• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Apple M1 Ultra

Soldato
Joined
6 Oct 2009
Posts
4,040
Location
London
New CPU from Apple.

Technically two M1 Max fused together using an 2.5 TB/s silicone interposer, not exactly a chiplet but similar enough.
  • 16 P-cores
  • 4 E-cores
  • Up to 64 core GPU
  • 114 Billion transistors
  • 32-channel DDR5-6400 (800 GB/s bandwidth)
  • Up to 128GB
  • 850mm2
And they confirmed a Mac Pro is coming another time. Likely 4x M1 Max chips fused together.

r8lacZZ.jpg

Performance in unspecified benchmark, according to Apple, versus 12900K (with DDR5). Apple expects it to comfortably beat the 12900K given the scales, at about 1/3 the power consumption.

Given that 12900K P-cores are about 5-10% faster than M1 Max P-cores that were in laptops, it's possible that these are clocked higher.

la59Kl5.jpg

And they expect the GPU to be equal to RTX 3090 (benchmarked alongside 12900K with DDR5). Again at about 1/3 the power of RTX 3090. Obvious disclaimer that these are for compute and not gaming. M1 Ultra will be awful at gaming, just like the M1 Max was.

GkARVQn.jpg

Releases 18 March. So we'll have benchmarks in a week or so.
 
Last edited:
New CPU from Apple.

Technically two M1 Max fused together using an 2.5 TB/s intercepter, not exactly a chiplet but similar enough.
  • 16 P-cores
  • 4 E-cores
  • Up to 64 core GPU
  • 114 Billion transistors
  • 800 GB/s DDR5 Ram (roughly 2x Adler lake)
  • Up to 128GB
r8lacZZ.jpg

Performance in unspecified benchmark, according to Apple, versus 12900K (with DDR5). Apple expects it to comfortably beat the 12900K given the scales, at about 1/3 the power consumption.

Given that 12900K P-cores are about 5-10% faster than M1 Max P-cores that were in laptops, it's possible that these are clocked higher.

la59Kl5.jpg

And they expect the GPU to be equal to RTX 3090 (benchmarked alongside 12900K with DDR5). Again at about 1/3 the power of RTX 3090. Obvious disclaimer that these are for compute and not gaming. M1 Ultra will be awful at gaming, just like the M1 Max was.

GkARVQn.jpg

Releases 18 March. So we'll have benchmarks in a week or so.

It's great, however the big elephant is the price. With threadripper coming out shortly, it will be a no brainer to have a TR system specd out, with top end ram etc GPu than this
 
It's great, however the big elephant is the price. With threadripper coming out shortly, it will be a no brainer to have a TR system specd out, with top end ram etc GPu than this

Yeah this is not cheap. M1 Ultra with 128GB of Ram will cost you £5000.

A 5950X + 128GB 4000Mhz Ram will cost you about £1200, leaving £3800 for GPU and the rest of the PC. Even a workstation build with the TR 5975WX (32 cores) and 128GB of Ram will end up less than £5000.
 
Yeah this is not cheap. M1 Ultra with 128GB of Ram will cost you £5000.

A 5950X + 128GB 4000Mhz Ram will cost you about £1200, leaving £3800 for GPU and the rest of the PC. Even a workstation build with the TR 5975WX (32 cores) and 128GB of Ram will end up less than £5000.

difference is size and power I guess.

m1 ultra powers a 60w device and the case is 3liters and weights 2.7kg. It's orders of magnitude smaller than a SFF PC, not that you can do Threadripper in SFF anyway
 
difference is size and power I guess.

m1 ultra powers a 60w device and the case is 3liters and weights 2.7kg. It's orders of magnitude smaller than a SFF PC, not that you can do Threadripper in SFF anyway

Yeah there really is no competition when it comes to perf/watt and perf/size.

--

First benchmark is out. It's Geekbench so take it for what it means (Geekbench 5 sub-scores are basically good and mirror their SPEC target reasonably well, the overall average score is just weighted absurdly. But in the end it should give a ballpark figure until we have proper SPECint/SPECfp benchmarks which are the gold standards)

https://browser.geekbench.com/v5/cpu/13330272

M1 Ultra:
  • ST: 1800
  • MT: 24000
AMD 5950X:
  • ST: 1800
  • MT: 18500
Intel 12900K:
  • ST: 1900
  • MT: 17500
ThreadRipper 3970X:
  • ST: 1250
  • MT: 23000
ThreadRipper 3990X:
  • ST: 1300
  • MT: 34000
Sdlp1NG.jpg
 
At first I thought it was something similar to the old quad core Core 2s then realised it's on another level. Looks very similar to Intel's EMIB. I can't imagine it being a direct 2x boost though, surely even with the super wide/fast interconnect there has to be some performance penalties. Is macOS treating this as a single die chip or 2 NUMA nodes like TR was? We know the 2 GPU areas are being treated as one so nothing like SLi/crossfire but will there be enough bandwidth between the two?

They have very bold performance claims but only time will tell on how it actually performs.

And they mentioned they're still working on the Mac Pro, so this isn't their best chip. M1 Ultra Max anyone?
 
At first I thought it was something similar to the old quad core Core 2s then realised it's on another level. Looks very similar to Intel's EMIB. I can't imagine it being a direct 2x boost though, surely even with the super wide/fast interconnect there has to be some performance penalties. Is macOS treating this as a single die chip or 2 NUMA nodes like TR was? We know the 2 GPU areas are being treated as one so nothing like SLi/crossfire but will there be enough bandwidth between the two?

They have very bold performance claims but only time will tell on how it actually performs.

Interesting to see if CPUs will be seen as NUMA nodes or whether it will still be seen as one. The BW on that UltraFusion interconnect is 2.5 TB/s, a little insane compared to Intel/AMD.

And they mentioned they're still working on the Mac Pro, so this isn't their best chip. M1 Ultra Max anyone?

Naming is just awful. M1 Ultra > M1 Max > M1 Pro. And one more is coming.
 
The BW on that UltraFusion interconnect is 2.5 TB/s, a little insane compared to Intel/AMD.

Assuming the "worst" case where it's both directions added up (1.25 TB/s each) that's significantly more than AMD's IF. And unlike AMD's it's die to die so latency should be very low. I can see Apple getting away with treating the CPUs as a single node in that case. Still a bit worried for the GPU which is more reliant on bandwidth and how much of that interconnect is dedicated to it.
 
Reviews are out.

M1 Ultra gets 24K in Cinebench R23.

Roughly equal to 5950X, narrowly loses to 12900K. Caveat that Cinebench heavily favours SMT (Intel, AMD) to aggressive OOO execution (Apple), although author said they will address this in future versions.

GgO8nLd.png

Premier Pro is roughly equal to 12900K. Though that is single-threaded for the most part.

6sfoBh9.png

Comfortably beats 12900K in Blender:

24TSerY.png

Also beats 12900K in Davinci Resolve.

jPIJrPS.png

To get a more balanced (and workload-specific) view we need to wait until someone does the full SPEC benchmark on it.
 
Last edited:
Reviews are out.

M1 Ultra gets 24K in Cinebench R23.

Roughly equal to 5950X, narrowly loses to 12900K. Caveat that Cinebench heavily favours SMT (Intel, AMD) to aggressive OOO execution (Apple), although author said they will address this in future versions.

GgO8nLd.png

Premier Pro is roughly equal to 12900K. Though that is single-threaded for the most part.

6sfoBh9.png

Comfortably beats 12900K in Blender:

24TSerY.png

Also beats 12900K in Davinci Resolve.

jPIJrPS.png

To get a more balanced (and workload-specific) view we need to wait until someone does the full SPEC benchmark on it.

gpu benchmarks are also out, it has half the performance of the 3090 - only mentioning as Apple hinted it would match a 3090
 
gpu benchmarks are also out, it has half the performance of the 3090 - only mentioning as Apple hinted it would match a 3090

Haven't seen reasonable GPU benchmarks yet. Most just used Geekbench compute which while convenient is a nonsensical benchmark as it measures different things on different platforms (also issues with burst length causing GPUs to never reach peak performance, so even a bad one between AMD and Nvidia). Gaming was always going to be awful. Some others showed benchmarks that M1 Ultra was better but those benchmarks clearly used other accelerators on the chip so not really a test of the GPU.

It's very difficult to compare Apple and Nvidia GPUs, there's not a single test that compares them fairly given you can't run metal on Nvidia or Cuda on Apple and the software that implement both well are quite rare. Which I guess is why Apple chose to compare against Nvidia :D they could claim anything and nobody can test it. Really dishonest claims.

GFXBench tests (1440p off-screen) show comparable results between M1 Ultra and 3090, but it's known to be massively CPU-bottlenecked so again it's meaningless.

This (like the M1 Max GPU) is again the case of "good if you use macOS productivity apps, completely useless if you don't".
 
Haven't seen reasonable GPU benchmarks yet. Most just used Geekbench compute which while convenient is a nonsensical benchmark as it measures different things on different platforms (also issues with burst length causing GPUs to never reach peak performance, so even a bad one between AMD and Nvidia). Gaming was always going to be awful. Some others showed benchmarks that M1 Ultra was better but those benchmarks clearly used other accelerators on the chip so not really a test of the GPU.

It's very difficult to compare Apple and Nvidia GPUs, there's not a single test that compares them fairly given you can't run metal on Nvidia or Cuda on Apple and the software that implement both well are quite rare. Which I guess is why Apple chose to compare against Nvidia :D they could claim anything and nobody can test it. Really dishonest claims.

GFXBench tests (1440p off-screen) show comparable results between M1 Ultra and 3090, but it's known to be massively CPU-bottlenecked so again it's meaningless.

This (like the M1 Max GPU) is again the case of "good if you use macOS productivity apps, completely useless if you don't".

https://www.techpowerup.com/293019/...aims-proven-exaggerated-by-mac-studio-reviews
 
Last edited:
Fixing Apple's claims about M1 Ultra... What they should have shown :D:cry::cry:

Like... yes. If you power limit a 3090 at 300w then M1 Ultra can match it at 100w in certain productivity workloads that Apple creators care about.

But 3090 can go higher than 300w, in fact a well-cooled 3090 can even get to 450w or more.

Nvidia GPus are still far far more versatile even at lower TDPs. They can do everything the Apple GPU can, and much more.

What they achieved was impressive enough, there was no reason to mislead anyone. Nobody expected them to beat Nvidia's best in first-gen desktop Apple GPUs. They could have just focused on performance/watt and everyone would have praised them.

PgylDUG.png
 



Apple live in their own little world of reality... :cry:

Very nonsensical tests to be honest, and poorly done, Apple did mislead people with the claims (especially GPU stuff) but doesn't excuse producing this garbage for clicks, lol. Even by the already very low YouTuber standards this was embarrassing.
 
Reviews are out.

M1 Ultra gets 24K in Cinebench R23.

Roughly equal to 5950X, narrowly loses to 12900K. Caveat that Cinebench heavily favours SMT (Intel, AMD) to aggressive OOO execution (Apple), although author said they will address this in future versions.

GgO8nLd.png

Premier Pro is roughly equal to 12900K. Though that is single-threaded for the most part.

6sfoBh9.png

Comfortably beats 12900K in Blender:

24TSerY.png

Also beats 12900K in Davinci Resolve.

jPIJrPS.png

To get a more balanced (and workload-specific) view we need to wait until someone does the full SPEC benchmark on it.

5950X reviews between 27,500 and 28,500 in R23. up to 32,500 overclocked.

dluGnLT.png

https://www.cpu-monkey.com/en/cpu-amd_ryzen_9_5950x-1749
 
Back
Top Bottom