NVIDIA Multi-GPU Support

[Ryan Chernin]

How much of a performance boost do you see using multiple graphics cards? Does it truly double your available VRAM, and more importantly, does Davinci properly utilize it?

I currently get perfectly smooth playback of RED raw, until I introduce a NR node (which is to be expected). I've been starting to wonder if a second card could be allocated to just effects/color/NR/etc.

 

[Michael Lindsay]

It all really depends on the ballance of your system and if you have Davinci using GPU for both dewavlet and debayer or just debayer??

Just recently I moved from a single Titan Xp card (for compute) to 1x RTX A5000 but then quickly to 2x RTX A5000..

I would say no it doesn't currently double your VRAM but that could change at some point ... BUT with ref to decoding Red 8k files my performance nearly doubled BUT NR can really change that.. TNR is different architecturally from Spatial NR for example. What kind of NR do you use?

 

[Ryan Chernin]

It all really depends on the ballance of your system and if you have Davinci using GPU for both dewavlet and debayer or just debayer??

Just recently I moved from a single Titan Xp card (for compute) to 1x RTX A5000 but then quickly to 2x RTX A5000..

I would say no it doesn't currently double your VRAM but that could change at some point ... BUT with ref to decoding Red 8k files my performance nearly doubled BUT NR can really change that.. TNR is different architecturally from Spatial NR for example. What kind of NR do you use?

I currently have it set to use GPU for both decompression and debayer.

For NR, I prefer to use Neat Video, since it seems to produce better visual results and the GPU acceleration seems to work a bit faster than Davinci's native NR plugin. Usually I select both TNR and SNR, but I'm also experimenting with just Temporal alone.

I just swapped out my (already formidable) 2080 ti for a new 3090 Founder's Edition, so I wouldn't say I'm currently lacking graphic performance haha. More so I was just curious how multi-gpu setups are supported in Davinci. If it's worth it, then it's something I'd do far down the line, but it was already hard enough getting hold of a single 3090, and the 24gb VRAM is performing *very* well at the moment.

 

[Michael Lindsay]

I can’t talk to neat videos performance as I haven’t tested in a while.. but Resolve is very multi GPU scaleable (not far off linear going from 1 to 2 GPUs) for many operations but unfortunately things like TNR are complicated to do well on more than one card.. you may also have a real hassle powering and cooling 2 x 3090 cards unless you are keen computer guy..

 

[Simon Dunne]

OS plays a major factor if you want to really see the benefits from multi GPU. Linux is what you're after.

 

[Brody_B]

It’s been a while so don’t remember all the details from when I was testing my pc build, but Resolve does do a pretty good job at dual gpu workload distribution, you won’t get full 2x everything with a second gpu but I would say it’s still with it if you have the $$.

The vram does not double, some ram usage is split between the cards but as soon as one is full your done, for ex there may only be 30-50% usage (not an exact number) on one gpu, with the other maxed and a restart required.

But it does a better job at spitting processing tasks. When I got to 3 gpus (titan rtx) was when one was never fully used but the other 2 could be 99% usage each.

I also use neat video, on my system it brings things to a studder, but I still prefer its quality over the others. I think you can allocate a single gpu for processing within Neat preferences. I haven’t tried that, but when I allocated all 3 gpus it didn’t like it, vram filled immediately, so I have it on CPU processing only. I need to go back and run some more tests with that.

Pretty much for all other processing tasks I have a raw workflow thru and thru with debayer reduction for when things get thick with lots of nodes etc. You could probably get that on 2 depending on how much more powerful the newer 3090’s are.

Side note, 24gb vram is great, not often does it max out, only sometimes here and there. I would consider it to be a minimum amount for 8k workloads. A lot of people learn the hard way that buying anything less your just not able to do some important intensive processing, soon as that’s maxed your done.