GPU benchmarks

[vlado]

Use CUDA...

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.

I check again with only the 2x780 OC = 1m17s instead of 61 s.
I didn't change anything on the scene, I just choose Cuda and hit render.

[fraggle]

Version 3.10.03 make it 1m34s, so it take longer on new vray RT version than the older one - vray 3.08!
Is it as expected?

[vlado]

For 3.10, it is expected (because of some new features that we added). The 3.20 service pack is faster though.

Best regards,
Vlado

[savage309]

[ATTACH=CONFIG]24298[/ATTACH]Version 3.10.03 make it 1m34s, so it take longer on new vray RT version than the older one - vray 3.08!
Is it as expected?

I am not sure if the time in the VFB reports only the render time or the total time (scene loading, etc). Will check that tomorrow.
If it is the second you might keep in mind that starting from 3.1 we do compile some stuff on the fly when you hit render. They are cached and reused and if this is the first time you run the test it might be the case that some stuff were compiled and this influenced the total render time (compilation usually takes between 5 and 15 seconds).
These on the fly compilations are working automatically and they are just some dev stuff we do to make vray faster and you should now worry about them at all, but if you do benchmarks it is good to pay attention on that.

[3LP]

Hello!

Hi guys,
Just wanted to update this little thread with the recent release of SP2!
In three works : "It is awesome"

Vlado and Chaosgroup, you guys really did a outstanding job on this one, it's truly amazing!
Beside all the new features you guys added, the speed increase is drastic!

So here we go for the benchmark scene we all know with the new SP2 :
Tri GTX970 : 56.9 sec in stead of my previous 1m 10.5s !
That's 20% speed increase!


From here on, in stead of adding those new benchmark info into the "old" chart and creating confusion, I would suggest now, to jump on to the new chart I added with the new benchmark scene who is designed for more robust configs.

Here is the chart :


The chart page on GDocs is still the same, just scroll down to the Benchmark 2.0.
https://docs.google.com/spreadsheets...A0k/edit#gid=0

Vlado, do you think you could let this one go on your VCA now that SP2 is out, hopefully this new scene will run for about 1min and this will be a good reference point.
Should we make a new topic for this?

Cheers
Stan

[Alan Iglesias]

...So here we go for the benchmark scene we all know with the new SP2 :
Tri GTX970 : 56.9 sec in stead of my previous 1m 10.5s !
That's 20% speed increase!...

Heya Stan, your new benchmark on the chart for 3 x 970 is 5mins+. Are you using a new benchmark file?

Thanks,

-Alan

[3LP]

Yes the one with the link in the Gspreadsheet :
https://dl.dropboxusercontent.com/u/368267/benchmark_scene.zip

It's designed to be running somehow close to 1min on a NVidia VCA (hopefulley) just waiting for the guys at Chaos to let it run on theirs.
That will give us a good base for comparison

Cheers
Stan

[Alan Iglesias]

Got it..thanks Stan...

-Alan

[3LP]

@Vlado : any news if you guys had a chance to run the benchmark on your VCA for about a minute with SP2?

Cheers
Stan

[savage309]

Benchmark #2 updated with Titan X - 683s and VCA results (8 x M6000) - 94s. VCA is almost 8 times faster (I should note that the driver on the VCA is pretty old).

[savage309]

Updated with just one of the VCAs M6000 - 746s. So, with 8 VCAs V-Ray is 7.94 times faster, which is pretty good. And apperantly we have to check for a newer driver, since this one is kind of slow.

[3LP]

Nice!
Thanks for your benchmarks, this will help quite a lot the community.
Would be great if you could try the M6000 with recent drivers if you think this should make a difference.

So at the time being, the VCA seems to be 7.2 faster than a single Titan X.

Can't wait to have our new GPU node build (7 Titan X) to see how it performs! Pretty sure the performance will be really close to the VCA.
Will keep you posted for sure.

Cheers
Stan

[savage309]

M6000, 353.06, Windows, V-Ray 3.20.02 - 674s.
Btw, we changed since 3.1 the way we ship V-Ray RT GPU so we could benefit from the speed ups that the new drivers can give (before you had to update driver *and* V-Ray), good to see that is working too.

[3LP]

Nice, you won about 10% there!
So the M6000 and the Titan are really close to each other now, the difference is barely perceptible, 10 secs on a 11min render...

Do you mind updating those infos on the gdoc spreadsheet? I guess the full 8 gpus are closer to the symbolic 1min render now.

Well done for the additional column with the "times slower" ratio, didn't thought about that one but really handy

Stan