more than 2x nvlink cards issue.

[super gnu]

Ive heard conflicting reports regarding this feature.

iirc, Vlado stated that although it currently works, this ability will be removed from future nvidia drivers, meaning that if you have more than 2x 2080 cards, nvlink memory sharing will not function at all, and you will be back to single card memory limit.

however ive seen numerous reports since then of people using many cards together without issue.


i cant imagine they could reasonably do this.. they would infuriate a number of people who currently use multiple cards with nvlink.. (customers who have spent a mountain of money with them)


so is there any more news re. this stupid "feature" they might introduce..?


my plan was 2x titan rtx to start, with option of another two later.. (spending £12k plus on a workstation is beyond my means at the moment...!)

another thought was 2x titan rtx, plus a third card (titanX maxwell i already have here) for viewport duties. would adding a third non-nvlink card like this *also* ruin nvlink functionality if these "improved" drivers come out?

[Muhammed_Hamed]

For NVlink to work you need to connect one of your Titan RTXs to a monitor.. you cannot use a third card for viewport, maybe you can use the Titan X for rendering too if your scenes fits in 12 GB of VRAM
Many people are using multiple pairs, maybe with old drivers ,but no idea how this will play with future drivers.
Nvidia wants users to use Quadro for any kind of rendering, so disabling this in future drivers will not be surprising

[super gnu]

For NVlink to work you need to connect one of your Titan RTXs to a monitor.. you cannot use a third card for viewport, maybe you can use the Titan X for rendering too if your scenes fits in 12 GB of VRAM
Many people are using multiple pairs, maybe with old drivers ,but no idea how this will play with future drivers.
Nvidia wants users to use Quadro for any kind of rendering, so disabling this in future drivers will not be surprising

ahh that is very handy to know. i didnt realise nvlink had this requirement.

havinng said that i have dual monitor setup.. could i have second monitor (used normally for web stuff/menus etc) connected to nvlink pair, and titan x for main monitor where i have max sitting?

[Muhammed_Hamed]

ya it should work, and after booting to windows and verifying the NVlink works you can remove the monitor from the Titan RTX

[danio]

So if you have (4) 2080 Ti’s, both pairs linked with NVlink, do you have to connect the monitor to both pairs for Windows to recognize both NVlinks?

[Muhammed_Hamed]

So if you have (4) 2080 Ti’s, both pairs linked with NVlink, do you have to connect the monitor to both pairs for Windows to recognize both NVlinks?

yes for old drivers that allows NVLink to work on multiple pairs...but in newer drivers you cannot use Nvlink with multiple pairs(only one pair)
Multiple pairs work fine with Quadros only
I only have 2 cards so I cannot really tell if this changed or not .. You have to ask Dabarti on Sebstian Michaliski on Facebook

[danio]

yes for old drivers that allows NVLink to work on multiple pairs...but in newer drivers you cannot use Nvlink with multiple pairs(only one pair)
Multiple pairs work fine with Quadros only
I only have 2 cards so I cannot really tell if this changed or not .. You have to ask Dabarti on Sebstian Michaliski on Facebook

Ok thanks for the information! What a crippling effect. I suppose a person could always do a pair of Nvlinked 2080 Ti’s and then add a 3rd Titan RTX and not be limited in RAM. I’m sure at some point VRay will require newer drivers so staying on the old seems like it might not be a permanent solution...

[zero-13]

super gnu I was wondering do you need to NV link to have more VRAM for larger scenes?

I ask because all the tests I've seen, and if I'm interpreting what I'm seeing correctly, is that NV Link gives you overall slightly slower rendering.

[squintnic]

barely barely slower, the extra VRAM makes a massive difference to what it can handle!

[Muhammed_Hamed]

super gnuis that NV Link gives you overall slightly slower rendering.

slower by 1-2% for few cases only.. See the tests that Chaos did awhile back, and Vlado said they will wok on optimizing it more

[zero-13]

Good info, thank you for clarifying guys.

So for sure the benefit is larger scene handling, but would a pair of NV Linked GPUs potentially ever be faster than unlinked after optimizations?

I have limited and dated understanding of linking GPUs from gaming way back when, but usually linked resulted in better FPS for games... so naively not fully understanding the technicalities, I thought it might improve rendering speed also but was surprised to see slightly slower results in some cases.

[a7az0th]

I wouldn’t expect it to be faster. For gaming SLI was faster as it allowed you to start using the second card’s firepower to assist in the game rendering, but V-Ray has always been able to do that without the need of SLI. Actually we recommend not using SLI for V-Ray as it could decrease perf, rather than increase it. Obviously this is not the case for non-Quadro RTX cards though, where NVlink can only be enabled IF the cards are in SLI mode.

so connecting cards with NVlink for rendering will always have some minor perf hit, maybe negligible, but it cannot make rendering any faster. We already use the firepower the GPU can give, even without the NVlink

[super gnu]

I wouldn’t expect it to be faster. For gaming SLI was faster as it allowed you to start using the second card’s firepower to assist in the game rendering, but V-Ray has always been able to do that without the need of SLI. Actually we recommend not using SLI for V-Ray as it could decrease perf, rather than increase it. Obviously this is not the case for non-Quadro RTX cards though, where NVlink can only be enabled IF the cards are in SLI mode.

so connecting cards with NVlink for rendering will always have some minor perf hit, maybe negligible, but it cannot make rendering any faster. We already use the firepower the GPU can give, even without the NVlink

could you shed any further light on the original subject, regarding newer driver changes to disable nvlink memory sharing on more than 2x consumer rtx cards..?

it will have some effect on my buying decisions in the next month or two..

[Macker]

Any update on this? I assumed (perhaps wrongly) that you could potentially create a GPU rendering node with 4 pairs of RTX titans giving you a total usable VRAM of 48GB across all cards?

[super gnu]

that would be a fine, if rather costly machine there is a distinct radio silence regarding the nvlink limitation that was supposed to occur in newer drivers..