GPU benchmarks

[Alan Iglesias]

Ah, thanks for the info, Vlado. Let's hope they get that bug fixed soon and you can optimize the code.

Yes, it sure seemed as though the slowdown was a function of the different driver version code (read: nVidia), so let's hope they can straighten that out. I know that as non-gamers, we are not their primary target market, but I'm hoping that eventually that with the advent of GPU rendering and its become more prolific, they will see how important of a market we will be and act accordingly.

I'm going to do a little research and see if I can't get the ear of some of their marketing and R&D people. Longshot, I know, but it couldn't hurt.

Again, thanks for the great 2.2 update and all the hard work the Chaos Group is doing.

-Alan

[vlado]

I'll be uploading a modified installation tomorrow where I added back the optimizations I mentioned - looks like the 290.xx driver works a better with them and the speed is very similar to the older drivers.

Best regards,
Vlado

[vlado]

So the modified installer is now on the web site; the V-Ray RT version is 2.20.02. With the 290.36 driver, it renders the test scene on my GeForce 580 GTX for 3m 5s, which is very similar to what I got before with V-Ray RT 2.10 and the 263.x drivers. Apologies for the inconvenience.

Note: the nVidia OpenCL compiler from the 290.36 driver package will take about 4.5 GB of RAM to compile the code. Make sure you have it

Best regards,
Vlado

[Alan Iglesias]

So I've installed the 290.36 driver and installed the new RT .2 code and yes, there is a definite difference! I am now rendering the benchmark scene at 2:04 which is a 17% improvement! I also know that my time is in the ballpark as my GTX 580 by itself renders the benchmark file in the same time as Vlado's did.

I would only caution folks out there that the initial OCL compile takes a considerable longer time (at least it did here), so please be patient. And as Vlado mentioned, it takes quite a bit of RAM to compile, so using a non-complex file for the initial compile is probably a good thing. My somewhat RAM-challenged test machine (6GB) pretty much ran out of RAM during the compile, but still eventually finished it.

All in all, an excellent update and well worth the effort. Thanks Vlado!

-Alan

[fraggle]

Here is a chart of the results so far:



I will be updating this as more results become available. An interesting fact is that the 580 card with the newest drivers is exactly as fast as the 480 with the older drivers...

Best regards,
Vlado

this is quite confusing. Number of Cuda core rules the game normally. So why a card with more cuda core (512 vs 480 core) did not perform better than the older one. Can the bad new driver eat all the benefit of more core?

[vlado]

this is quite confusing. Number of Cuda core rules the game normally. So why a card with more cuda core (512 vs 480 core) did not perform better than the older one. Can the bad new driver eat all the benefit of more core?

Yes. The optimizations done by the compiler are very important. However the drivers mentioned in the table are quite old now; it would be interesting to do the test with newer drivers (using the same GPU code though).

Best regards,
Vlado

[3LP]

Yeah a new list with the latest drivers and cards would be great, any news about the new ati ?

I was also thinking about having a similar chart for the cpu benchmark.
The thread is somehow a little dead sometimes and there are still every day some new cpu coming out.
Would be really handy for ppl searching for creating some new nodes or workstations.

[lukx]

how is it possible that 560 is slower than 480 ?!

[Paul Oblomov]

Any news on latest radeon cards ?

[vlado]

It's sitting on my desk Will try to get you a result a bit later today.

Best regards,
Vlado

[Paul Oblomov]

Great . . . .

[vlado]

Great . . . .

Well, on a first test not so great - the 12.2 video driver just crashes...

Best regards,
Vlado

[Dariusz Makowski (Dadal)]

Heya

Nice test! I just thought it drop in that question here since is kinda related its been bugging me for a while now...

Anyway Vlado do u think that Chaosgroup did a bit of a mistake by investing in to OpenCL mode and not the CUDA giving the fact that ATI Radeons are a failure for GPU processing? I guess openCL is also good because then we can use custom shaders from openGLSL instead of CUDA?

Thanks, bye.

[vlado]

Anyway Vlado do u think that Chaosgroup did a bit of a mistake by investing in to OpenCL mode and not the CUDA giving the fact that ATI Radeons are a failure for GPU processing?

Well, performance-wise there isn't much of a difference - in fact, for us OpenCL tends to be a little faster than CUDA on the same hardware for V-Ray RT GPU. So from that point of view, it wasn't a mistake. On the other hand, it certainly was much more difficult to develop for OpenCL than it would have been for CUDA and we could have implemented the same features much faster. In any, case, see the snapshot below

Best regards,
Vlado

[Delineate]

Does Graphic card memory make a difference? 1Gb vs 1.5GB vs 3GB?