3D Rendering on Apple Silicon, CPU&GPU

[treehuggerpro]

@sirio76, it's a complex enough scene, over 10 minutes on a single W5700X. That's reasonable and the best guide we've got. The Mac example above is probably from two 16 core MPs and likely as good as you're going to get, like for like, in the wild. Still, if you look at the two RTX 2070 Super 8GB x 2 in the list, you've got two OS's, two different CPUs, running at 3.5 and 3.0 GHz respectively, with 1 second difference in the result. It's a GPU renderer, after all.

 

[ader42]

I guess we’ll have to wait for actual render times to come out to see if the Ultra GPUs actually do anything of note - it might be that the main reason to buy an Ultra is for the extra CPU cores.

 

[Xiao_Xi]

Apple still has a lot of work to do on Blender. Currently, M1 Max (704) performs like an Nvidia GTX 1660 SUPER (766).

M1 Max results:

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

Nvidia GTX 1660 SUPER results:

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

 

[treehuggerpro]

I guess we’ll have to wait for actual render times to come out to see if the Ultra GPUs actually do anything of note - it might be that the main reason to buy an Ultra is for the extra CPU cores.

Yes, I think you're right. I also assume the 48 core Ultra won't trail too far behind a fully specced Ultra in Redshift, or similar tasks, due to where Apple has likely targeted their noise / heat management to be in the Studio.

 

[ader42]

I’m also wondering how the Studio will compare to the MBP with the same SOC - i.e. MBP 16” M1 Max 64GB RAM 32 GPU Cores versus Studio M1 Max 64GB RAM 32 GPU Cores…

I’m sure the MBP only draws about 45 watts but the studio looks like it draws more power as the slides showed 60 watts…

 

[treehuggerpro]

Yep, I'm sitting on my hands until next week, to be sure. I expect Apple's charts are correct though, based on how the Max's GPU was managed in the MBPs. CPU should be consistent, but I think it's a little clearer now, Apple tunes the Max's GPU to suit the thermal and/or noise constraints of their enclosures. The 32x Max in the Studio looks like it will have its GPU fully utilised (first example), whereas the 64x Ultra will likely be somewatt managed.

From their marketing, I'm guessing sustained rendering is where the 64x Ultra gets constrained most, in other tasks it looks to scale well. No matter, the Studio is good value, the 48x Ultra may just work out the better choice where GPU rendering is concerned. One caveat, Twinmotion is supposed to gain Path Tracer support for the M1 Macs soon, it may prove a better use case for the 64x Ultra?

 

[Boil]

Yep, I'm sitting on my hands until next week, to be sure. I expect Apple's charts are correct though, based on how the Max's GPU was managed in the MBPs. CPU should be consistent, but I think it's a little clearer now, Apple tunes the Max's GPU to suit the thermal and/or noise constraints of their enclosures. The 32x Max in the Studio looks like it will have its GPU fully utilised (first example), whereas the 64x Ultra will likely be somewatt managed.

Someone will swap out their aluminum M1 Max heat sink for a copper M1 Ultra heat sink at some point...?

From their marketing, I'm guessing sustained rendering is where the 64x Ultra gets constrained most, in other tasks it looks to scale well. No matter, the Studio is good value, the 48x Ultra may just work out the better choice where GPU rendering is concerned. One caveat, Twinmotion is supposed to gain Path Tracer support for the M1 Macs soon, it may prove a better use case for the 64x Ultra?

The M1 Ultra sounds awesome, but that price tag...!

I was one of those expounding for a taller Mac mini (actually for splitting the Mac mini line into two different chassis designs, a new shorter for the Mn SoC, and a new taller for the M1 Pro/Max SoCs) after no new Mac mini appeared at the Fall 2021 MBP event...

All the h8rs: "Apple will never do that."

And now we have the Mac Studio, which is basically a taller Mac mini...! ;^p

Point is, seems the Mac Studio is the perfect chassis to let the full-die M1 Max SoC shine...?

Now if Apple could just do something about these danged shipping dates...!

 

[treehuggerpro]

Point is, seems the Mac Studio is the perfect chassis to let the full-die M1 Max SoC shine...?

Yeah, I'm assuming the Mac Studio gives you the maximum M1 Max (as is), and that the 64x Ultra is a little cramped for some loads.

The M1 Ultra sounds awesome, but that price tag...!

By my estimate, you're still paying less than half of what you would to acquire a Mac Pro with similar performance.

Now if Apple could just do something about these danged shipping dates...!

Yep, that hurts!

 

[sunny5]

Blender 3.1 released but still need to wait till 12.3 release.

 

[Boil]

Excited to see Blender on Mac Studio...!

 

[sunny5]

Apple still has a lot of work to do on Blender. Currently, M1 Max (704) performs like an Nvidia GTX 1660 SUPER (766).

M1 Max results:

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

Nvidia GTX 1660 SUPER results:

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

Blender is not yet ready to support Apple Silicon's GPU. You need Blender 3.1 AND macOS 12.3.

 

[mi7chy]

Blender is not yet ready to support Apple Silicon's GPU. You need Blender 3.1 AND macOS 12.3.

AS is supported with Blender 3.1 and MacOS 12.2. MacOS 12.3 is required for Metal support on AMD GPU.

 

[sunny5]

AS is supported with Blender 3.1 and MacOS 12.2. MacOS 12.3 is required for Metal support on AMD GPU.

Blender 3.1 update adds Metal GPU rendering for M1 Macs - 9to5Mac

Blender 3.1 now has Metal GPU rendering for Macs with the M1 chip. The app renders up to 2x faster compared to standard CPU rendering.

9to5mac.com

Blender 3.1 released from yesterday and macOS 12.3 is also required for Apple Silicon chip. You better check the info again.

 

[mi7chy]

Blender 3.1 update adds Metal GPU rendering for M1 Macs - 9to5Mac

Blender 3.1 now has Metal GPU rendering for Macs with the M1 chip. The app renders up to 2x faster compared to standard CPU rendering.

9to5mac.com

Blender 3.1 released from yesterday and macOS 12.3 is also required for Apple Silicon chip. You better check the info again.

Time to complete the hackintosh. Curious to see what's faster on Blender, HIP or Metal, with AMD GPU.

https://www.blender.org/download/releases/3-1/

 

[sunny5]

https://www.blender.org/download/releases/3-1/
View attachment 1971441

I'm on the latest version but it performs worse than GTX 1660TI. I dont think it's ready yet unless you get 12.3 which isn't release yet.

Also does it even say M1 Max or Pro? It clearly mentioned M1 MBA, not M1 Pro and Max so I'm still right.

 

[mi7chy]

View attachment 1971443
I'm on the latest version but it performs worse than GTX 1660TI. I dont think it's ready yet unless you get 12.3 which isn't release yet.

Also does it even say M1 Max or Pro? It clearly mentioned M1 MBA, not M1 Pro and Max so I'm still right.

https://devtalk.blender.org/t/cycles-apple-metal-device-feedback/21868

 

[l0stl0rd]

Blender is not yet ready to support Apple Silicon's GPU. You need Blender 3.1 AND macOS 12.3.

Ooh but it is and you don't need 12.3 you only need 12.3 for AMD cards.

I have been using GPU rendering on the M1 for months now.

M1 Pro double binned and my and old
5700 XT and the M1 Max for the screenshots

And some more tests I did… same scenes that the benchmark uses.
Those are with the 32 core Max.
Ok here we go, all done using Blender 3.1 final on both Systems

Monster under the bed.
M1 Max 173 sec
5700 XT 156 sec
5700 XT = 10% than M1 Max

Classroom
M1 Max 108 sec
5700 XT 91 sec
5700 XT = 18% faster than M1 Max

Junkshop
M1 Max 34 sec
5700 XT 36 sec
5700 XT = nearly 6 % slower than M1 Max

Also if you want a prediction I am saying slightly over 1000 point in the open data bench for the 48 core.

An even if 3.1 is not fully optimized for metal I would not expect a gain of more than 10-20 %.

 

[sunny5]

https://devtalk.blender.org/t/cycles-apple-metal-device-feedback/21868

That's beta dude.

 

[sunny5]

Ooh but it is and you don't need 12.3 you only need 12.3 for AMD cards.

I have been using GPU rendering on the M1 for months now.

M1 Pro double binned and my and old
5700 XT and the M1 Max for the screenshots

And some more tests I did… same scenes that the benchmark uses.
Those are with the 32 core Max.
Ok here we go, all done using Blender 3.1 final on both Systems

Monster under the bed.
M1 Max 173 sec
5700 XT 156 sec
5700 XT = 10% than M1 Max

Classroom
M1 Max 108 sec
5700 XT 91 sec
5700 XT = 18% faster than M1 Max

Junkshop
M1 Max 34 sec
5700 XT 36 sec
5700 XT = nearly 6 % slower than M1 Max

Also if you want a prediction I am saying slightly over 1000 point in the open data bench for the 48 core.

An even if 3.1 is not fully optimized for metal I would not expect a gain of more than 10-20 %.

Apple compared M1 Max to laptop RTX 3080 in terms of performance so that's really low.

 

[l0stl0rd]

Apple compared M1 Max to laptop RTX 3080 in terms of performance so that's really low.

Ooh I know also rasterization is not the same thing as compute.

Also they compared it to a 3080 mobile which was tdp limited and they probably used some game or other benchmark.

I have seen it beat some 3080 mobile cards in some games.

Same reason I don’t believe the Ultra is faster than a 3090 in rendering.

I games who knows.

I am not going to blame it on software because in Redshift the M1 Max is disappointing too.
I think it was about 1080ti level speed in there.

 

[NT1440]

Apple compared M1 Max to laptop RTX 3080 in terms of performance so that's really low.

Did they? I thought it was YouTubers saying it was the laptop 3080 while Apple compared it to the actual 3080?

I could be misremembering though. Any clarity here?

 

[sunny5]

Did they? I thought it was YouTubers saying it was the laptop 3080 while Apple compared it to the actual 3080?

I could be misremembering though. Any clarity here?

I'm talking about a laptop version of 3080. They clearly mentioned a high-end PC laptop GPU which is RTX 3080.

 

[NT1440]

I'm talking about a laptop version of 3080.

I understand that, what I’m asking is whether Apple’s own comparisons were to the desktop or the laptop version.

 

[sunny5]

I understand that, what I’m asking is whether Apple’s own comparisons were to the desktop or the laptop version.

HIGH-END PC LAPTOP GRAPHICS. I've been saying the same thing. With Blender 3.1, M1 Max is not even close to a laptop RTX 3080 which is a shame. It only performs just like GTX 1660.

 

[sunny5]

Blender community on Reddit believe that Apple's marketing is a lie and M1 will never compete with dedicated GPUs as M1 Max proves to be worse than 1660.

If Apple cant improve this, then their high end market will be limited only for video and photo.