@Ozu I haven’t had the chance to do another test yet, but I think you’re right, it seems it is not multi-device yet. So, those render time I got are quite something compared to 2.93.

Note the cycles-x builds on are made with an older CUDA toolkit version. We got better performance with CUDA toolkit 11.3, and that’s how both master and cycles-x were built for the graphs in the blog post. We’ll upgrade the buildbot CUDA toolkit to fix this.


I’m glad that Cycles is back on track with these improvement even though i mostly use Eevee & UE4 nowadays but I do hope they keep the ball rolling because that speed is much needed.
No better way to start 3.0 train than this.

Some interesting snippets.

  1. Volumes. They plan to add ray marching algorithms. Does this open up the possibility of generating surface normals for volumetric materials, allowing volume materials to exhibit glossy/specular reflections. The link below suggests it’s possible to apply surface shading to a volume defined via ray marching algorithms:
  1. Path Guiding. Will this finally give Cycles the opportunity to generate reasonable caustics and possible make portals redundant?

  2. Introduce more advanced rendering algorithms. May we see MLT or Bidirectional options at some point?

Overall, i’m pretty excited to see where this all leads. Bring on the next 6-12 months :slight_smile:


Yes. Brecht clearly said that in livestream.

1: I have no idea. But currently, you can use Volume to Mesh modifier and composite render of a volume and render of a corresponding surface.


It seems Cycles X does not work with multi GPU at the moment (both Windows and Linux).

My internal scene (open space, 16 internal lights, HDRI), 500 samples:
Cycles X
1 GPU: 1:10,72
6 GPU: 1:24,70

Cycles (2.93)
1 GPU: 1:39,44
6 GPU: 0:18,42

I didn’t see any mentions of this but are they thinking of updating the Disney principled shader ? I feel it’s a bit outdated… , to me the fresnel especially behaves better in other renderers default universal materials (redshift, octane) and getting some kind of micro roughness approximations would be nice…


Yes. It was explained that was not supported into Blender.Today livestream with Pablo.

I am re-posting link to it because it is essential and give a lot of answers.

no it doesn’t, they mention it multiple time in the stream, if you re going to test the branch at least read the blog post and the technical presentation ( link in the blog post ) so you’ll know about the current limitation

Yep - but being able to do it via a material would be much simpler and would save having huge amounts of additional geometry to contend with, especially for very detailed volumetric materials (like foam).

1 Like

here the document on PathGuidinginProduction

I hope AMD will help blender developers to get it working on their hardware asap.


Well they are swimming in that Zen money, so they don’t have any excuses.


Hi, it was mentioned in the presentation, BF is working with developer from AMD and Intel to get it back to work on there GPU´s.

Cheers, mib


Nice. Too bad they don’t expect it to work on release.
I wonder how are they gonna do it - vulcan compute?

Tested it on a rtx 3060 and the speed is crazy. Magic to me.

CUDA took 2 min longer.


I’ve made a test with the junk shop scene, i had to remove some light and change the setting

Setting changed ;
render samples to 2048
Activated AS
Removed denoiser
Clamp indirect at 10
Removed compositing

Also changed a shader from volume scatter to a default diffuse, there’s still a bit of a difference between them, apparently cycles x have a problem with MIS, to fix some light i had to desable MIS but it introduced a lot of noise so i removed them

Hardware : Ryzen 3800X and 3060Ti on linux, rendering with Optix only

2.93 render time : 3min14sec

cycles X 3.0 render time : 2min2sec

Viewport Performance

Cycles X

Regular cycles


Just tried it on a 1070, this old interior scene took 20min on 2.93, and only 9min on the Cycles X build :nerd_face:


One more test with an old scene, 1024 samples for both: