I am interested in how this approach could realistically evolve - before we invest in any new hardware. For me, starting multiple denoise jobs on multiple nodes, using skip existing, sounds workable via the command line. But needing to copy all the frames to each node first will likely add another big time overhead. Is it likely to be changed so that the denoiser can work on frames on mapped drives? As I write this, I think I must be missing something - how can skip existing work when the denoised output is saved locally?
Results from my first sequence denoise:
330 frames multichannel exr @ 1280x720, c. 86Mb each
Denoise took 2.5 hours on a Nvidia GTX 780, which we are pretty happy with: an average of 27s per frame that produced a cleaner result than the previous frames we had rendered, that took an hour longer each.
As an aside, has anyone done a GPU denoise on a large (6K+) render yet? If so, what hardware? Our large renders typically have many elements and whilst I understand that only the RGB is denoised, I figure the GPU will need to load a lot, if not all, the elements into memory first? Please put me right, if I am being dim.