Page MenuHome

Cycles: Allow variable render tile size for mixed gpu/cpu
Closed, InvalidPublic


Blender now has the ability to render using both CPU and GPU in a single pass ( However, cpus desire smaller tile sizes (32x32) and gpus prefer larger (256x256). On several machines I've tested, it's often the case that the mixed mode of rendering yields to no benefit to using either pure cpu or gpu (on the same machine) .I'd like to propose a major optimization where the tile size in a single render pass are different between cpu and gpu.

One method could be to divide the rendering area first into larger tiles (like 256x256) that would be optimal for the GPU to render each tile. However, tiles can become owned by the cpu renderer, which would subdivide these large blocks into subdivisions (32x32 or smaller) to process.



Event Timeline

I've looked into this and it's definitely doable. The basic changes are easy, but there are two tricky parts:

  • The render engine API expects uniform tile sizes, that would need to be changed,
  • For denoising a uniform grid is needed. However, it wouldn't be too hard to merge smaller tiles back into large tiles before queueing them for denoising.

That being said, the bugtracker is not the right place for feature requests, so I'll be closing this.

Also the whole 'gpu likes larger tiles' no longer holds true since rB6da6f8d

This comment was removed by Jonathan Dunlap (jadbox).

@Lukas Mandok (Lukas) Stockner (lukasstockner97) thanks for the info, where's the right place for this?

The default paste for a feature requests has some links

Hi, while we love to hear these ideas, this is not the right place for them. This website is mainly used to track bug reports and not feature requests. But luckly, there are some other nice solutions you can take, so please use one of the other >forms of communication listed below:
Fun board mailing list:

It is impossible to stress how important this is. Most people dont understand enabling CPU+GPU uses nearly twice the power to render 3 seconds faster. If this is not optimized blender is solely responsible for global warming xD

Also the whole 'gpu likes larger tiles' no longer holds true since rB6da6f8d

That is absolutely incorrect. I use blender every day of my life on production scenes and never is a tile size below 196x196 faster than 256x256

When this was initially committed you could see a huge improvement especially at 32x32 it was even faster than rendering at 256x256

Now that is not true at all 256x256 is once again optimal for some reason

Yes, I noticed it too lately, there is not much of a difference, at least not like before.

Maybe some regression?