MacBook Pro's M1 Max GPU is Over 3x Faster Than M1 in First Metal Benchmark

[newbmacuser2021]

I can pull 2.2K gaming using Metal / SceneKit on an iPad Pro 10.5” (2017)

2224-by-1668

vs

1920-by-1080

So… this benchmark makes you wonder.

 

[McG2k1]

Any word on whether or not it can run Crysis?

 

[the-msa]

Any word on whether or not it can run Crysis?

running minecraft is the real benchmark my man

 

[newbmacuser2021]

running minecraft is the real benchmark my man

Minecraft is not a 3D game

 

[madsonit]

99% the 24 core. We should see 80.000-90.000 for the 32 core version.

 

[PPietra]

It definitely is the 24 core GPU version because the M1 Pro score in the database is twice as fast as the M1, on the spot!
This leads to believe that the 32 core GPU would be a little better than a RTX 3060 for laptops in this particular benchmark, not exactly as rosy as in Apple’s charts.

 

[ingambe]

It's not the raw GPU performance that is interesting, but the dedicated chip to encode audio/video/signal faster for less compute power
I'm waiting for PyTorch to use the neural engine and I will buy it

 

[vladi]

Agreed. But I don’t think you can actually utilize Metal 8K textures if the game is cross platform.

Maybe I am mistaken?

You can't utilize 8K textures anyway unless it's a linear demo scene or made for rendering. Once you throw in game physics, light dynamics, post and bunch of other things one good game requires there is no GPU that can handle all of that. Have you ever wondered why there was an obvious visual disparity between Unreal 3/4/5 stage demos and actual games built in it? That's the reason why.

 

[vladi]

Edit: I just saw that he is comparing OpenCL results again
OpenCL is abandoned on macOS, so not comparable.

But everyone on the board wants to compare newest GPUs to nVidia latest despite nVidia not supporting Metal and M1 GPU not supporting CUDA. How is that comparable then?

 

[SicTransitGloriaMundi]

To everyone that hopes this is the 24c and not the 32c version: no. The benchmark clearly states M1 Max.

you do know there's an m1max config with a 24c gpu?

 

[Kylo83]

Apple said 4x gpu compared to 5600m so this doesn’t seem right

 

[macpot]

But everyone on the board wants to compare newest GPUs to nVidia latest despite nVidia not supporting Metal and M1 GPU not supporting CUDA. How is that comparable then?

Still not accurate, but AMD Radeon RX 5700 XT has a similar Metal result to this M1 variant (68610)

That gpu (RX 5700 XT) scores 73672 on OpenCL.

So I guess you could say that this M1 Max GPU (I really think it is the 24-core model or it's the 14 inch model and throttling) is "computationally" similar to GeForce RTX 2070 Max-Q since that GPU also scores 73485 similarly.

If the M1 Max with 32-core scores ~88,000 with Metal as I think it will, then it will be on par with RTX 2080 Super Max-Q

Almost like the Desktop GeForce RTX 3060.

Like I said, still not very accurate, I don't know which benchmark can compare them head-to-head.

 

[EPO75]

well this score divided to 24 and then x32 gets you up to 90k which seems to me the score you should get with a 32core variant.

 

[iBug2]

So this M1 Max when will be placed in the bigger iMac will be slower in gpu department than the current 2 years old Gpu?
There is any chances that this could be the 24gpu ?

5700XT in the iMac has 59629 metal score. It's lower than this. And I doubt we are seeing the actual performance of this SoC on these benchmarks yet. Apple said it's 4X faster than 5600M. These numbers are not 4x more than 5600M.

 

[StoneJack]

thats a good thing? 580 i think its 4 years old...you are talking about the mobile 580 from the mac?

Asus RX 580 OC 8GB desktop version.
So I got about 41000 from that card on Intel i9 machine, while the Macbook (notebook) got 68000 score.

68000 corresponds to more modern AMD Radeon Pro Vega 56 8GB desktop version.

Apple GPU is equal to quite powerful desktop graphic cards. If it is 24 core, it plays like most of current discrete cards. The best out there RX6800 etc, so may be 32 GPU may be closer to RX6600-6800. RX6800 alone retails for 1.5K dollars, so Apple GPU is as good as 1.5K modern GPU.

In other words, we got a desktop-class graphic cards in new Macbooks.

 

[iBug2]

Apple said nearly 3x gpu compared to 5600m so this doesn’t seem right

They said 4x.

 

[Kylo83]

They said 4x.

Oh so how is this right?

 

[Michael Scrip]

Did Apple say which benchmark they used to measure the 4X increase or whatever?

Did they mention GeekBench 5 by name?

I didn't watch the keynote... so I don't know.

 

[Bandaman]

Or if you’re Sony…

Sony reveals Alpha 1 50MP full-frame camera capable of 30fps and 8K

Sony has announced the Alpha 1, a fast, high-res 50MP camera capable of shooting at 30 fps with no blackout. It can capture oversampled 8K video, has refined menus and comes in a weather-sealed body.

www.dpreview.com

That's nothing ...

https://www.dpreview.com/news/78519...capable-of-8k-120fps-16-bit-raw-video-capture

 

[macpot]

5700XT in the iMac has 59629 metal score. It's lower than this. And I doubt we are seeing the actual performance of this SoC on these benchmarks yet. Apple said it's 4X faster than 5600M. These numbers are not 4x more than 5600M.

I think that part (these GPUs being 2.5-4X the 5600M) was a mistake somehow, I don't see how that's possible...

I know TFLOPs aren't a good way to measure GPUs, but 5600M is 5.8 TFLOPS, while M1 Pro is 5.2 TFLOPs
and they said M1 Pro is 2.5X that card! even with difference in architecture and unified memory, I don't see how that could happen while M1 Pro is 2X M1, and M1 != 5600M!

Even the difference in TFLOPs of the ancient GCN vs RDNA 2 isn't that much!

P.S. I took the 5700XT's score from here.

 

[PPietra]

To everyone that hopes this is the 24c and not the 32c version: no. The benchmark clearly states M1 Max. Also this benchmark is likely run on a laptop given to reviewers, who obviously receive the most powerful version.

M1 Max has 2 versions, one with 24 GPU cores and another with 32 GPU cores.
The score for the M1 Pro (you can search the database) is exactly twice as fast as the M1... if with 16 cores is exactly twice as fast you wouldn’t get just 3 times as fast with 32

 

[rukia]

This is definitely the 24 core. 32 core is around 85K.

 

[iBug2]

I think that part was a mistake somehow, I don't see how that's possible...

I know TFLOPs aren't a good way to measure GPUs, but 5600M is 5.8 TFLOPS, while M1 Pro is 5.2 TFLOPs
and they said M1 Pro is 2.5X that card! even with difference in architecture and unified memory, I don't see how that could happen!

Even the difference in TFLOPs of the ancient GCN vs RDNA 2 isn't that much!

P.S. I took the 5700XT's score from here.

View attachment 1873259

Doubt that's the one in the iMac. iMac cards are always downclocked versions of retail AMD's. I just checked the iMac 20.2 scores in geekbench metal.

Also even M1 Pro should be faster than 5600M. I think we'll have to wait for testers like Anandtech to see the actual numbers. Apple's claims don't match these scores.

 

[zonoskar]

Put 2 (or 4?) of these M1max CPUs in a Mac pro case and you have a winner.

 

[ScottKinOC]

I hope this is not the 32-core gpu otherwise this would be a bit disappointing

Edit: I already pre-ordered my full spec MacBook with M1 Max anyways

Looks like system tested packed maybe 32GB RAM (unified memory, right) vs 64GB top spec and that may be the issue on benchmarking. Wouldn't think cores = linear in blocks of 8 vs effect of doubling RAM the way M1 architecture allows its use.