3D Rendering on Apple Silicon, CPU&GPU

[singhs.apps]

According to Anandtech’s article, the M1 Max CPU clusters tops out at 250GB/s. Although super impressive, it shows that the internal fabric may be bandwidth limited. It could be by design as GPUs also need bandwidth.

A single M1 core on the other hand could saturate the entire RAM bandwidth, although only at 60GB/s. So I think Apple is balancing the cost of their design. It’ll probably be too expensive to go high core counts. I do not know what’s the magic number tho. ?

I have a feeling Apple will increase bandwidth above 400 GBps for the 128 core GPU…that may have a nice side effect of higher bandwidth for the CPU too.

Also the maxed out MBP 2021 hovers around the 6k USD mark. An equivalent Mac Pro might be somewhat more expensive, even if the MBP has features (screen, battery, keyboard etc) that the Mac Pro wouldn’t (these costs will be external to the unit)
So a straight up 4x m1 max might be around the 25-30k mark, total guesswork of course, but far less than the maxed out 2019 Mac Pro (though it’s RAM capacity, MPX modules and the afterburner card [redundant on AS] would be still unaccounted for )

Let’s see.

 

[jeanlain]

What about ray tracing hardware? It seems that the M1 Pro/Max don't have anything like that.
And if we trust rumours, Mac Pros will use the same GPU cores, just more of them.
If M2 variants just use the A15 GPU cores with minor tweaks, then we won't have dedicated ray tracing accelerators for a loooong time.

 

[singhs.apps]

What about ray tracing hardware? It seems that the M1 Pro/Max don't have anything like that.
And if we trust rumours, Mac Pros will use the same GPU cores, just more of them.
If M2 variants just use the A15 GPU cores with minor tweaks, then we won't have dedicated ray tracing accelerators for a loooong time.

I am hoping Apple gets some dedicated RTX card module that’s built just for offline (non gaming)raytracing. It will smoke anything out there…but very unlikely.
If anything it will be Standard compute units + RTX units on a dedicated GPU..that’s if Apple sees sense in creating MPX modules for an AS Mac Pro

 

[quarkysg]

What about ray tracing hardware? It seems that the M1 Pro/Max don't have anything like that.
And if we trust rumours, Mac Pros will use the same GPU cores, just more of them.
If M2 variants just use the A15 GPU cores with minor tweaks, then we won't have dedicated ray tracing accelerators for a loooong time.

Maybe Apple is cooking up some drivers in their labs to pair the GPU with the NPU for ray tracing? The NPU is a matrix calculator so it’s possible?

 

[jeanlain]

Maybe Apple is cooking up some drivers in their labs to pair the GPU with the NPU for ray tracing? The NPU is a matrix calculator so it’s possible?

They may well do that already. Who knows what's behind Metal APIs like the "intersector" (which Apple initially described as a compute shader). But whatever's in the M1 variants, there's no module dedicated to finding ray intersections.

 

[hefeglass]

I did some more testing today, I am able to get much better viewport playback on both cycles and Eevee with the M1 Pro vs zephyrus g15. Is there something set incorrectly on the zephyrus? I have gpu compute set for cycles and it seems to be using the GPU , but the performance is poor. I am using CPU compute on the M1. Eevee seems to utilize GPUs in both systems, the M1 does a bit better here as well…but this is a simple test and I need to get a more complicated scene to really test viewport performance.

 

[Macintosh IIcx]

I think there a very high chance that Apple will add hardware ray tracing eventually. They have been working closely with Redshift and Otoy Octane to get Metal API and Metal drivers optimized and they wouldn't do that just for fun. But probably not something we will see before 2023 at the earliest.

 

[crazy dave]

They may well do that already. Who knows what's behind Metal APIs like the "intersector" (which Apple initially described as a compute shader). But whatever's in the M1 variants, there's no module dedicated to finding ray intersections.

Yeah you generally need hardware for both intersections (RTX) and denoising (tensor cores because the number of GPU rays is too small). Given the unified setup maybe the NPU can be used for the latter when doing graphics … unclear until Apple does it!

 

[crazy dave]

I think there a very high chance that Apple will add hardware ray tracing eventually. They have been working closely with Redshift and Otoy Octane to get Metal API and Metal drivers optimized and they wouldn't do that just for fun. But probably not something we will see before 2023 at the earliest.

True and I think they will too but keep in mind there is still CPU ray tracing which a number of renderers actually still use

 

[Boil]

I've been wondering this myself. How well does the M1 architecture suit itself to being "doubled" or "quadrupled" in the sense that you could stick two xeons in my Mac Pro or just 1 depending on your budget.

Wouldn't it depend on how much latency is introduced moving the ram off SOC?

The SoC is the CPU / GPU / NPU / etc. die itself; the RAM is immediately adjacent to the SoC, so it is "off SoC" already.

Does the current Mac Pro support 2 CPUs? According to WikiPedia, only the 1st gen Mac Pro sports a dual CPU design, as well as the Power Mac G5.

I don't think current macOS support NUMA tho.

Apple most likely will stay with a single SoC design and stitch multiple M1 Max together via a high speed interconnect fabric to take care of cache coherencies between all the M1 Maxs chiplets. Probably faster than two M1 Max as separate SoCs tied together externally and requiring macOS to support NUMA.

Obviously the 2019 Mac Pro does not (physically) support dual CPUs, there is not a secret hidden second socket somewhere. The earlier usage of dual CPUs in the PowerMac G5 & the 1st Gen Mac Pro, IMHO, was more because of low core counts at the time; easier to double up on the CPUs than to get Motorola or Intel to magically make high core count CPUs?

Wondering aloud:

Considering that macOS currently doesn’t support more than 64 threads and AS doesn’t have hyper threading (essentially one thread per core ), it still leaves Apple enough headroom to max out its Mac Pro AS system (instead of the current - and most likely - 40 core speculation) ?

Potential combinations for a M1 Mac Pro max :

14 firestorm + 2 ice storms x 4 SOCs = 64 cores (56 firestorm + 8 icestorm )

12 firestorm + 4 ice storms x 4 SOCs = 64 cores (48 firestorm + 16 icestorm ) = 16 LP cores are a waste in the Mac Pro IMO.

12 firestorm + 4 ice storms x 2 SOCs + 16 firestorm x 2 SOCs = 64 cores ( 56 firestorm + 8 icestorm )

16 firestorm x 4 SOCs = 64 cores ( full fat 64 firestorm ) = who needs LP cores on wall powered workstations



Of course I am pulling this out of thin air and have no clue about chip designs …but hey… why not ?

All rumors point towards dual & quad M1 Max SoC systems, how Apple ties these together we do not really know right now. But I really doubt Apple is going to make custom SoCs for a lineup (Mac Pro / Mac Pro Cube / <possible> 32" iMac Pro) that has low sales volume?

According to Anandtech’s article, the M1 Max CPU clusters tops out at 250GB/s. Although super impressive, it shows that the internal fabric may be bandwidth limited. It could be by design as GPUs also need bandwidth.

A single M1 core on the other hand could saturate the entire RAM bandwidth, although only at 60GB/s. So I think Apple is balancing the cost of their design. It’ll probably be too expensive to go high core counts. I do not know what’s the magic number tho. ?

Some feel the memory bandwidth will somehow magically double with a dual SoC model, and quadruple with a quad SoC model; but that doesn't seem like the way it would work? I don't have knowledge of how these things work, so I remain skeptical of 800GB/s & 1.6TB/s memory bandwidths in the M1 Max Duo & M1 Max Quadro systems.

But hey, if we do end up with systems that have those astounding bandwidth numbers, all for the better?

Also the maxed out MBP 2021 hovers around the 6k USD mark. An equivalent Mac Pro might be somewhat more expensive, even if the MBP has features (screen, battery, keyboard etc) that the Mac Pro wouldn’t (these costs will be external to the unit)
So a straight up 4x m1 max might be around the 25-30k mark, total guesswork of course, but far less than the maxed out 2019 Mac Pro (though it’s RAM capacity, MPX modules and the afterburner card [redundant on AS] would be still unaccounted for )

I would think nearly half of that US$6k figure would be the WAY overpriced 8TB SSD and the "laptop stuff" (chassis / display / keyboard / trackpad / batteries / etc.).

So a M1 Max Duo Mac Pro might start at the same price as the 2019 Mac Pro, about six grand; and a M1 Max Quadro Mac Pro might be about ten grand or so?

 

[mi7chy]

Have seen Blender benchmarks for the new Macbook Pro M1 Max 32-core iGPU 10-core CPU pop up and wanted to see how the Macbook Air M1 compares. Surprisingly, only about half slower. For dGPU comparison in Blender, Nvidia OptiX > Nvidia CUDA > OpenCL in performance.

Blender for Apple Silicon
https://www.blender.org/download/Blender2.93/blender-2.93.5-macos-arm64.dmg/

Windows x86-64
https://www.blender.org/download/Blender2.93/blender-2.93.5-windows-x64.msi/

Splash Fox demo render
https://cloud.blender.org/p/gallery/5f4cd4441b96699a87f70fcb

For demo renders it's usually best to avoid anything short. Classroom is the standard for benchmarking or even Splash Fox which is used in the YouTube video below for benchmarking Macbook Pro 16" M1 Max 32-core iGPU.

After installing Blender, load Splash Fox demo render, make sure to change render engine to Cycles from Eevee, select GPU (preferably) or CPU then render image.

Summary rendering times in min:sec.subsec (lower is better)

Macbook Air M1 8-core CPU 16:03.90
Macbook Air M1 7-core iGPU 8:30.30
AMD 5800H 8-core 3.2GHz base-clock no-boost CPU 5:13.47
Macbook Pro 16" M1 Max 32-core iGPU 10-core CPU (about same for either CPU or iGPU from YouTube video) ~4:
RTX3060 70W using OptiX 1:19.65

Macbook Air M1 8-core CPU 16:03.90

Macbook Air M1 7-core iGPU 8:30.30

AMD 5800H 8-core CPU base-clock no-boost 5:13.47

RTX3060 70W 1:19.65

 

[vel0city]

Looking good.

 

[jeanlain]

The GPU test is impressive. How can the M1 Max be >5 times faster than the radeon Pro 5600M?
I'm not a 3D artist, and I don't know which renderer is used here. I thought blender did not support Metal?

 

[jeanlain]

Looking good.

The mouse lag is interesting. The difference with the PC is night and day and I suppose this is where UMA really helps.
We tend to only rely on benchmarks, but this thing right there may be more important to may users than rendering time. Yet, it is not quantified.

 

[vel0city]

We tend to only rely on benchmarks, but this thing right there may be more important to may users than rendering time. Yet, it is not quantified.

Absolutely this 100%. Working on a 3D scene and dealing with mouse lag, viewport lag, "heavy" feel, really spoils the workflow and how involved you feel in your work. 3D is about constant optimisation and compromise which impacts on your creativity. Reducing those bottlenecks is way, way more important than the time it takes to render a final frame.

For most artists it's all about lookdev and getting the right look. Final renders aren't that important, you can just farm or overnight your finished frames. I want to know how quickly dynamics simulations solve, whether heavy displacement textures are responsive, how many millions of particles can I sim before the viewport freezes, sculpting on high density mesh responsiveness. This is all way more mission critical than render times.

 

[TrueBlou]

I think there a very high chance that Apple will add hardware ray tracing eventually. They have been working closely with Redshift and Otoy Octane to get Metal API and Metal drivers optimized and they wouldn't do that just for fun. But probably not something we will see before 2023 at the earliest.

I have little doubt that they will, at some point expand their GPU hardware to include hardware ray tracing.

But we must remember that we are still on Gen1 of Apple Silicon and likely will be for at least another year or so. Though it’s probably the most impressive Gen1 Apple hardware I’ve used in a very, very long time.

Perhaps in the second generation pro desktop class Apple Silicon we’ll see it. But at this point, it’s all wishful thinking and speculation.

They know they’re going to need it, so it will come, just when is the big question.

 

[diamond.g]

They may well do that already. Who knows what's behind Metal APIs like the "intersector" (which Apple initially described as a compute shader). But whatever's in the M1 variants, there's no module dedicated to finding ray intersections.

Which (as we have seen with AMD) is a very costly thing with tracing rays realtime.

 

[diamond.g]

I have little doubt that they will, at some point expand their GPU hardware to include hardware ray tracing.

But we must remember that we are still on Gen1 of Apple Silicon and likely will be for at least another year or so. Though it’s probably the most impressive Gen1 Apple hardware I’ve used in a very, very long time.

Perhaps in the second generation pro desktop class Apple Silicon we’ll see it. But at this point, it’s all wishful thinking and speculation.

They know they’re going to need it, so it will come, just when is the big question.

Do we know if the GPU in the A15 is the same as the GPU in M1? @leman?

 

[jeanlain]

load Splash Fox demo render

Does it render with the Apple GPU? Commenters on the video you linked to say Apple GPUs are not supported yet.

 

[jeanlain]

Do we know if the GPU in the A15 is the same as the GPU in M1?

It's not the same. There are new features like lossy texture compression. Oddly, these features may be missing from the latest M1 variants.

 

[diamond.g]

It's not the same. There are new features like lossy texture compression. Oddly, these features may be missing from the latest M1 variants.

So if they use the same GPU blocks in M2/M2P/M2M then there would be no hardware RT acceleration there. Unless they diverge the GPU blocks (again) like they did for M1 and A14. Hardware RT (aactual RT vs RQ that Metal currently uses) and Mesh Shaders are like the last bits that D3D12U (and Vulkan) have over Metal

It appears we have to wait for the GravityMark devs to see how much M1M improves the rendering time for RT scenes (I am not aware of any other apps/devs that have done a comparison of the various RT API's).

 

[jeanlain]

Thinking about it, the GPU cores used in the M1 Pro/Max may not tell us much about ray tracing in future Apple SoCs. Apple can add a ray intersector module in there while keeping the same GPU cores. They just added ProRes codecs, and they didn't have to change anything in the GPU for that.

 

[leman]

Do we know if the GPU in the A15 is the same as the GPU in M1? @leman?

It’s different.

 

[diamond.g]

Thinking about it, the GPU cores used in the M1 Pro/Max may not tell us much about ray tracing in future Apple SoCs. Apple can add a ray intersector module in there while keeping the same GPU cores. They just added ProRes codecs, and they didn't have to change anything in the GPU for that.

Yeah I am hoping they go the nvidia/intel route with the RT hardware. The way AMD is doing it probably would be more flexible in the end, but current workloads point to the nvidia/intel way being much faster.

 

[TrueBlou]

Do we know if the GPU in the A15 is the same as the GPU in M1? @leman?

To make what was becoming a long boring post shorter, basically there are differences between A-Series and Apple Silicon.

But as far as hardware ray-tracing is concerned on this architecture, they can add it whenever they like. And most likely will add it as soon as they deem the demand for it is high enough.
That may be in a second (or later) generation design, or it could make an appearance in a further optimised M1 variant next year when the Mac Pro is launched.

It's all speculative for now.