Hi all,
I’ve noticed when rendering that CPU+GPU is slower than just cpu alone, could someone shed some light on this please?
Any help would be greatly appreciated
5800x and rtx3070
Always has been, unless you have a slow GPU and a good CPU, or a scene that prefers CPU majorly over GPU
Hmm weird, do you know why that is the case?
Because the implementation is horrible. it is a known issue.
One thing you can do to maximize the usage of your PC is start a GPU render and a CPU render in 2 different blender instances.
Here is an example where I render with 2 blender instances, each instance gets 1 GPU (I have 2) this allows my CPU to get utilized more to denoise and to save images to disk.
Fair enough, do other software’s manage to do this better or is this just kind of one of those general 3D things
I am not sure, the only time it has a real problem is on powerful machines, a lot of mid range and low end machines (laptops) benefit from CPU + GPU but not very efficiently
CPU+GPU rendering in Cycles makes sense on the hardware where the CPU render times are similar to GPU render times. But with Cycles X and OptiX GPU times can be much lower compared to CPU, so in most cases with modern GPUs, CPU+GPU will not make sense.
1 Like
Correct me if I’m wrong but:
3070 with optix has 3417 score on opendata.
5800x score is 220.
Even in ideal world that only would be increase of 6.4% (3417->3637). But because of difficulty of balancing fast gpu rendering and slow cpu rendering you usually lose performance.
1 Like
Sounds about right to me
There are not that many renderers that support both GPU+CPU, or at least its something new so the tech isn’t there (yet).
Renderman XPU and Clarisse/Angie can do it and use most of the system resources, but this feature is still pretty new so there are bugs and instabilities.
Houdini’s new Renderer Karma is also going in that direction but AFAIK its still in beta.
Neither Arnold, V-Ray, Redshift nor Octane can do it.
In the high end segment of the industry there seems to be not that much demand for it. I don’t know for sure but I guess Renderfarms are either optimized for CPU or GPU, not both simultaneously.
V-Ray had hybrid render since V-Ray 3.6, year 2017.
I stand corrected.
I’ve never touched V-Ray, all the others I took for a test-ride - never V-Ray.
I remember testing this feature a while back and, as you, surprised it was actually slower. From my understanding my CPU was a bottleneck as my GPU is much faster. When the GPU was done rendering its part it was waiting for for the CPU to finish as well.
I think when the GPU is done rendering its part it needs to take over where the CPU currently is in the progress. How hard this would be to program I don’t know, apparently it’s not easy or the devs would probably already done so.
That had already been implemented in Cycles, and I guess in Cycles X some of that exists as well.
https://developer.blender.org/D9324
But when the GPU is much faster than the CPU I suppose that it is still not optimal, and I suppose the same happens with any render engine in these cases.
Ok, I see AMD helped optimizing this, that’s nice. Will be interesting to see if they ever manage to remove all bottlenecks no matter the speed difference on GPU and CPU. Maybe it’s impossible with just software solutions, maybe they need to create new hardware for that.
Will be interesting to follow this tech evolve nevertheless.