Vray GPU and NVLink question

[Markus]

Hey everyone,

I am about to test out GPU rendering, so far I am using Vray (cpu) and corona. I started with reading a lot about which video cards to choose. One thing I am still not sure is the amount of VRAM I really need!

My usual CPU scenes take around 20-50 GB (interior) and up to 100 GB exterior. How does this translate to GPU rendering? I see a lot of really complex looking rendering from people and I really doubt, that they all have GPUs (NVLinked) with this amount of VRAM.
Is there like a rule of thumb of how this translate to GPU rendering?

Speaking about NVlink brings me to my next question. What I understood, NVLink enables one video card accessing data stored in the memory of the 2nd nvlinked card. It is not automatically doubling the ram, more granting access between each other, which comes with a slight penalty in performance.

What if I have 4x 2080ti cards in one System. Can I have two NVlink sets, 2x 2 Nvlinked Cards or does this only work for one pair in a system and I have like 1x Nvlink pair and 2x single cards, or does such a setup only work with 4x single video cards (no NVlink)?


Thanks a lot for some infos

[matanov]

Hey Markus,

NVLINK works with pairs. In your case 4x2080 Ti will have VRAM limitation of the each pair = 22GB (2x2080Ti). Mind that some data needs to be copied on both cards, so it's not straight double the memory rule.

And for the other question, there is no rule of translating CPU to GPU memory usage. You can try one of your scenes with CUDA CPU and check the memory stats in the log to see how much it would take on GPU.

[Markus]

Thank you for the answer Peter. That is already good news, because someone told me, you cant have two nvlink pairs with geforce cards. this would only work with quadro. In run some test yesterday with my old 980ti and must say I was suprised what i was already able to render in the 6GB vram. So 11GB or even something around 19/20GB with nvlink should be enough for me needs for now.

[Muhammed_Hamed]

I'm able to fit scenes that take like 30 GB of RAM on CPU, in less than 8 GB of VRAM
with 8GB you can fit something like 80 million polies, displacement is one thing that consumes a lot of VRAM though
other than that, you should be fine

[Markus]

Cool, that´s great to hear Muhammed. Sounds like 2080ti is really a good starting point for me and with NVLink I have enough option to increase my vram limit.

[Markus]

Hey Markus,

NVLINK works with pairs. In your case 4x2080 Ti will have VRAM limitation of the each pair = 22GB (2x2080Ti). Mind that some data needs to be copied on both cards, so it's not straight double the memory rule.

And for the other question, there is no rule of translating CPU to GPU memory usage. You can try one of your scenes with CUDA CPU and check the memory stats in the log to see how much it would take on GPU.

hey Peter, I need to go back to this possible 2 pair NVlink setup. I read a couple posts on the Octane group on facebook, that having more than one Nvlink pair doesn't really work with geforce cards. This post was several months old, so I wanted to ask if that has changed and you guys at chaosgroup made different experiences?

thanks for clarification on this topic ☺️

[Muhammed_Hamed]

Multiple pairs of NVlink work nicely, many people are running this currently
Check this out, you will need dummy DP cables for one card of each pair at least.. that is all you need to do
https://www.facebook.com/groups/Octa...5531279957281/

[Markus]

Seems the multipair NVlink with Geforce is over for good:

Here is what we heard back from OTOY's developers:

According to feedback we got from NVIDIA regarding this, more than one NVLINK pair is only supported with the Quadro product line. We are under the impression that probably this was not ever intended to work with GeForce in the fist place and the reason why older drivers allowed it could have just been an oversight.

So, I guess the old statement from october isnt valid anymore

[matanov]

Yes, NVIDIA changed this in the newer drivers.

[squintnic]

my god what a stupid move.
really limits Vray GPU on complex scenes