3D Rendering on Apple Silicon, CPU&GPU

[leman]

I think we are done with these types of comparative discussions many, many pages ago. Nothing new here...

It is still interesting to look at different benchmarks, as they can show us relative strengths and weaknesses of various implementations. AMD seems to perform better on CB24 GPU relative to others than they do on Blender for example. Apple seems to perform better on Blender (while obviously still trailing far behind Nvidia).

I would speculate that metal code in Redshift is not yet as much optimised as it could be. I hope Mason is using Apple's RT API so that next-gen hardware can benefit from hardware RT (if it ships with it, hopefully).

FWIW: M2 Ultra 24 CPU, 60 GPU

- cpu multicore: 1927
- gpu: 7441

Thanks, missed this!

 

[leman]

From the Redshift Forums . . . assume this is the 76 core GPU.

View attachment 2256404

Yeah, this shows that Maxon's Metal implementation lacks crucial optimizations. M2 Ultra is nominally 50-60% faster than the W6800 Pro — and almost 2x faster in Blender benchmarks.

 

[iPadified]

It is still interesting to look at different benchmarks, as they can show us relative strengths and weaknesses of various implementations. AMD seems to perform better on CB24 GPU relative to others than they do on Blender for example. Apple seems to perform better on Blender (while obviously still trailing far behind Nvidia).

I agree regarding relative gains of software (and hardware) on the Mac with 4090 as a reference. However, we have discussed price competitiveness and RAM size availability many many times.

 

[avkills]

Yeah, this shows that Maxon's Metal implementation lacks crucial optimizations. M2 Ultra is nominally 50-60% faster than the W6800 Pro — and almost 2x faster in Blender benchmarks.

Please post a link supporting this claim. Every benchmark I have seen shows the Ultra falling short of the W6800.

 

[diamond.g]

Please post a link supporting this claim. Every benchmark I have seen shows the Ultra falling short of the W6800.

I'd think those benchmarks are from when the framebuffer on the W6800 isn't big enough.

 

[leman]

Please post a link supporting this claim. Every benchmark I have seen shows the Ultra falling short of the W6800.

You can find official Blender benchmark results here:

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 results appear to scale very well with the compute capability, e.g. 20TFLOPS Nvidia and AMD GPUs have comparable scores (not taking Nvidia's hardware RT into account, just looking at compute). Since m2 Ultra nominally has ~50% more compute throughput than W6800, I expect it to be faster in compute workload.

 

[avkills]

I'd think those benchmarks are from when the framebuffer on the W6800 isn't big enough.

This I would agree with.

 

[avkills]

You can find official Blender benchmark results here:

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 results appear to scale very well with the compute capability, e.g. 20TFLOPS Nvidia and AMD GPUs have comparable scores (not taking Nvidia's hardware RT into account, just looking at compute). Since m2 Ultra nominally has ~50% more compute throughput than W6800, I expect it to be faster in compute workload.

I have not seen that bear out in real world use comparisons. But perhaps Blender is an outlier since they seem to be getting a lot of help from Apple; and other stuff just isn't optimized enough to take advantage of how the Apple Silicon chip wants stuff done.

I'll probably have Apple silicon here soon enough; kind of depends on how much of a step up the M3 is going to be.

 

[deconstruct60]

You can find official Blender benchmark results here:

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

Not sure why Cuda is thrown in there. Dump the extra stuff and drop down the HIP and Metal in the chart.

Yeah the W6800X is 1811 but the 6800XT is 2408 ( 2156 Metal) , 6950 2741 with the Ultra 76 core is 3447 .
So the leap from W6800 is about 2x , but that isn't the particularly optimized starting point from which card could be thrown into a MP 2019 with that die. The W6800X is being power managed. Not really getting max performance out of that die with Metal either.

Metal has isn't really a good cross architecture compare anymore. It is on the fast track to being a one shot arch wonder.

 

[leman]

Not sure why Cuda is thrown in there. Dump the extra stuff and drop down the HIP and Metal in the chart.

Yeah the W6800X is 1811 but the 6800XT is 2408 ( 2156 Metal) , 6950 2741 with the Ultra 76 core is 3447 .
So the leap from W6800 is about 2x , but that isn't the particularly optimized starting point from which card could be thrown into a MP 2019 with that die. The W6800X is being power managed. Not really getting max performance out of that die with Metal either.

Metal has isn't really a good cross architecture compare anymore. It is on the fast track to being a one shot arch wonder.

I am not comparing different Metal implementations, I am comparing the GPU scores with different backends. As you say, Metal scores on AMD are irrelevant. HIP scores are interesting. CUDA is also interesting because it's a mature backend that shows how Nvidia GPUs perform with their raytracing functionality disabled (so that we can compare the compute capability of various GPUs).

If you look at compute-only scores (CUDA for Nvidia, Metal for Apple, HIP for AMD), you will observe that they are well predicted by the nominal GPU compute throughput. A 20TFLOPS GPU from different vendors will have similar scores for example. And this makes a lot of sense. The way how different mainsteam GPUs execute code is very similar these days, micro-architectural differences only creating very small variation (e.g. Apple M2 has slightly more efficient compute utilisation, Nvidia has slightly less, but these are all known and explained facts).

 

[treehuggerpro]

Yeah, this shows that Maxon's Metal implementation lacks crucial optimizations.

Please post a link supporting this claim.

On Maxon's side, they've indicated they're committed to optimising for Apple's GPU, while making clear it's not a small undertaking. They've also remained tightlipped about it, citing NDAs.

https://forums.macrumors.com/threads/3d-rendering-on-apple-silicon-cpu-gpu.2269416/post-32148155

 

[terminator-jq]

Based off the A17, it looks like we are getting hardware raytracing in the M3 chips!! Glad Apple made the jump! The M3 Max will be very interesting!

 

[Boil]

The M3 Ultra will be very very interesting...! ;^p

 

[phobos]

Based off the A17, it looks like we are getting hardware raytracing in the M3 chips!! Glad Apple made the jump! The M3 Max will be very interesting!

Yep that was definitely the most exciting thing out of the whole presentation! Apple might have a chance of competing with Nvidia now. Let's see!

 

[sirio76]

A low power chip may compete on performance per watt but not on absolute performance, since Nvidia GPU will reach soon 600W there’s no chance to compete.
Regardless I’ll take a silent reliable machine any day of the week, even if it means to give up on some performance.

 

[leman]

The M3 Ultra will be very very interesting...! ;^p

If the claims of 4x RT performance translate to production RT, M3 Ultra could be as fast as a 4090 RTX in Blender. Of course, I am very skeptical. But 2-2.5x improvements are probably not unrealistic. That would already be phenomenal and make Mac laptops best choice for CAD and 3D (at least platform-wise).

I wonder what other GPU improvements they made. They mentioned a full redesign. Could be something to do with closing the feature parity gap to Nvidia chips?

 

[iPadified]

4x A16 ray tracing speed. I wonder what that translate to in real life using optimised software?

 

[leman]

4x A16 ray tracing speed. I wonder what that translate to in real life using optimised software?

If they are talking about RT heavy scenes, then, 4x If they are talking about limited use of RT (e.g. for a game with a few ray-traced lights or shadows), probably under 2x. We will see.

 

[leman]

A low power chip may compete on performance per watt but not on absolute performance, since Nvidia GPU will reach soon 600W there’s no chance to compete.
Regardless I’ll take a silent reliable machine any day of the week, even if it means to give up on some performance.

Yes, but there is a but: Apple‘s RT approach is inherently much more power-efficient than Nvidia‘s. At least based on the patents I’ve read. It could be possible for Apple to beat a much larger and power-hungry GPU on the RT metric. In pure shader power, not really (here die area and power consumption constraints reign supreme). But with smarter tech Apple could punch way above its weight in certain applications. Like we already see M2 performing surprisingly well for rasterization thanks to TBDR.

 

[jeanlain]

If the claims of 4x RT performance translate to production RT, M3 Ultra could be as fast as a 4090 RTX in Blender.

When they say it's 4X faster, are they comparing A17 to A16, or to A17 without the use of the dedicated RT hardware?

 

[diamond.g]

When they say it's 4X faster, are they comparing A17 to A16, or to A17 without the use of the dedicated RT hardware?

Probably

 

[leman]

When they say it's 4X faster, are they comparing A17 to A16, or to A17 without the use of the dedicated RT hardware?

Who knows… but I doubt it matters much. It’s the ballpark that’s interesting.

Its compared to A16, as @NT1440 mentions

 

[NT1440]

When they say it's 4X faster, are they comparing A17 to A16, or to A17 without the use of the dedicated RT hardware?

4x faster than software based RT on the A16, as stated explicitly in the presentation (I made a note of it).

 

[Xiao_Xi]

When they say it's 4X faster, are they comparing A17 to A16, or to A17 without the use of the dedicated RT hardware?

Unfortunately, there is no footnote on how Apple arrived at the number.

iPhone

Designed for Apple Intelligence. Discover the new iPhone 16 Pro, iPhone 16 Pro Max, iPhone 16, and iPhone 16 Plus.

www.apple.com

 

[richinaus]

The M3 Ultra will be very very interesting...! ;^p

Did you ignore all the rest of the news like me and just focus on A17 GPU raytracing when you saw it