Let the GPU take over the last Tiles from CPU (combined rendering)
Cycles now has combined rendering of CPU and GPU. While this option (at least for me) usually is faster at least for heavy scenes, the bottleneck seems to be, that the last tiles take ages to converge, because thenow idle GPU already did it's last tile and the CPU is still rendering the last tiles alone (much slower than the GPU would).
So my suggestion is, that as soon as the GPU is finished with it's last tile, it "takes over" the next best unfinished Tile that is currently rendered by a CPU thread and finishes it. Then it jumps to the next Tile and so on.
I imagine this is very hard and probably hacky to implement - it's just something that I thought could sqeeze out some more juicy rendertime.
4 Comments
And V-ray has a cool trick that dived the last tiles to smaller tiles and use all free cores for render them.
blender.community/c/rightclickselect/CWbbbc/previous-frame-performance-tile-order-in-cycles-to
That would be a nice thing
It is painful to see a render that is complete minus a handful of tiles. Sometimes the last four tiles take 15-25% of the render time. I think this proposal could be very doable. It may take longer to render in some small percentage of cases, but I think if you base the decision to steal a tile from the CPU or not on average render times I think in the vast majority of cases it would speed up rendering.