Resolve 18.6 - Slow GPU renders

[Ferdinand13]

Hi there,

I've got a bunch of related questions about GPU usage in NV using version 5.5.11 inside Resolve 18.6.2 on a Windows 10 machine. Let's get going.

I've been doing a couple of denoising tests over the last week, in order to speed up my ProRes RAW workflow. For reference, I'm recording ProRes RAW on my Atomos Ninja V then converting each clip into a Cinema DNG sequence, so I can do RAW post production inside Resolve.

My system - using three NVIDIA 2080 Super cards - is playing the CDNG timeline in full resolution (essentially 4K DCI) in realtime. Of course, once you add Neat Video to denoise the grainy RAW footage, things slow down to a snail's pace. No surprise here.

My issues are as follows. First of all, using the "Use GPU only" mode is yielding the fastest results, 6.96 frames/sec according to the "Check Speed" button. Having already discovered how unstable Resolve gets when NV has got access to all three GPUs (the dreaded "GPU memory full" error), I've settled on a setup, where GPU 1 is reserved to Resolve, while 2 & 3 are used for NV only.

Screenshot 2023-10-14 193935.png (38.7 KiB) Viewed 1030 times

However, this still causes "GPU memory full" errors about 50% of the time, plus the actual render speed upon export stays between 3 and 3.5 frames/sec, so half of what was promised.

Even stranger, when I dedicate 2 GPUs (1 & 2) to Resolve and only one (GPU 3) to NV, I get a slightly faster speed test (7.01 frames/sec) and the actual render speed stays at the 3 to 3.5 frames/sec mentioned above.

So my findings so far are that Resolve becomes unstable when I allocate more than one GPU to NV and there is actually no benefit of doing this in the first place, as one GPU yields the same frames/sec as two.

I mean, this of course perplexes me a bit. After all there is more processing power, so shouldn't an extra GPU give me at least some boost? But even if I would let this slide, why am I not getting the 7 frames/sec that I've been promised, but instead only 3.5 frames/sec max?

Using NV is slow enough as it is, so if I could get at least the 7 frames/sec, I'd already be a lot happier, even though I am really displeased by the constant "GPU memory full" errors that I'm getting, even when I'm just clicking the "Optimize Settings..." or "Check Speed" buttons. NV is definitely not playing nice with the hardware.

Bottom line, how can I make NV faster and more stable? Thanks for letting me know.

[Ferdinand13]

OK, a quick update. I restarted my system - as I've done dozens of times during my tests - and still ended up with the same results speed-wise. Annoyingly so, even a restart didn't "flush" the memory on my GPU 3, which happens all the time. Any idea why this happens?

Anyway, just for fun I decided to allocate GPU 2 to NV and GPUs 1 & 3 to Resolve. The "Check Speed" button gave me the same result as before, around 7 frames/sec.

Screenshot 2023-10-14 204015.png (42.72 KiB) Viewed 1028 times

However, to my complete surprise, when I did my test render, I actually got between 6.5 to 7 frames/sec this time, resulting in a render of 39 seconds instead of 1 minute and 18 seconds. What is going on?

I then disallocated GPU 3 from Resolve and did another "1 GPU only" test with NV, using GPU 3. This yielded a slightly slower result (6 to 6.5 frames/sec), which was still faster than before. I then set GPU 2 as my NV GPU again and now I can't reach 7 frames/sec anymore, but things stay at the 6 to 6.4 frames/sec I got with GPU 3.

Suddenly things are working, when they didn't during all my other tests. I'm happy things have sped up, but equally confused as to what actually caused this.

[NVTeam]

Resolve uses the GPUs, its codecs may use the GPUs, Neat Video uses the GPUs and possibly other software in the system use the GPUs, so there is some overlap and competition for the same resources that can take place. So there is nothing unusual that they affect each other in somewhat unexpected ways.

Please remember that the speed measured by Neat Video in Preferences is the speed of Neat Video alone, without any overhead of Resolve and its part of the workflow. The full workflow will always be slower (take more time to process each single frame) than Neat Video alone, which is just one part (just one part of the frame processing time) of the workflow. This is described by efficiency of the host application and can be measured directly as described in this article.

To optimize editing and processing specifically in Resolve, please also see the ideas from this article.

Thank you,
Vlad