I enabled Cuda for 1x3090 and 2x2080ti, but all cards are running quite low only. Is there a way to get the full GPU power?
-Micha
I disabled the 2x2080ti and now I get ~42% at the 3090.
RH-57576 Multi-device broken. Currently working on it.
Do you know why I don’t get 100% per single 3090?
Maybe not heavy enough scene.
If the scene needs 20s at 40% for rendering, why should it not be done in 8s at 100%? For animations I hope to get frames rendered within seconds.
Higher GPU utilization doesn’t necessarily mean faster rendering, since it could be just a sign of wasted energy.
The internal tilesize setting is set to work well for most scenes and such that it is fast.
You could make it smaller and get higher GPU utilization - but then the rendering could also be quite a bit slower. I did just a test with a model that utilizes ~79% with the default setting. When I decrease the tile size I get up to 90% GPU utilization, but the rendering speed is ~5 times slower.
You’ll see higher utilization for more complex scenes where complex means more rays to bounce. Lots of reflective and transmissive materials will do that. If you have only an object with a simple diffuse material and not that complicated lighting you will not see high GPU utilization.
Interesting.
If the usage is 40% and the GPU stay cool, than the power isn’t used. Right, wasted power wouldn’t be nice. Best would be 100% of the power and speed up the rendering. It’s a pity that simple scenes can’t be rendered at full speed within a few seconds. I hope in the future it will be possible. Here a test with different single GPU.
20s per 3090
35s per 2080ti
Micha, could we get this scene, I’m curious to compare on iray.
Dunno, on my RTX A6000 simple scenes do render in matter of seconds (and GPU utilization will stay very low).
Anyway, I’m going to chip away at the multi-device problem now. Hopefully I’ll have a workable fix sooner rather than later.
I found the reason for the slow down. There is a throttle option. I didn’t expect something like this. It was at default 5ms and at 0ms I get ~90% GPU usage. Without throttle I don’t get any display problems. I’m not sure it good to set it a 5ms per default.
Interesting is that the Intel denoiser is a very strong throttle. It looks like I get ~10% GPU usage (it jumping up and down between 1%…70%). The Nvidia denoiser is a little better at the beginning, but also not good.
At V-Ray is an option to use the denoiser only after the rendering is finished. Where is this option for Cycles?
Solid red area without denoiser and two last renderings was Intel and Nvidia.
5ms is to cater for those with not so powerful GPUs, especially if there is only one GPU which also drives the monitors.
For the viewport I don’t think there is an option. Using Rhino Render you can just let the render go, and at the end toggle the denoiser. But that is manual.
I logged RH-63211 Option for denoiser to only apply at the very end for this feature request.
1 Like
Multi-GPU rendering will be back in Rhino 7.7. Service Release Candidate due next Tuesday-ish (barring any stop-ship issues).
2 Likes