NVLink Question

[Bighouse]

I don't seem to have the ability to post to the GPU forums, so I'm posting this question here. Chaosgroup seems to do some pretty amazing work for making rendering possible within programs, taking advantage of GPUs to do so. Why can't identical cards that use NVLink, be used on the same system to take full advantage of ALL the combined GPUs on each cards? I understand, from what I read, that it will ONLY really make use of the added/combined memory on the cards, but not the GPUs.
Is this a limitation that imposed on the rendering engine by NVidia, a technical impossibility, or is it a business decision on the part of Chaosgroup? It seems other developers can access all GPUs at once for other purposes, can't V-Ray do the same?
I've got two RTX2080ti cards that do have NVLink on them and I'd really, really, really like to use both of them to render simultaneously with all GPUs with V-Ray.

[Muhammed_Hamed]

I don't seem to have the ability to post to the GPU forums, so I'm posting this question here

Send an Email to support@chaosgroup.com , requesting access to GPU forums. They will help you

Why can't identical cards that use NVLink, be used on the same system to take full advantage of ALL the combined GPUs on each cards?

This doesn't make sense to me, let me explain
Each Video card has only one GPU(except for rare cases like the Titan Z, GTX 590, and Tesla cards)
So talking about your RTX 2080Ti for example, this card has one GPU.. If you have 2 cards, Vray GPU can make full advantage of both of them in GPU rendering, make sure you select both of them in the render devices menu.
At work I have 7 1080Tis, Vray GPU can take advantage of all of them and I get 7 times faster render times compared to using one card. And btw, you don't need to use identical cards, so If you have 1080Ti and 2080Ti in one system, Vray GPU can still use both in rendering without issues
For NVlink though, you will need identical cards that support Nvlink like RTX 2080, RTX 2080Ti, GP100 ,GV100 and others

I understand, from what I read, that it will ONLY really make use of the added/combined memory on the cards, but not the GPUs

This not the case, Vray GPU can make use of multiple cards(GPUs) since long time
NVlink bridge can combine the VRAM, but this doesn't affect how Vray GPU uses multiple cards

Is this a limitation that imposed on the rendering engine by NVidia

All GPU renderers outthere can use all GPUs installed in your system, This is been the case with GPU rendering for a decade
But Nvlink only works in Vray so far, which is nice

It seems other developers can access all GPUs at once for other purposes, can't V-Ray do the same?

like other renderers Vray can use/access all GPUs, Many many people have been using Vray GPU in production for years, read above

[Bighouse]

Muhammed,

Thank you for your reply. I'm not an IT person, so excuse my errors. But, let me ask you a couple of questions to confirm what you're saying and then I can clear up my own confusion- which exists because I've either misread or misinterpreted the information I've read in other discussion groups or websites.

WIth NVLink, V-Ray can combine both cards' VRAM to allow for larger rendering requirements. Is this correct?

With NVLink, V-Ray can use both cards' CUDA cores combined to perform rendering calculations. Is this correct? (This is the one that I'm not certain about as I've read the only real advantage is that VRAM can be combined.)

As I understand it, there is a bit of a performance hit that is taken when NVLink has to do its work- so it's not a perfect mathematical combination of VRAM or cuda numbers. Is this also correct?

I'm mostly concerned about finding a rendering hardware solution that will minimize my rendering times with V-Ray with a budget of around $2500 (USD) for that rendering card(s). I've purchased two RTX2080ti cards and I'm about to water-cool them (voiding my warranty?). Before I do, I want to make certain I wouldn't have received better rendering times had I gone with a different option.

Thank you.

[savage309]

Yes, this is correct.
Keen in mind that to get nvlink working you need to have SLI setup for GeForce cards, if you are on windows.
nvlink with GeForce works only on 2080 and 2080ti. The memory is not exactly doubled, as there is some overhead, meaning 2 x 2080ti will be close to 22GB, but less than 22GB. It is hard to say how much exactly, as it depends. Nvlink requires separate bridge, which Nvidia sells for about 79$.
This might help https://www.chaosgroup.com/blog/what...endering-v-ray

Best,
Blago.

[Bighouse]

Savage309, Thank you for your reply. I do have the NVLink device. I was confused because I saw a lot of mention about doubling the VRAM for large scenes, but didn't see anything much about doubling the CUDA Cores that the renderer can take advantage of. I do see, in the article you pointed me to, that it does say that "....double your speed..." so I assume that means it doubles the amount of CUDA cores the renderer can use.

Thanks! Excuse my ignorance, I'm learning.

[savage309]

People often say double the memory and don’t mention double the cores, as doubling the used cores has always been the case and it is just assumed to just work at this point. As (almost) doubling the memory is what nvlink brings as new, it is what get mentioned.

Best,
Blago.

[Muhammed_Hamed]

Thanks for all info Blago ^^