Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

3D Rendering on Apple Silicon, CPU&GPU

Xiao_Xi said:
Last Blender 3.1 alpha has the Metal support for Cycles, and first impressions are excellent. Has anyone tried it?
devtalk.blender.org

OpenCL on Apple M1

And now it is finally done!!! I just tested the upcoming Metal patch and it works nicely with the BMW. M1Max 32GB/24GPU: 48 seconds. Classroom: 2:06 min Monster under Bed: 3:35 min (MacBook completely silent and on battery, nice!!)
devtalk.blender.org devtalk.blender.org
Click to expand...
no waiting for the official build. I do not feel like trying to compile it myself.
 
Last edited:
Xiao_Xi said:
Last Blender 3.1 alpha has the Metal support for Cycles, and first impressions are excellent.
Click to expand...

Nice! As a tl;dr from the blender thread, the times on the BMW scene looked like:

29s - 2070 Super (Optix)
48s - M1 Max 24GPU
51s - 2070 Super (CUDA)
2.04 min - Mac mini M1

Which seems pretty good (although also from the thread, the results may not be representative because alpha builds/comparing between 2.9x and 3 etc). Now just waiting for it to make it's way into builds so I can give it a go (curious, but not optimistic, to see how my vega64 fares).
 
  • Like
Reactions: throAU
mi7chy said:
Going from 24 to 32-core iGPU M1 Max would be at ~36s if scaling was perfect.

Also, looking forward to AMD dGPU support that could carry over to hackintosh. People are reporting about double the performance or half the time with AMD HIP vs OpenCL.

https://code.blender.org/2021/11/next-level-support-for-amd-gpus/
Click to expand...

I was pretty surprised at how bad OpenCL did. It's good that AMD is putting in the effort to create a workable cross vendor platform (because sure as hell NVIDIA isn't going to bother). While it's assumably not going to have any impact on macOS, anything that makes CUDA less dominant is probably a good thing.

They appear to be claiming that "HIP is very thin and has little or no performance impact over coding directly in CUDA mode." Nothing like allowing programmers to code once for multiple platforms to get a technology adopted (see also electron, although hopefully this is a bit better :p).

One interesting thing on that page was the statistic that 80% of Blender users are on Windows (hence windows is getting the implementation first). Always naively assumed because of Blender's roots and that 90% of the large FX studios are on linux, that a significant amount of their users would be...
 
Last edited:
BootLoxes said:
Im actually quite surprised at the scores. I was expecting much better. Hopefully it gets better as the metal rendering is further developed
Click to expand...

What kind of scores were you expecting?

It seems like a pretty reasonable score to me; faster than a desktop gpu, without the dedicated raytracing hardware and in a shouting distance of the same gpu with the dedicated hardware.

To put that in perspective a GeForce 3090 render the same scene (using CUDA) in around 20s, or around 10s with Optix. As a side note a Ryzen 9 3950X takes over a minute, so gpu ftw.

Definitely feel that the BMW scene doesn't really play to the strengths of the Mac GPU; it's a relatively simple scene, so not going to really benefit from the increased unified memory available. But having those speeds on a near silent machine seems like a pretty big win.

As you say, hopefully it will get faster as the metal implementation improves. Still hoping that apples going to add some raytracing hardware in the not too distant. If the iMac Pro (as rumoured) is 2x M1 Max it should be pretty competitive (although the price would be pretty astronomical one assumes).
 
  • Like
Reactions: iPadified and Macintosh IIcx
jujoje said:
As you say, hopefully it will get faster as the metal implementation improves. Still hoping that apples going to add some raytracing hardware in the not too distant. If the iMac Pro (as rumoured) is 2x M1 Max it should be pretty competitive (although the price would be pretty astronomical one assumes).
Click to expand...

Then again a GeForce 3090 with 24GB seems to be around 2800usd so possibly not that expensive when you take that into account :p
 
jujoje said:
Nice! As a tl;dr from the blender thread, the times on the BMW scene looked like:

29s - 2070 Super (Optix)
48s - M1 Max 24GPU
51s - 2070 Super (CUDA)
2.04 min - Mac mini M1

Which seems pretty good (although also from the thread, the results may not be representative because alpha builds/comparing between 2.9x and 3 etc). Now just waiting for it to make it's way into builds so I can give it a go (curious, but not optimistic, to see how my vega64 fares).
Click to expand...
Great! What was the time for CPU on M1 Max?

Really surprising how competitive the M1 Max is in the GPU department.
 
B4DF82F4-8B32-4026-81C3-EABE0D839F13.jpeg
About Subtance Stager and GPu support.
 
BootLoxes said:
Im actually quite surprised at the scores. I was expecting much better. Hopefully it gets better as the metal rendering is further developed
Click to expand...

Why? I mean, it would be unrealistic to expect it to outperform a high-end desktop GPU with dedicated RT hardware. For a compact mobile workstation, this looks very good to me.
 
  • Like
Reactions: Xenobius and Macintosh IIcx
leman said:
Why? I mean, it would be unrealistic to expect it to outperform a high-end desktop GPU with dedicated RT hardware. For a compact mobile workstation, this looks very good to me.
Click to expand...
Like I said I was expecting about 25% better on the M1 (base not pro or max)
 
iPadified said:
Great! What was the time for CPU on M1 Max?
Click to expand...

According to that thread 3min 13s. So significantly slower.

BootLoxes said:
Like I said I was expecting about 25% better on the M1 (base not pro or max)
Click to expand...

The goal of the M1 mini GPU was to be significantly better than Intel Integrated GPUs, and by and large, it does a good job. It is bit disappointing, and as someone who owns a M1 MBA would love it to be that 25% faster. GPU rendering on the base M1 chips is more of a bonus than feature (still surprised at how well it runs 3D Apps).
 
  • Like
Reactions: iPadified
jujoje said:
According to that thread 3min 13s. So significantly slower.



The goal of the M1 mini GPU was to be significantly better than Intel Integrated GPUs, and by and large, it does a good job. It is bit disappointing, and as someone who owns a M1 MBA would love it to be that 25% faster. GPU rendering on the base M1 chips is more of a bonus than feature (still surprised at how well it runs 3D Apps).
Click to expand...
Honestly at the moment pretty much all my software is rosetta and without gpu support and the experience on m1 8gb hasnt been that bad. I just finished this model entirely on an m1 air and I can see the potential of apple silicon. If devs bring native apps with gpu support it will be a gamechanger
 

Attachments

  • D0C68038-1581-4311-AC67-4E7D68D6D425.png
    D0C68038-1581-4311-AC67-4E7D68D6D425.png
    706.8 KB · Views: 115
  • Like
Reactions: kamikazeeMC, jujoje and Boil
BootLoxes said:
Like I said I was expecting about 25% better on the M1 (base not pro or max)
Click to expand...
Which does not answer Leman's question. Why do you expect it to be 25% better? Why 25% in particular?

The BWM scene takes about 4:30-5 min to render on openCL on the M1 GPU. Now it takes about 2 min with the current Metal build. Not sure why you expect "25% better".
 
Last edited:
jujoje said:
According to that thread 3min 13s. So significantly slower.



The goal of the M1 mini GPU was to be significantly better than Intel Integrated GPUs, and by and large, it does a good job. It is bit disappointing, and as someone who owns a M1 MBA would love it to be that 25% faster. GPU rendering on the base M1 chips is more of a bonus than feature (still surprised at how well it runs 3D Apps).
Click to expand...
Yes same I did try it on the M1 Pro yesterday was about 3 min on Cpu.

Still waiting for that Metal build…
 
jeanlain said:
Which does not answer Leman's question. Why do you expect it to be 25% better? Why 25% in particular?

The BWM scene takes about 4:30-5 min to render on openCL on the M1 GPU. Now it takes about 2 min with the current Metal build. Not sure why you expect "25% better".
Click to expand...
Its an improvement yes but not as big as I was expecting. 25% wasnt a particular number but rather a range. I guess I just overestimated the m1 gpu potential.
 
BootLoxes said:
I guess I just overestimated the m1 gpu potential.
Click to expand...
The performance per watt of Nvidia and Apple GPUs is very similar for deep learning tasks.

So, if an Nvidia GPU consumes eight times what the Apple GPU consumes, that Nvidia GPU tends to finish eight times faster.
 
iPadified said:
Scaling to 16 core GPU would be 1 minute 15 sec. Still not bad compared to CPU.
Click to expand...
yeah would be happy enough with that for a portable machine and silent machine.

Even after the 3dmark stress test which is 20 loops the fans are only at 1700ish rpm and 99,5 % stability ;)
 
Last edited:
  • Like
Reactions: iPadified
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back