripped of any textures, maps etc to maximize the speed that the scene would be transferred to the GPUs.
At the same time I though that the Greeble plugin would help make a lot of little geometry so that BF is used as much as possible.

Here is a little quick test :


https://dl.dropboxusercontent.com/u/368267/Bench.max

This has been rendered in 5m35 on my 3x970s where I'm doing 1m18 with the "old" Bench, so about 4 times more.
Max PPP is 2048 but we could push the ress and the PPP more if we would like to have a more heavy scene.
We might need a pretty hard one to render as once we will have some heavy setup kick in like quad Titan (X or Z) or even NVidia VCA, that bench will quickly fall down.

Just thinking out loud here,
what do you think?

Stan

[fraggle]

61s
dual 780 GTX (oc from factory asus) + 1 GTX 670 (oc from factory asus).

Thanks sounds strange compare with the brand new card!

[3LP]

What 61 seconds?
Can you check you've got PPP set at 2048?
Sure you should do better than me, but not 5 times more.

How much do you do with the "old" benchmark?

[vlado]

You actually need the textures and so on. The GPU is affected by the complexity of materials; comparing plain grey materials doesn't really tell you anything. That's of course true for any render engine comparisons; it's often strange to see people making conclusions from scenes with one light, 5 polygons and grey materials. For GPUs, different architectures handle code complexity in different ways, so you really need to have that. The benchmark scene has bitmaps, color correction textures, bump maps, different materials (refractive, reflective, diffuse) so it's a good test.

Best regards,
Vlado

[fraggle]

What 61 seconds?
Can you check you've got PPP set at 2048?
Sure you should do better than me, but not 5 times more.

How much do you do with the "old" benchmark?

That's the old one, the one provided by Vlado.

[3LP]

You actually need the textures and so on. The GPU is affected by the complexity of materials; comparing plain grey materials doesn't really tell you anything. That's of course true for any render engine comparisons; it's often strange to see people making conclusions from scenes with one light, 5 polygons and grey materials. For GPUs, different architectures handle code complexity in different ways, so you really need to have that. The benchmark scene has bitmaps, color correction textures, bump maps, different materials (refractive, reflective, diffuse) so it's a good test.

Best regards,
Vlado

Right Vlado,
well then maybe lets keep that scene but let's update it by adding more PPP and a higher ress.
maybe 1280px and 2048 ppp?

With those seetings, I just did 8min 10sec on 3x970s, would that be long enough?
Here is the link :
https://dl.dropboxusercontent.com/u/...mark_scene.zip

I was wondering at the same time, Vlado, do you guys have access to a VCA? Do you think there is any chance you'd run that benchmark on one of them?

Cheers

Stan

[vlado]

I was wondering at the same time, Vlado, do you guys have access to a VCA? Do you think there is any chance you'd run that benchmark on one of them?

We do, actually... updated recently to eight M6000's (about 30-40% faster than the K6000's). It would be interesting to run the benchmark there, yes. I will try to figure out settings that would make it run for about a minute. Perhaps then we can measure the performance of other cards relative to that.

Best regards,
Vlado

[3LP]

Nice!
Yes that a great idea actually!

Stan

[3LP]

Actually the new benchmark should not be far from that.
From my estimations, that bench should be done in 68 sec on you VCA if it's 35% faster than the k6000.

Let's see if Im' right

[wyszolmirski]

I will deviate a bit from the new benchmark talk.

Does anyone tested RT on GTX 960 ? Wonder how they perform, 4GB versions are quite cheap as well.

Best,

[fraggle]

What 61 seconds?
Can you check you've got PPP set at 2048?
Sure you should do better than me, but not 5 times more.

How much do you do with the "old" benchmark?

What means PPP?

[3LP]

@Vlado, did you had a chance to send the benchmark on your new VCA for 1min?
We are in the process of building a GPU node with 7 Titan X and I'd like to be able to see how it performs against it.

Cheers

Stan

[savage309]

@Vlado, did you had a chance to send the benchmark on your new VCA for 1min?
We are in the process of building a GPU node with 7 Titan X and I'd like to be able to see how it performs against it.

Cheers

Stan

We have tested that (with only one Titans vs one Quadro) and they seems to be in pair (nvidia-smi reports that Titan X can be started with TCC driver, however my tests so far are not working. So if the Quadro is started with TCC driver, it is faster.).

[vlado]

Sorry for the delay, I hope to get some time tonight to do the VCA test...

Best regards,
Vlado

[Paul Oblomov]

Finally did myself a present - gtx980 strix. And failed to compile opencl kernels Latest drivers of course