Apple M1 Ultra With 20 CPU Cores, 64 Core GPU, 32 Core Neural Engine, Up To 128GB Memory

[calc]

The only disappointing thing for me to is the relatively low core counts. Judging by their annotated die shots something like 10% of the die is cores. I would have preferred a different balance of cores and cache vs media/gpu/npu that increased it to 30 or 40% and cut back the others.

of course, I don’t think I’m their target audience. Maybe they’ll do a CPU focused IC for the Mac pro that uses discrete GPU and socketed ddr5 ram

This seems highly unlikely.

Assuming they are actually going to release a Mac Pro it would probably be at least 4 M1 Max dies connected which would be much faster than any discrete GPU available. And to get higher memory bandwidth than 400GB/s per die (so 1600GB/s+) would likely not be possible with DDR5 DIMMs.

[coder]

Maybe they’ll do a CPU focused IC for the Mac pro that uses discrete GPU and socketed ddr5 ram

Ooh. What'd be nuts is if they used the M1 Ultra in a dGPU card. There have got to be some with too many bad CPU cores, so they could just harvest those for their GPUs.

So, Apple could then make essentially a server CPU and even enter that market, as well. I'm not sure they really want to play in the server space, but it's interesting to think about.

[nranger]

Probably a low core count due to it being such a new arch. Intel and AMD have had the multi-core thing down for a while now. I gave Apple a few more years before we see 32 and 64-core models. Of course, as a CPU design, it could already be pretty efficient with just the 20 cores. Maybe they're saving 32 / 64 cores for 4 or 3nm process where it'll yield the most benefit?

My bet would be Apple bets on more accelerators on chip in future Apple Silicon generations rather than massively expand cpu and gpu core counts. More, bigger, fancier media engines, AI processing, matrix math blobs, etc. would do more to accelerate the specific work loads Apple wants customers to focus on, and do it on specific apps like Final Cut Pro, Logic Pro, Photoshop, etc. They can get more performance, with less power, while also locking customers into the Mac ecosystem, while also avoiding direct comparisons with PC software.

[coder]

Assuming they are actually going to release a Mac Pro it would probably be at least 4 M1 Max dies connected which would be much faster than any discrete GPU available. And to get higher memory bandwidth than 400GB/s per die (so 1600GB/s+) would likely not be possible with DDR5 DIMMs.

Highly unlikely.

• Many pro users require TB's of memory. Not the mere 256 GB that would provide.
• You end up with the same sort of NUMA problems of 1st gen EPYC, when you simply scale up compute dies without a central I/O die.
• GPU scaling will also be hampered by NUMA topology, particularly for actual graphics rending workloads.

[Mario Junior]

...

You forgot:

60Hz

[Slartifartblast]

Say what you want about Apple as a company, its walled gardens, the markup on its products, its hype machine, or its cult-like fanbase... but one thing that's undeniable is their actual lead, in the hardware race. Even if you go back and compare other mobile SoC's on the same process node as an Apple chip that used it, Apple was still ahead on microarchitectural sophistication, alone.

To quote Newton (Apple ?)

If I have seen further than others, it is by standing upon the shoulders of giants.

[calc]

Highly unlikely.

• Many pro users require TB's of memory. Not the mere 256 GB that would provide.
• You end up with the same sort of NUMA problems of 1st gen EPYC, when you simply scale up compute dies without a central I/O die.
• GPU scaling will also be hampered by NUMA topology, particularly for actual graphics rending workloads.

Those pro users aren't using even workstation systems, but multirack systems like HPE Superdome's that can handle 10s of TB of ram, which Apple has never targeted.

And going off package to DDR5 DIMMs would be a huge tradeoff as it only does ~ 50GB/s per channel.

But it will be interesting to see how they transparently scale their GPU with the Mac Studio.

[coder]

My bet would be Apple bets on more accelerators on chip in future Apple Silicon generations rather than massively expand cpu and gpu core counts. More, bigger, fancier media engines, AI processing, matrix math blobs, etc. would do more to accelerate the specific work loads Apple wants customers to focus on, and do it on specific apps like Final Cut Pro, Logic Pro, Photoshop, etc. They can get more performance, with less power, while also locking customers into the Mac ecosystem, while also avoiding direct comparisons with PC software.

They already have:

• fancy media engines - they claim the Ultra can decode 18 streams of 8K video (unsure of the codec, so probably H.264)
• AI processing - already has a 32-core Neural unit
• Matrix math blobs - something called "AMX" (unrelated to Intel's): https://medium.com/swlh/apples-m1-se...r-6599492fc1e1

Still, there are lots of apps that want general-purpose cores. So, true workstation users will indeed want more. We just have to wait and see what the next Mac Pro will look like, but more cores and support for external memory are both safe bets.

[coder]

To quote Newton (Apple ?)

If I have seen further than others, it is by standing upon the shoulders of giants.

Sure. They didn't invent computing. However, they have piled absolute fortunes into buying the best talent in the industry and supporting it on a well more than decade-long quest to build the best CPUs in the industry.

[jacob]

For those who want an open system (not making a judgment here, simply looking at the PoV of those for whom openness is a more important criterion than compatibility, price or convenience), does this have any potential advantage over, say, a Threadripper? Of course Threadripper is far from being exactly open, but it's still less egregious than Apple in that regard.