GPU benchmarks

[Guest]

Sure!
here it is :
https://dl.dropboxusercontent.com/u/...scene_2012.rar

Jesus mate, 2012, that seems far away
I'm adding the download link to the benchmark spreadsheet just in case.

Thanks
Stan

Yeah I know, not all workstations are up to date yet. That scene rendered in 5 minutes 24 seconds no over clocking. Both run at about 68% power with GPU Z stating they are restricted by reliability voltage under full load.

[super gnu]

anyone tested the fury x? not even considering it (ive just ordered a titan x and cooling block) but id be interested to know if AMD are moving in the right direction..

[vlado]

We tested it, Blago can provide more details...

Best regards,
Vlado

[savage309]

anyone tested the fury x? not even considering it (ive just ordered a titan x and cooling block) but id be interested to know if AMD are moving in the right direction..

The bad news is that the AMD driver fails to compile the current V-Ray versions (3.2 and 3.1 for sure, I don't remember for 3.0). So you can't test that.

The good news is that since a couple of weeks we (with some help from AMD) finally seem to have found so far stable and reliable solution for that (at least on Windows & Linux). So, the OpenCL kernels in the recent internal V-Ray builds are compiled fast and with low memory usage and most importantly, without any issues. It is even better that so far the changes in the code that we regularly make are not affecting that. We can remove, add and change stuff and the kernels are still working. This is the case for all AMD GPUs for which there is driver 15.7 (meaning almost all AMD GCN GPUs - 1.0, 1.1, 1.2 and 1.3), which is a lot of GPUs and it is a big relief.

I've tested the benchmark 1.0 scene with a custom build and through the standalone (so the times are a bit different).
Titan X using CUDA was rendering the scene in about 73sec.
Titan X using OpenCL was rendering the scene in about ~180sec, and Fury X using OpenCL was with roughly the same speed as Titan X using OpenCL.

My suspicions are that the slower render times with OpenCL has nothing to do with the hardware or the language features, but more with the compiler and the driver. So the problem is software, not hardware. This is good, because it is fixable. The bad is that this software fix does not depend on us. If my suspicions are correct, and if they will be fixed we are yet to see.
And as a disclaimer, I have tested only with the benchmark scene so far.

But this is without a doubt steps in the right direction, even that render time is not at all bad. I just hope it is not too late.

[super gnu]

very interesting post.. thanks Blagovest. i must admit im surprised that cuda is sooo much faster than opencl. i didnt realise they had diverged quite so much. hopefully this gap can be closed at some point (rather than opencl gradually being relegated to "legacy feature"). -a good performing AMD card is a good motivator.

[savage309]

very interesting post.. thanks Blagovest. i must admit im surprised that cuda is sooo much faster than opencl. i didnt realise they had diverged quite so much. hopefully this gap can be closed at some point (rather than opencl gradually being relegated to "legacy feature"). -a good performing AMD card is a good motivator.

The gap is actually closing.
We will hopefully have at least 2 GPU vendors that can run our GPU code in one form or another without much hassle (compared to one before) with the next V-Ray SP.
And the RT OpenCL running on AMD is slower compared to the nVidia CUDA, but the Fury X is also 350$ cheaper (649$ vs 999$). And with OpenCL you can use the CPU to help with the render, too. We will have to test with other AMD GPUs like 390x, since I doubt that the HBM is changing that much at the moment.
And this result is still much faster compared to running V-Ray RT OpenCL on a CPU, for example.

[Alan Iglesias]

All this is great to hear...nVidea needs some competition for Vray GPU rendering out there and the sooner the better...

-Alan

[emartin8907]

Where is the latest Benchmark scene for Maya? Would love to try this Titan X, I think I am installing a second one tomorrow to test as well would love to compare. This benchmark thing confuses me though

[3LP]

I never heard about a scene for maya.
Not sure how I can export and keep all the settings though.
Would fbx do the trick?

Stan

[georgi.zhekov]

You can run the scene through VRay Standalone (it comes with your vray installation) - you just give it a vrscene with -rtEngine=5 and it will use CUDA to render it.

[Booska]

Would love to try this Titan X, I think I am installing a second one tomorrow to test as well would love to compare.

I'm still hearing a lot of good things about the Titan X even though the 980Ti is now available. As far as I know, the two perform similarly except that the Titan X has twice as much onboard RAM. I just bought 2 980Tis for my new build. Please tell me that I did the right thing...

[savage309]

You can run the scene through VRay Standalone (it comes with your vray installation) - you just give it a vrscene with -rtEngine=5 and it will use CUDA to render it.

You can download the scene from here http://ftp.chaosgroup.com/support/vr...h_interior.zip
The command for the standalone should be something like " vray -sceneFile="C:\path\to\vrayrt_gpu_bench_interior.vr scene" -rtEngine=5 "
When the render is finished, you should check "RT Frame Time" in the log (please note that this is different from "Frame Time" which include the time needed to prepare the scene). "RT Frame Time" is the one that is printed in the VFB.

@Booska : You can check benchmarks with V-Ray here https://docs.google.com/spreadsheets...A0k/edit#gid=0

[matanov]

You can run the scene through VRay Standalone (it comes with your vray installation) - you just give it a vrscene with -rtEngine=5 and it will use CUDA to render it.

Here is the scene - http://ftp.chaosgroup.com/support/vr...h_interior.zip

[3LP]

I'm still hearing a lot of good things about the Titan X even though the 980Ti is now available. As far as I know, the two perform similarly except that the Titan X has twice as much onboard RAM. I just bought 2 980Tis for my new build. Please tell me that I did the right thing...

Well, I have here one node with 4 Titan X and have 3 other Titan X to stick somewhere
They are awesome, incredibly fast and stable. I even OC them 20% without any issues.

I guess the 980 Ti are close and will perform well, the only point is the ram.
I found that for interiors, I really never had above 4-6 Gb so you should be right, exteriors though, where you have massive maps and loads of forest packs, I think we do need more than 6Gb.

So really happy we went for Titan X but it all depends on your workflow and projects.

Stan

[emartin8907]

1m37s with single Titan X, driver 353.30

Curious about that x9 580 card setup in the benchmark on original page- how many cards can vray see/ how do you build a machine with 9 video cards? 0.o

EDIT: just ran this test at home on dual 780ti classified setup and got 1m.13s