GPU benchmarks

[stilgarna]

Some Test

Hello here are some results at work this afternoon :

--------------------------------GPU :
Vray 2.40.3
Drivers 314.07 (only Nvidia)

2x GeForce 580GTX 3Gb : 1min28s
1x GeForce 580GTX 3Gb : 2min44s
1x GeForce 680GTX 2Gb : 3min03s
1x GeForce 570GTX 1.26Gb : 3min12s
1x GeForce 470GTX 1.26Gb : 3min57s
1x GeForce 560Ti 1Gb : 4min50s
1x Geforce 460GTX 1Gb : 6min40S

--------------------------------CPU : (just for fun )
2x Xeon E2650 : 16Min
1x I7 980 : 32min20s
1x I7 920 : 60min40s



If someone can test one 670GTX, it could be nice, thank you !
I want to know if it's better to get a 690GTX or 2x 670GTX ??

Best regards...

[Alan Iglesias]

Thanks for the info. If GPU rendering was on par with regular rendering such thing may be worth investigating/investing in. But for me personally I won't bother getting gpu config until vray rt gpu is ready to handle all the aspects of our production and maybe, when it is, there are really good/reasonable cards available at that time.

Makes perfect sense...and I wouldn't be surprised if you saw your requirements start to crystalize in the next 12-48 months...

Best,

-Alan

[Morbid Angel]

Thanks for the info. If GPU rendering was on par with regular rendering such thing may be worth investigating/investing in. But for me personally I won't bother getting gpu config until vray rt gpu is ready to handle all the aspects of our production and maybe, when it is, there are really good/reasonable cards available at that time.

[dtolios]

Vlado is it safe to assume that they will come up with similar config like gtx 690 for titan?

I would say no, because GPU manufacturers are trying to stay within reasonable TDP envelopes with their cards.
590 was more than 350W, 690 is around 300W and Titan is around 250W. A dual Titan/K20 based card would be flirting with 450-500W TDP (usually dual-GPU cards are clocked lower, so consume a bit less - that's why 2x 680s > 1x 690).

450-500W would be at least 3x 8pin PCIe aux connectors, and nearly impossible to cool down with a 2 slot air-cooler. We had slower and much lower TDP fast Fermi GPU cards pushing high 90s or more under load, when even the extreme O/C CPU ppl go crazy with CPU temps above 85oC thinking those are too high.

Getting a 3x slot cooler design would be problematic in case fitting, and in the long run you would have to go to great lengths to cram more Titans in than you would with existing 2-slot designs. If you have this kind of patience and money, you can engineer a custom or modify a case and use flexible pcie risers/extenders to fit more than 4x titans that current motherboards can accommodate in an affordable 8-9 expansion slot mid or full tower. In theory you can get 6-7 of them on at least a couple X79 boards out there.

Some kinda dated 2P Xeon barebones from Tyan can do 8 dual slot cards for GPU computing. I believe the 2400W PSUs in those can support 8x titans (bit tight), as long as you don't do 100% CPU + 100% GPU utilization for prolonged time.

[alexyork]

Vlado is it safe to assume that they will come up with similar config like gtx 690 for titan?

I've heard this is highly unlikely. More likely is a 7xx series, with dual GPU config. But those might fall back to 3GB/4GB RAM.

I'd say 2 x Titans is the way to go, if you can afford it!

[vlado]

Vlado is it safe to assume that they will come up with similar config like gtx 690 for titan?

Heh how would I know

Best regards,
Vlado

[Morbid Angel]

Vlado is it safe to assume that they will come up with similar config like gtx 690 for titan?

[vlado]

So we got a Titan card to test; as has been posted somewhere else, it renders the benchmark scene in 1m 46s. Here are some other numbers:

GTX Titan: 1m 46s
GTX 580: 2m 39s
GTX 680: 3m 2s
GTX 690: 1m 31s

Best regards,
Vlado

[alexyork]

No worries I guess we would need two scenes now, one to test memory and one to test speed.

I think it's necessary to have multiple scenes that will work for multiple GPU types based on RAM. I would like to see different scenes for 1GB, 2GB, 3GB and 5GB. This should cover all the major cards I think. If we only have scenes that required 1GB then we're never going to be pushing Titans etc., similarly if we only have 1 or 2 scenes that are 3GB or so then older cards with small VRAM won't run the scenes at all.

No worries I guess we would need two scenes now, one to test memory and one to test speed.

[vlado]

I wouldn't mind hosting a benchmark chart like this on MintViz Workshop, only if Chaos Group were to not do this of course.

We probably won't get to it anytime soon, so feel free

Best regards,
Vlado

I wouldn't mind hosting a benchmark chart like this on MintViz Workshop, only if Chaos Group were to not do this of course.

[3LP]

Hey,

I know it's un-loyal competition but here is a nice bench chart

http://www.randomcontrol.com/arionbench

Do you think you could set something up like this for VrayRT, Vlado?

Would make it way easier to track.

Thanks

Stan

[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 !!!

[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