3D Rendering on Apple Silicon, CPU&GPU

[leman]

Just read the 7900 xtx creators review on techgage. Would be good to do these tests for Macs as well. https://techgage.com/article/amd-radeon-rx-7900-xt-radeon-rx-7900-xtx-creator-review/
I tried out 2 of the tests since the blender scenes are just standard scenes that are freely available.
Maybe I did something wrong but just testing the first blender example scene renders in about 9 mins(canceled the render at 20% after 2 mins) on my 16" m1 Max (32). On a 7900xtx that render takes less than 2 minutes in total and a 4090 finishes in just 21 secs... I guess that is a scene that is extra well suited for RT HW with all the small details but using quite simple shading. However, the diff in speed between the M1 series and PC based solution is quite extreme and shows why there is a need for RT hardware on a pro machine if it is to be used for 3D beyond a few stills from time to time. I also quickly tested the eevee scene and got render times for 100 frames in around 290s (slightly slower than a nvidia 3050 but faster than a radeon 6500xt. 7900 xtx did the 100 frames in 65 sec)
Eevee tests done on Blender 3.5 alpha using metal viewport.

I know some of you like to say that render times doesn't matter and we should not compare to PC. I totally disagree and we must know what the alternative solution is at. All in all, we better hope that Apples m2 pro/max/ultra chips have HW RT or we will have to wait another 1.5 years hoping for the M3 series.

Just for curiosity, I did a quick test and can confirm your results (using latest released Blender 3.4.1). I got 20 min on CPU and 9 min on GPU on M1 Max when rendering the Scanlands scene.

One thing to note is that hardware utilisation is poor on both the CPU and the GPU. For example, the GPU only draws 16-18 watts when rendering this scene suggesting that there are a lot of stalls preventing the ALUs from being utilised fully (under full utilisation the GPU can draw 40 watts).

 

[aytan]

Just for curiosity, I did a quick test and can confirm your results (using latest released Blender 3.4.1). I got 20 min on CPU and 9 min on GPU on M1 Max when rendering the Scanlands scene.

One thing to note is that hardware utilisation is poor on both the CPU and the GPU. For example, the GPU only draws 16-18 watts when rendering this scene suggesting that there are a lot of stalls preventing the ALUs from being utilised fully (under full utilisation the GPU can draw 40 watts).

Blender Version 3.4.0 (Release date was 2022-12-07), 48 GPU core Ultra Scanlands scene took 5:15:02 render out, with Metal RT on same scene took 3:57:01
looks like Metal RT get some improvement base on that scene.

 

[Xiao_Xi]

looks like Metal RT get some improvement base on that scene.

Does Metal RT improve GPU utilization?

I also quickly tested the eevee scene and got render times for 100 frames in around 290s (slightly slower than a nvidia 3050 but faster than a radeon 6500xt. 7900 xtx did the 100 frames in 65 sec)

Isn't the Eevee benchmark a better metric of how powerful a GPU is?

 

[aytan]

Does Metal RT improve GPU utilization?


Isn't the Eevee benchmark a better metric of how powerful a GPU is?

I have no idea, these are the results I get. Someone who knows entire Metal backend in Blender could explain

 

[innerproduct]

I have no idea, these are the results I get. Someone who knows entire Metal backend in Blender could explain

Both eevee and cycles are good tests for the GPU. Cycles is pure compute while eevee also uses the rasterization hw and currently doesn’t benefit from RT hw. But, different scenes tax the system very differently so it is good to try a few different ones. For example, scenes with massive amounts of geometry but having simple shading is prime candidate for RT hw. Theoretically the m1* series should be great at scenes that needs many large assets and textures and maybe more complex shading. Would be fun to find that kind of perfect scene.
We find very interesting that your 48 core ultra scales so well on that scene and that metalRT is faster for once.

 

[ader42]

I enjoyed this today:

 

[aytan]

Both eevee and cycles are good tests for the GPU. Cycles is pure compute while eevee also uses the rasterization hw and currently doesn’t benefit from RT hw. But, different scenes tax the system very differently so it is good to try a few different ones. For example, scenes with massive amounts of geometry but having simple shading is prime candidate for RT hw. Theoretically the m1* series should be great at scenes that needs many large assets and textures and maybe more complex shading. Would be fun to find that kind of perfect scene.
We find very interesting that your 48 core ultra scales so well on that scene and that metalRT is faster for once.

I have tried other scenes, junkyard scene there is no difference Metal GPU/MetalRT GPU, BMW scene there is no difference Metal GPU/MetalRT GPU, but in the White Lands Scene MetalRT GPU is slower than Metal GPU.
White Lands MetalRT GPU 2:50:38 / Metal GPU 2:34:50 even tough it's a similar scene as Scanlands by memory usage.
If there is a kind of modeling/shading/texturing formula for MetalRT it could be useful for everyone. I want to understand why MetalRT behaves faster result on that particular scene Scanlands but this subject is beyond my knowledge.

 

[Xiao_Xi]

Both eevee and cycles are good tests for the GPU.

Cycles uses ray tracing, while Eevee uses rasterization. In Cycles, Nvidia GPUs get much better results than any other GPU because they are the only GPUs with hardware acceleration support for ray tracing.

I have tried other scenes, junkyard scene there is no difference Metal GPU/MetalRT GPU, BMW scene there is no difference Metal GPU/MetalRT GPU, but in the White Lands Scene MetalRT GPU is slower than Metal GPU.
White Lands MetalRT GPU 2:50:38 / Metal GPU 2:34:50 even tough it's a similar scene as Scanlands by memory usage.
If there is a kind of modeling/shading/texturing formula for MetalRT it could be useful for everyone. I want to understand why MetalRT behaves faster result on that particular scene Scanlands but this subject is beyond my knowledge.

Someone in the thread for Cycles on macOS in the Blender forum could shed some light on the situation. It might help Apple to optimize Cycles.

 

[jujoje]

Looks like Otoy have released OctaneX for iPadOS:

Octane X App Launches for M1 and M2 iPads Bringing State-of-the-Art Cinematic GPU Path-Tracing to Mobile Workflows

OTOY is proud to officially release Octane X for M1 and M2 iPads bringing the industry’s first and fastest spectral GPU renderer to mobile production workflows – with an all new UX supporting the Multi Touch design of iPad and with support for Apple Pencil – for next generation visual effects...

home.otoy.com

Don't have an M powered iPad to test it on, but it looks pretty neat; would be pretty cool to go Nomad Sculpt -> Procreate -> Octane (wonder how painful the workflow ends up being). Nice to have on the iPad either way though.

 

[avkills]

Cycles uses ray tracing, while Eevee uses rasterization. In Cycles, Nvidia GPUs get much better results than any other GPU because they are the only GPUs with hardware acceleration support for ray tracing.


Someone in the thread for Cycles on macOS in the Blender forum could shed some light on the situation. It might help Apple to optimize Cycles.

AMD 6000 and 7000 series cards also support hardware raytracing; whether it is supported in that particular software package is another thing...

 

[jmho]

AMD 6000 and 7000 series cards also support hardware raytracing; whether it is supported in that particular software package is another thing...

I believe AMD 6000 / 7000 cards are getting HW RT in cycles in 3.5 when the HIP backend is ready.

 

[jeanlain]

I have tried other scenes, junkyard scene there is no difference Metal GPU/MetalRT GPU, BMW scene there is no difference Metal GPU/MetalRT GPU, but in the White Lands Scene MetalRT GPU is slower than Metal GPU.

What exactly is the difference between Metal GPU and MetalRT GPU? The former uses Metal for ray tracing, but doesn't use the Metal ray-tracing APIs?

 

[leman]

What exactly is the difference between Metal GPU and MetalRT GPU? The former uses Metal for ray tracing, but doesn't use the Metal ray-tracing APIs?

Yes, I believe this to be the case. What's interesting is that the traditional compute APIs are, well, compute-only. CUDA does not expose the RT functionality of Nvidia GPUs for example, you have to use a different toolkit if you want to use Nvidia raytracing. So Apple engineers started to adapt the CUDA code to make it compatible with Metal shading language; adding RT is a different.

 

[aytan]

What exactly is the difference between Metal GPU and MetalRT GPU? The former uses Metal for ray tracing, but doesn't use the Metal ray-tracing APIs?

As I mentioned before I have no idea what is the difference, but there is render time differences without any logic. So I asked if there is anyone who could have correct information about subject. There had to be a difference otherwise there is no logic this options ready on Blender UI. By the way on Blender 3.5.0 Alpha release you can select your GPU backend as OpenCL or Metal.

 

[Xiao_Xi]

So Apple engineers started to adapt the CUDA code to make it compatible with Metal shading language; adding RT is a different.

MetalRT follows OptiX implementation.

MetalRT closely follows the Optix implementation, and in some cases (notably handling of transforms) it makes sense to extend Optix special-casing to MetalRT.

Blender Developer - developer.blender.org

Join the global community of Blender developers.

developer.blender.org

 

[iPadified]

Looks like Otoy have released OctaneX for iPadOS:

Octane X App Launches for M1 and M2 iPads Bringing State-of-the-Art Cinematic GPU Path-Tracing to Mobile Workflows

OTOY is proud to officially release Octane X for M1 and M2 iPads bringing the industry’s first and fastest spectral GPU renderer to mobile production workflows – with an all new UX supporting the Multi Touch design of iPad and with support for Apple Pencil – for next generation visual effects...

home.otoy.com

Don't have an M powered iPad to test it on, but it looks pretty neat; would be pretty cool to go Nomad Sculpt -> Procreate -> Octane (wonder how painful the workflow ends up being). Nice to have on the iPad either way though.

This is crazy. Think that we actually have GPU rendering on iPads.

 

[Xiao_Xi]

What exactly is the difference between Metal GPU and MetalRT GPU?

Tech talk 2020:

Discover ray tracing with Metal - WWDC20 - Videos - Apple Developer

Achieve photorealistic 3D scenes in your apps and games through ray tracing — a core part of the Metal graphics framework and Shading...

developer.apple.com

Tech talk 2021:

Enhance your app with Metal ray tracing - WWDC21 - Videos - Apple Developer

Achieve photorealistic 3D scenes in your apps and games through ray tracing, a core part of the Metal graphics framework and Shading...

developer.apple.com

Tech talk 2022:

Maximize your Metal ray tracing performance - WWDC22 - Videos - Apple Developer

Learn how to simplify your ray tracing code and increase performance with the power of Metal 3. We'll explore the GPU debugging and...

developer.apple.com

 

[Xiao_Xi]

It is looking more and more likely that "Full Metal" Blender will be Blender 3.6, and not Blender 3.5 because the Cycles optimizations and Viewport compositing support are still missing.

Cycles Apple Silicon optimisations

**Tasks** For tracking Apple Silicon specific optimisations (macOS 13.0+). - [x] Architecture specific parameter tuning - [x] Threads / Threadgroup - [x] Num concurrent states / busy states - [x] Work scheduling improvements - [ ] Add new multi-device abstraction for sharing WorkTileSchedul...

developer.blender.org

• Metal backend can be enabled in Blender 3.5 builds via the user preferences. People have already been tested and giving feedback. No show-stoppers found and Metal backend can be targetted as a secondary backend for Blender 3.5
• The metal backend doesn’t support Viewport compositing yet. For this SSBO’s and Compute shaders support should be added. This is high on the priority list and looks similar to previous changes. Most impact would be platform support to make sure that it is stable on all Apple devices. Goal is to do it as soon as possible. Indirect drawing would be post-poned to be tackled later.

2023-01-09 Eevee/Viewport Module Meeting

Practical Info This is a weekly video chat meeting for planning and discussion of Blender Eevee/viewport module development. Any contributor (developer, UI/UX designer, writer, …) working on Eevee/viewport in Blender is welcome to join. For users and other interested parties, we ask to read...

devtalk.blender.org

 

[BootLoxes]

ZBrush 2023 | ZBrush Docs

Zbrush is finally native!

 

[singhs.apps]

ZBrush 2023 | ZBrush Docs

Zbrush is finally native!

Yes. And they bothered to add local symmetry and charge you 450 bucks for that.

 

[l0stl0rd]

Yes. And they bothered to add local symmetry and charge you 450 bucks for that.

Yes

 

[ader42]

I think the Redshift inside ZBrush is going to be brilliant for me.

To be honest, I think most of the new features will be useful for me.

 

[Xiao_Xi]

Apple's optimization of Blender is already paying off.

Introducing the Blender Metal Viewport — Blender Developers Blog

Next to Metal-accelerated computing in Cycles, it is now possible to enjoy improved EEVEE viewport (and UI) performance as well.

code.blender.org

Do other 3D programs show a similar improvement?

 

[aytan]

Apple's optimization of Blender is already paying off.

View attachment 2143149 View attachment 2143148

Introducing the Blender Metal Viewport — Blender Developers Blog

Next to Metal-accelerated computing in Cycles, it is now possible to enjoy improved EEVEE viewport (and UI) performance as well.

code.blender.org

Do other 3D programs show a similar improvement?

I will give a try, I have my render times for previous versions and could figured out improvment.

 

[avkills]

Metal works way different than OpenGL; so other than stubborn developers who refuse to update to use Metal; things seem to be looking good. Here's hoping Apple releases a monitor PCI-E based GPU with the new Mac Pro that also works in the older 2019 model as well. (I know, doubtful...lol)