4 GPU Box Render Settings

[bendreality]

Hey guys,

we are testing a 4 GPU Extention Box (Cubix Xpander Elite) to render with VRay RT and some questions emerged out of our session. First of all we are not using all the same graphics cards. Our atual setup consists of 2 Titan Black and 2 1080 cards. Yea, I know not the best setup to mix up G-cards, but it's all we have available.

so Question one:
The Log says
[2017/Mar/10|17:09:23] V-Ray: Querying for CUDA devices...
[2017/Mar/10|17:09:23] V-Ray: Device GeForce GTX 1080 has compute capability 6.1
[2017/Mar/10|17:09:24] V-Ray: Setting CUDA stack size to 14000
[2017/Mar/10|17:09:24] V-Ray: Device GeForce GTX 1080 has compute capability 6.1
[2017/Mar/10|17:09:24] V-Ray: Setting CUDA stack size to 14000
[2017/Mar/10|17:09:24] V-Ray: Device GeForce GTX TITAN Black has compute capability 3.5
[2017/Mar/10|17:09:25] V-Ray: Setting CUDA stack size to 14000
[2017/Mar/10|17:09:25] V-Ray: Device GeForce GTX TITAN Black has compute capability 3.5
[2017/Mar/10|17:09:25] V-Ray: Setting CUDA stack size to 14000

what does the term "compute capability" mean and what would be the consequences of two different values for the rendering?

The Second Question is regarding the rendersettings.


1. Should the Ray Bundle size be mutch higher when we use 4 G-cards? What would be a "good" or better recommended setting here?
2. How are the noise threshold and the Max Subdivs are related?
The assumption on our side was that when the sample threshold is reached the render stops regardless how many Max subdivs are set. But even when the threshold was reached the rendering went on. So, now we are not sure anymore.^^

3rd question:
We were wondering how you set up the render settings to leverage the full power of the 4 cards. Are there any general rules that can be applied?

All in all it's definetly a lot of fun to get feedback on shading and light so fast. Great work guys.

greetz
Ben

[savage309]

On the questions.
1. Compute capability number is just a number associated with NVIDIA architecture. 2.X is Fermi, 3.X is Kepler, 5.X is Maxwell, 6.X is Pascal. You can mix those without any worries.
2.1 It is best to try for yourself. Values between 16 and 64 for rays per pixel could work best.
2.2 Whatever is reached first - max samples per pixel (subdivs) or noise threshold will make the render not sampling this pixel anymore (it is per pixel). It is not a good idea to leave "Max subdivs" to 0, it will make some scenes much slower.
3. It should work quite straight forward without need of special care.

Btw, V-Ray GPU 3.5 for Maya (and current nightlies) are quite faster compared to 3.4 - there was a specific thing about the Maya materials which was improved recently.

Best,
Blago.

p.s. make sure to check these as well.

[bendreality]

Hey Blago,

sorry for the late answer and thank you very much for clearing the questionmarks. I'll have a close look at the thread you linked.

Cheers
Ben