What usual review Bench tool is computing the closest/similar way to RT GPU?

[3LP]

Hi guys,

I was wondering how I could compare any benchmark tool that are used to try out every hardware that come out.
As a example, Anandtech has done a review of the GTX 980 here : http://www.anandtech.com/show/8526/n...-980-review/20
Loads of bench-tool : LuxMark, CompuBench, Folding @ Home, SystemCompute, etc

They all have different results with CG and they are not consistent.
If I would like to use any of those review/tool to asses what CG I should use for rendering with RT GPU, what should that be?

Thanks

Stan

[3LP]

At the same time, I'm looking into what would be the best bang for bucks if I want to buy few cards. If I'm looking into Luxmark benchmarks (Stil waiting to see if this is somehow the closest benchmark related to RT GPU) the GTX 980 and the GTX 970 are not that far for each other.
CF : http://www.anandtech.com/show/8568/t...w-feat-evga/14
Having both 4Gb of ram, it seems to be more worth wile buying more GTX 970 than less GTX 980 for the same price and have more horse power, what do you guys think? Stan

[3LP]

Hey guys,
No one know what benchmark tool is similar to RT in terms of performance charts?
I just want to know how I can compare CG cards.

In doubt, I'll consider Luxmark as a good place to start...

Stan

[savage309]

I can't tell you what benchmark is closest to V-Ray RT, but we will test a bunch of GPUs here with the latest V-Ray version and we will share the results (having these numbers will tell, in fact, which benchmark is closest to the RT).

[matanov]

Scene used - http://ftp.chaosgroup.com/support/benchmark_scene.zip with max paths per pixel = 512, so we are looking at the time of completion .
V-Ray version official 3.10.02 for 3ds Max
Note the different times with different driver versions.

Soooo here are the results:
Quadro K2200 / 341.05 - 9m 43,7s
Quadro K6000 / 341.12 - 2m 13s
Quadro K5200 / 341.05 - 3m 9,7s
Quadro K4200 / 341.21 - 4m 54,7s
GTX 550Ti / 340.52 - 11m 25s
GTX 560 / 340.52 - 6m 49,7s
GTX 560 / 335.23 - 6m 48s
GTX 770 / 344.75 - 4m 22,6s
GTX 770 / 340.52 - 5m 17,5s
GTX 770 / 344.60 - 4m
GTX 750 / 340.52 - 8m 55s
GTX 570 / 334.89 - 4m 22s
GTX 980 / 344.16 - 2m 57,7s
GTX 980 / 347.09 - 2m 30,9s
GTX 980M / 344.45 - 4m 32s
GTX 480 / 340.52 - 4m 9.5s
GTX 660Ti / 331.82 - 6m
GTX 660Ti / 344.48 - 5m 45,7s
GTX 690 / 340.52 - 2m 11s
GTX 590 / 334.67 - 5m 12s
Tesla K40 / 332.76 - 3m 10,6s

[3LP]

Great info, thanks a lot.

Would you mind doing a few bench on few cards with Luxmark 2.0? Mainly the newer cards would be great : K40/gtx980/k6000
The Room scene would be great.
http://www.luxrender.net/wiki/LuxMark#Download

The purpose would be to see how they scale compared to Vray RT.
That way, in the future, we would have a performance index that would let us asses how they would perform in RT based on those results we can see all over the net.

ie : My GTX 780 Ti does 1222 points

Thanks a lot

Stan

[matanov]

Luxmark's Room - OpenCL GPU only

Quadro K5200 / 341.05 = 795
GTX 980 / 347.09 = 1938

Knowing your results and mine do the math ... you can't rely on this benchmark software to compare and scale GPUs for V-Ray RT CUDA rendering.

[vlado]

If Luxmark is based on OpenCL, then it's best to ignore it. nVidia has not paid attention to their OpenCL performance and it is not an indication of the true capabilities of the cards.

Best regards,
Vlado

[3LP]

Thanks Slizer for that.

Vlado, you're right, (once again ) this opencl benchtool is not usable as I have this ratio :

GTX 780 TI : Vray : 134 / Luxmark : 1220 / Ratio : 9.10
GTX 980 : Vray : 150 / Luxmark : 1938 / Ratio : 12.92
K5200 : Vray : 190 / Luxmark : 795 / Ratio : 4.18

It's weird because you could expect the 980 to be better at cuda than the 780 Ti but it not, although it's better in opencl...

Vlado, what would you think of creating a little Vray benchtool based on cuda?
That way, features sites that trys out hardware could potentially integrate that in their benchtools.
they all seems to be based on opencl for now, so useless for us...
The bench scene seems to be the same for 4 years now and although Vray does evolve, we just need to asses the cards themselves.
Having a consistent benchtool will make it easier to see the evolution of cards over time as well. Currently the benchmarks page has results from different Vray versions so it's hard to ratio them out.

Stan

[ivaylo.ivanov]

Another issue with Luxmark is that it's too simplistic. Newer generation cards like 980 may have some specific optimizations, but when you get to a complex code, the overall performance is what matters. And in that respect 980 is just slightly faster then 780. The only true benchmark for anyting is to run the actual software, there is no other reliable way...

[3LP]

Yes, I get that now.
This is exactly why I was asking if it was possible to create a bench tool using Vray. In the same idea as arionbench.

[3LP]

Hey guys,

Any idea if a easy look to benchmark GPUs with Vray would be possible and come out someday?

I was hoping to use Arionbench in the mean time but it doesn't seems to act as Vray so I'm not sure if any of this data is ok to use.
I used the GPU benchmark as usual and Arion to compare and here as some examples :

GTX 970 :
Vray : 198 sec
Arion : 222 sec


GTX 780 Ti :
Vray : 124sec
Arion : 119sec

Dual GTX 680 :
Vray : 129 sec
Arion : 128 sec

As you can see, some cards are faster with Arion and some with Vray, this seems to be like OpenCL and the previous debated Luxmark tool, it can't be used to asses cards with Vray, that's why a tool proper to Vray would be great. Any info on this?

Maybe it's just a mistake with the 970.
If someone could run few tests with Arion so we could have a better idea if it's usable or not compared to Vray, it would be great.
Here is the link to try it out :
https://dl.dropboxusercontent.com/u/...-%20v2.5.0.zip

Thanks

Stan

[vlado]

I thought Arion uses the CPU to help the calculations, could that mess up the benchmarks?

Best regards,
Vlado

[3LP]

I guess it could but it doesn't use excessively the cpu, it seems to be used only to handle the gpu, or they would have coded it really badly to be working single threaded for the cpu to help rendering :



What I found is that my 780ti and my dual 680 are really close rendering with Vray and arion, but the cpu are completely different.
I would bet more on something to do with the 970 that waks out (either with arion or with vray)

Would be good to have some other data from arion to compare with Vray so we can see if it's only on few specific cards (maxwell architecture?) or if it's completely not relevant.

Or having a tool done by vray as simple to use as arion and then we can throw that arion out of the window