Apple Sillicon behind in GPU both RT and Raster and Pref/$.

[exoticspice1]

Apple is on TSMC 5nm and later this year Nvidia and AMD wil release RTX 4000 and AMD 7000 both will be on TSMC 5nm based nodes.

My question is, will Apple match a supposed RTX 4070 or RX 7700XT with the M2 Max in the upcoming Mac Studio?

Apple is already WAY behind RT and raster compared to AMD and Nvidia.

Also why is Apple using laptop SoCs and therefore laptop GPUs in the Mac Studio, the Mac Studio has insane cooling but its being limited.

I hear the RTX 4070 will match the RTX 3080 Ti. If priced at $600-$650 it will be insansely powerful and cheap to build a compable PC compared to an M2 Max if we follow the M2 scaling for 3D and content creation and as a bonus you can game too.

M2 10 core = 30,000 Metal
M2 Pro 20 core = 60,000 Metal
M2 Max 40 core = 115,000 Metal

Of couse this is assuming perfect scaling.
Thoughts?

With Metal 3 supported on 6000 series I want Apple to support 7000 series as well. Apple's GPUs are not as impressive as their CPUs.

 

[Cmaier]

Apple is on TSMC 5nm and later this year Nvidia and AMD wil release RTX 4000 and AMD 7000 both will be on TSMC 5nm based nodes.

My question is, will Apple match a supposed RTX 4070 or RX 7700XT with the M2 Max in the upcoming Mac Studio?

Apple is already WAY behind RT and raster compared to AMD and Nvidia.

Also why is Apple using laptop SoCs and therefore laptop GPUs in the Mac Studio, the Mac Studio has insane cooling but its being limited.

I hear the RTX 4070 will match the RTX 3080 Ti. If priced at $600-$650 it will be insansely powerful and cheap to build a compable PC compared to an M2 Max if we follow the M2 scaling for 3D and content creation and as a bonus you can game too.

M2 10 core = 30,000 Metal
M2 Pro 20 core = 60,000 Metal
M2 Max 40 core = 115,000 Metal

Of couse this is assuming perfect scaling.
Thoughts?

With Metal 3 supported on 6000 series I want Apple to support 7000 series as well. Apple's GPUs are not as impressive as their CPUs.

I don’t think it’s fair to call the Mac Studio CPUs “laptop cpus.” That’s the old way of thinking about it, brought on by terrible power consumption by CISC chips.

 

[exoticspice1]

I don’t think it’s fair to call the Mac Studio CPUs “laptop cpus.” That’s the old way of thinking about it, brought on by terrible power consumption by CISC chips.

mm. ok, can you tell me why the scaling is bad when moving to M1 Max and then to M1 Ultra?

Maybe Apple can fix it in M2 generation

 

[Cmaier]

mm. ok, can you tell me why the scaling is bad when moving to M1 Max and then to M1 Ultra?

Maybe Apple can fix it in M2 generation

What makes you say the scaling is bad?

 

[exoticspice1]

What makes you say the scaling is bad?

Does not scale well with the doubling each with SKU.

 

[Cmaier]

View attachment 16411

Does not scale well with the doubling each with SKU.

Ah, you mean the GPU. I don’t know, I am not a GPU designer. You referred to the CPU, so that’s what I was talking about.

 

[Yoused]

View attachment 16411

Does not scale well with the doubling each with SKU.

Man, that is cool. The Metal score converges. At some point it will reach saturation and adding more cores will not measurably improve performance.

 

[exoticspice1]

RDNA 2 scales way better.

 

[Cmaier]

View attachment 16412
RDNA 2 scales way better.

View attachment 16413

Different architecture entirely, isn’t it?

 

[exoticspice1]

Different architecture entirely, isn’t it?

Yes, AMD's architecture different buts its better and scaling is good. Also has better features and RT. I want Apple do better, otherwise the Mac will be behind in the GPU space. Apple already kicked out Nvidia and if any of the future Mac products will have no AMD GPUs, Apple needs to make sure they are at least on par with pref.

 

[SuperMatt]

mm. ok, can you tell me why the scaling is bad when moving to M1 Max and then to M1 Ultra?

Maybe Apple can fix it in M2 generation

Buy a PC.

 

[leman]

Apple is already WAY behind RT and raster compared to AMD and Nvidia.

Apple is behind RT because they don't have hardware RT. I am not sure what you mean with them being behind on "raster". If you are talking about traditional rasterisation, Apple is around 2-4x faster per watt than Nvidia GPUs, depending on configuration.

View attachment 16411

Does not scale well with the doubling each with SKU.

GB severely underestimates the performance of Apple GPUs. Apple needs at least 10ms of warmup period to enter the high performance mode on the GPU and GB GPU workloads are too short. Also, these results suggest that GB does not use enough threads to saturate the larger M1 GPUs. It all boils down to how they do work dispatch. Without seeing the code and profiling their pipeline it's impossible to say more.

mm. ok, can you tell me why the scaling is bad when moving to M1 Max and then to M1 Ultra?

Because you are using a flawed benchmark. Apple has nothing to fix here. Rather, it's Geekbench that has to be fixed.

M1 GPUs show linear scaling in synthetic throughput benchmarks (I got 2.6TFLOPS/s for 10W with M1 and 10TFLOPS/s for 40W with M1 Max, just like advertised) as well as gaming benchmarks (3dmark)

 

[jbailey]

View attachment 16411

Does not scale well with the doubling each with SKU.

You should never make comparisons based on a single benchmark. As @leman says above, GeekBench Metal test is known not to saturate the M1 and M2 GPUs.

 

[diamond.g]

Apple is on TSMC 5nm and later this year Nvidia and AMD wil release RTX 4000 and AMD 7000 both will be on TSMC 5nm based nodes.

My question is, will Apple match a supposed RTX 4070 or RX 7700XT with the M2 Max in the upcoming Mac Studio?

Apple is already WAY behind RT and raster compared to AMD and Nvidia.

Also why is Apple using laptop SoCs and therefore laptop GPUs in the Mac Studio, the Mac Studio has insane cooling but its being limited.

I hear the RTX 4070 will match the RTX 3080 Ti. If priced at $600-$650 it will be insansely powerful and cheap to build a compable PC compared to an M2 Max if we follow the M2 scaling for 3D and content creation and as a bonus you can game too.

M2 10 core = 30,000 Metal
M2 Pro 20 core = 60,000 Metal
M2 Max 40 core = 115,000 Metal

Of couse this is assuming perfect scaling.
Thoughts?

With Metal 3 supported on 6000 series I want Apple to support 7000 series as well. Apple's GPUs are not as impressive as their CPUs.

As of right now the 4070 is rumored to match a 3090ti in Time Spy Extreme score (supposedly using less power to boot). Apple has pitted the Ultra vs the 3090vanilla not the Max. I am not sure the M2 Max is going to get the kind of performance bump needed to rival the M1 Ultra.

 

[Joelist]

Yeah, Geekbench is known (even by GB themselves) to have an issue with its benches on M Series because they fail to utilize all the resources.

 

[Jimmyjames]

Everything @leman said above basically.

I'd add for raster performance you can look at benchmarks like Gfxbench for more positive scores. The Max and Ultra are very competitive with the best AMD and Nvidia have to offer on that benchmark.

Finally, I've already posted this in another thread, but
https://www.twitter.com/i/web/status/1538362629818703872/
If this guy thinks Apple are onto something, (and he was very high up in Nvidia before leaving for Apple) then he's probably correct.

 

[Renzatic]

I'm still waiting on Apple to add some RT cores to their GPU mix. That's really the only thing that's keeping me away from a Mac Studio at this point.

 

[diamond.g]

Everything @leman said above basically.

I'd add for raster performance you can look at benchmarks like Gfxbench for more positive scores. The Max and Ultra are very competitive with the best AMD and Nvidia have to offer on that benchmark.

Finally, I've already posted this in another thread, but
https://www.twitter.com/i/web/status/1538362629818703872/
If this guy thinks Apple are onto something, (and he was very high up in Nvidia before leaving for Apple) then he's probably correct.

He probably is right, sadly OEM/ODM do not appear to want 3090ti's attached to the same die as a 12900k (or in the case of AMD a 5950 + 6950XT). That is where Apple gets to blaze a trail because they own the whole stack (ignoring Fab stuff) and are in control of their own destiny.

 

[exoticspice1]

Apple is around 2-4x faster per watt than Nvidia GPUs, depending on configuration.

This is because Apple uses a better node. TSMC 5nm is very good and RTX 3000 uses Samsung 8nm. We should wait and see how the RTX 4000 and RX 7000 do as they will be based on 5nm from TSMC.


Everything else you said I agree. But 3D and GFX has been Apple's area. They JUST made Metal better and added APIs that have been in PCs and Consoles for years. RDNA 3 and Lovelace will once again bring forth new advancements. It's not good to put all your eggs in one basket.

 

[Cmaier]

This is because Apple uses a better node. TSMC 5nm is very good and RTX 3000 uses Samsung 8nm.

No. Going from 8nm to 5nm doesn’t explain 2-4x faster per watt. Going from 8nm all the way down to 4nm wouldn’t explain 2x faster per watt, let alone 4x.