Latest Nividia RTX 2080 ti concerns

[chuwaiting]

Hi there, I am planning to upgrade my workstation to the latest Nvidia RTX 2080 ti but I still have few concerns before I invest this graphic card.

1. Does the vray rt 3.60.04 with 3ds max 2016 support the RTX's real-time ray tracing functions?
2. Does the amount of cuda cores of the graphic cards affect the rendering time? or which spec should i look into for a faster rendering time?
3. I am using a GTX 1080 with 2560 cuda cores right now. Does it run 2x time faster if I use RTX 2080ti with 4352 cuda cores?

Thank you!

[tashko.zashev]

Hi,

1. Does the vray rt 3.60.04 support the RTX's real-time ray tracing functions?

A. No, this is a new architecture and we will support the specially designed RT cores only in V-Ray 4.x versions.

2. Does the amount of cuda cores of the graphic cards affect the rendering time? or which spec should i look into for a faster rendering time?

A. Yes, generally this is the parameter which has influence over the render speed the most.

3. I am using a GTX 1080 with 2560 cuda cores right now. Does it run 2x time faster if I use RTX 2080ti with 4352 cuda cores?

A. We haven't had a chance to test this card yet and make render time comparisons, but we expect to be faster.

[savage309]

Tashko is right, just few notes:
1. V-Ray 3.4 and newer support Turing GPU as of today, using only the CUDA cores.
Nvlink requires V-Ray 3.6 or newer.
TensorCores are used in NViDIA AI denoiser in V-Ray Next and later.
The additional RT Core requires experimental V-Ray GPU Next build.
2. Keep in mind that between generations, cores are different. Turing cuda cores are much faster Pascal CUDA Cores.
3. We have a pre-release version of Turing with 4K CUDA cores for a while now. It runs very fast, but we can’t share performance numbers for pre-release hardware and drivers.

Best,
Blago.

[Nicinus]

Is NvLink significantly different than SLI? I remember that SLI never really mattered in anything but games as applications would have to be programmed directly in order to benefit. Also, since Vray already can use multiple GPUs, how would it benefit to have them linked? I guess the cards still have to have the same amount of memory.

[savage309]

Is NvLink significantly different than SLI? I remember that SLI never really mattered in anything but games as applications would have to be programmed directly in order to benefit. Also, since Vray already can use multiple GPUs, how would it benefit to have them linked? I guess the cards still have to have the same amount of memory.

NVLINK is different than SLI - it is much faster and allows the GPUs to share memory. So, 2 x 2080ti will give you 22GB of memory in theory, not 11GB. And it is supposed to work without any major slowdown.
This is new version of NVLINK though, and we haven't had the chance to test it yet, as our pre-release GPUs and drivers do not have it yet.

Best,
Blago.

[super gnu]

well thats the best news ive had all day.. i didnt realise 2080 would have nvlink. i tuned out of all the news about these cards when i realised they only had 11gb ram.

i hope x1000 that its not a nerfed version of nvlink, although given past nvidia pricing structures, i wouldnt be surprised if it was a highly cut-down version, otherwise they would take sales from their insanely expensive pro cards. ( no more than 2x cards sharing memory, castrated bandwith etc...)

if its proper nvlink it could finally be a good time to do gpu renderfarms.

[Jeffrey]

Sharing memory requires NVSwitch. I'm not sure if the 2080 has this capability.
https://www.nvidia.com/en-us/data-center/nvlink/

[super gnu]

i thought nvswitch was only required to link large numbers of gpus together.. id imagine a 2gpu system should not require a switch...? i mean, heck you can do unified memory over pcie now if you want things to go really slow.

[savage309]

NVSwitch is completely different thing, it is used in super computers.

Sharing memory in GPUs for rendering requires just NVLink.
We are trying to get the full details on the GeForce RTX NVLink and we will share them when possible.

Best,
Blago.

[savage309]

Long story short, GeForce RTX NVLINK work on Windows as long as you can enable SLI https://www.chaosgroup.com/blog/prof...idia-rtx-cards

Best,
Blago.

[Morbid Angel]

soooo complicated Just put 128 gb of ram on gpu already!

[Micha]

I think about what would be my next steps to upgrade my hardware. At the moment I use a GTX1080ti 11GB. I think, NVLINK will be the future, so my next step could be to buy 2x2080ti. So, I would get 22GB and approx. 4 times more cuda cores. Sounds like a good deal for a single workstation.

How could I use my GTX1080ti in this case? If my scene isn't heavy, could I use all three cards at 11GB and for heavy scenes only 2x2080ti at 22GB? Would I need to remove the NVlink or can I simple enable/disable the cards at the V-ray options?

[Muhammed_Hamed]

So, I would get 22GB and approx. 4 times more cuda cores. Sounds like a good deal for a single workstation.

2x 2080Tis are only 15% faster than 2x 1080Tis in GPU rendering now. They should be much faster whenever RT cores are supported in GPU renderers, but this will probably take a lot of time.

How could I use my GTX1080ti in this case? If my scene isn't heavy, could I use all three cards at 11GB and for heavy scenes only 2x2080ti at 22GB?

Having multiple cards with different VRAM, Vray GPU will be limited to 11 GB (memory of your 1080tis)
Also keep in mind that the 2080Tis produce a lot of heat. You will need blower style cards which are very expensive and not easy to find. FE cards doesn't use blower style coolers anymore
And considering the 1300+ USD price tag for the 2080Ti, it probably makes sense to buy these only if you need the extra VRAM from Nvlink.
Speed wise, the 1080ti is still a better value.

[Micha]

This helps me a lot again. Thank you very much.