Future M2 Ultra chip

[sam_dean]

Last year's ‌M1‌ Ultra is essentially two ‌M1 Max‌ chips that have been connected together through an Apple-developed method, so it offers double the performance of the ‌M1 Max‌.

Using the same method this year the M2 Ultra would have 2x the performance of a M2 Max.

Possible M2 Ultra specs vs M1 Ultra

Apple Silicon 5nm chip	Possible M2 Ultra	M1 Ultra
CPU	24-Core	20-Core
High-performance	16×	16×
High-efficiency	8×	4×
GPU	76-Core	64-Core
Neural Engine	32‑Core	32‑Core
Transistors	134 billion	114 billion
Max unified memory	192GB	128GB
Memory bandwidth	800GB/s	800GB/s

I can see the 2023 Mac Studio M2 Ultra being released by WWDC 2023 in June.

A more substantial improve of an Ultra chip would probably come with the M3 Ultra when the die shrink changes from 5nm to 3nm next year right after 3nm A17 Bionic chip that will be 1st used on the iPhone 15 Pro in Sep 2023.

With improved industry-leading performance per Watt I expect the max power consumption would be dramatically less than 200W.

 

[Appletoni]

Last year's ‌M1‌ Ultra is essentially two ‌M1 Max‌ chips that have been connected together through an Apple-developed method, so it offers double the performance of the ‌M1 Max‌.

Using the same method this year the M2 Ultra would have 2x the performance of a M2 Max.

Possible M2 Ultra specs vs M1 Ultra

Apple Silicon 5nm chip	Possible M2 Ultra	M1 Ultra
CPU	24-Core	20-Core
High-performance	16×	16×
High-efficiency	8×	4×
GPU	76-Core	64-Core
Neural Engine	32‑Core	32‑Core
Transistors	134 billion	114 billion
Max unified memory	192GB	128GB
Memory bandwidth	800GB/s	800GB/s

I can see the 2023 Mac Studio M2 Ultra being released by WWDC 2023 in June.

A more substantial improve of an Ultra chip would probably come with the M3 Ultra when the die shrink changes from 5nm to 3nm next year right after 3nm A17 Bionic chip that will be 1st used on the iPhone 15 Pro in Sep 2023.

With improved industry-leading performance per Watt I expect the max power consumption would be dramatically less than 200W.

It's about time that the M2 Ultra finally makes its way into the MacBook Pro 20-inch M2 Ultra.

The M1 Ultra has: 16x and 4x
Hardly anyone will buy an M2 Ultra that only has 16x and 8x.
The competition has significantly increased the number of CPU cores as well as the number of performance cores.

For example, Stockfish achieves about the same performance with 2 high-efficiency cores as with 1 high-performance core.
M1 Ultra: 16x + 4x = 18
M2 Ultra 16x + 8x = 20
This corresponds to a performance increase of only about 10%.

If Apple had introduced an M2 Ultra with 20x + 4x = 22 instead, we would now have an increase of 20%!!!

The GPU only has 12 cores more.
That's only about +18%.

Only the RAM has become noticeably larger.
The RAM unfortunately does not even reach the 256 GB, let alone 512 GB RAM.

Unfortunately, the memory bandwidth has remained the same and the number of cores in the Neural Engine has not increased either.

16 TB SSD are still not available!!
Raytracing cores are still not available.
Tensor cores are still not available.
AVX-512 (VNNI) is still not available!!!
...

 

[leman]

It's about time that the M2 Ultra finally makes its way into the MacBook Pro 20-inch M2 Ultra.

Oh no, he is back.

Tensor cores are still not available.

Apple had matrix accelerators for years.

AVX-512 (VNNI) is still not available!!!

AVX512 is an x86 extension. What are you even talking about? The functionality offered by VNNI (which by the way is not available on any of customer Intel CPUs) is implemented by the NPU on Apple Silicon.

 

[jujoje]

I can see the 2023 Mac Studio M2 Ultra being released by WWDC 2023 in June.

A more substantial improve of an Ultra chip would probably come with the M3 Ultra when the die shrink changes from 5nm to 3nm next year right after 3nm A17 Bionic chip that will be 1st used on the iPhone 15 Pro in Sep 2023.

Given the improvements in gpu scaling and the nice cpu bump with the M2 Max the Ultra should be a pretty nice improvement, particularly the 20-30% improvement in GPU.

There was vague conjecture that the Ultra will skip M2 and go straight to M3 - particularly if it's going to be the basis of the Mac Pro - a Mac Pro which is just an Ultra chip + some expansions slots would be pretty weak for the market it's targeting. Although given the (rumoured) delays, probably wishful thinking. Would like some raytracing core though.

 

[TDTOMW]

I have a Question about Apple's system on a chip Technology and the concern people have about the lack of GPU performance of the rumored Mac Pro. I’m not very technical. But I would like to ask you if Apple could separate the CPU and other media engines on one chip and the GPU on another chip and use the ultrafusion technology to fuse the two independent chips together making it easier to increase the core counts of the GPU, and CPU.


Given that the ultrafusion technology of the Apple system on a chip that it would give it the appearance of being one chip of having a very powerful, GPU. This may mean a redesign of the CPU architecture, and a separate GPU architecture, but couldn't it allow Apple to use the die size of the M series chips to accomplish greater core counts for both CPU and GPU?

 

[sam_dean]

I have a Question about Apple's system on a chip Technology and the concern people have about the lack of GPU performance of the rumored Mac Pro. I’m not very technical. But I would like to ask you if Apple could separate the CPU and other media engines on one chip and the GPU on another chip and use the ultrafusion technology to fuse the two independent chips together making it easier to increase the core counts of the GPU, and CPU.


Given that the ultrafusion technology of the Apple system on a chip that it would give it the appearance of being one chip of having a very powerful, GPU. This may mean a redesign of the CPU architecture, and a separate GPU architecture, but couldn't it allow Apple to use the die size of the M series chips to accomplish greater core counts for both CPU and GPU?

I guess that could be done but would it be economical to do? Apple has teams of people who focus on performance and use case. That's why Apple Silicon is designed to address key workload types that typical Mac users do.

Also reason why it does not do well for unicorn-like unique Windows-specific workloads that is largely absent on Macs.

UltraFusion as implemented matches identical Max chips together so that it just doubles everything. On the macOS-end it simplifies things.

What you are proposing may require additional software development resources.

 

[TDTOMW]

I guess that could be done but would it be economical to do?

UltraFusion as implemented matches identical Max chips together so that it just doubles everything. On the macOS-end it simplifies things.

What you are proposing may require additional software development resources.

As you say economically, I believe Apple was thinking of cost over function. they could get more use out of one chipset and it works to a degree, but as they use the same chipset, it bottlenecks. when the that chipset works fine for consumer products, you need a second chipset for the pro models where you can independently create larger core counts for both CPU and GPU that may also give you the possibility to increase de unified ram.

 

[leman]

I have a Question about Apple's system on a chip Technology and the concern people have about the lack of GPU performance of the rumored Mac Pro. I’m not very technical. But I would like to ask you if Apple could separate the CPU and other media engines on one chip and the GPU on another chip and use the ultrafusion technology to fuse the two independent chips together making it easier to increase the core counts of the GPU, and CPU.


Given that the ultrafusion technology of the Apple system on a chip that it would give it the appearance of being one chip of having a very powerful, GPU. This may mean a redesign of the CPU architecture, and a separate GPU architecture, but couldn't it allow Apple to use the die size of the M series chips to accomplish greater core counts for both CPU and GPU?

Yes, absolutely. Assuming the technical part (the interconnect) is solved, the main remaining problem would be economical. Every new die in production comes with a significant cost overhead and you need to make a lot of them to amortize this cost. Making separate GPU die just for the low volume Mac Pro is probably not feasible. Making a small GPU die that will be used across different devices is more interesting.


By the way, Apple is way ahead of you: https://patents.google.com/patent/US20210125967A1/en

UltraFusion as implemented matches identical Max chips together so that it just doubles everything. On the macOS-end it simplifies things.

What you are proposing may require additional software development resources.

Why would software care? Apples MCM simply stitches together multiple chips into one big chip. From the software perspective, it doesn’t make any difference how many actual dies are there. Of course, the system needs to be aware of the overall topology for scheduling etc. purposes, but that’s also the case for something like Max.

 

[sam_dean]

As you say economically, I believe Apple was thinking of cost over function. they could get more use out of one chipset and it works to a degree, but as they use the same chipset, it bottlenecks. when the that chipset works fine for consumer products, you need a second chipset for the pro models where you can independently create larger core counts for both CPU and GPU that may also give you the possibility to increase de unified ram.

Depends on the business model of the company making the chips.

What you described works for parts companies like Intel/AMD/Nvidia as they sell their processors to OEMs, so they only get profit from the processor not the finished system. So they have to prioritize cost, by optimizing their designs for Area first and then focus on power. This is why both AMD and Intel use smaller cores, which allows them for smaller dies. But which have to be clocked faster in order to compete in performance, unfortunately that also increases power.

In contrast; Apple is a system's vendor. Meaning that they sell the finished product, not just the processors. So they can use several parts from the vertical process to subsidize others. In this case, Apple can afford to make very good SoCs because they don't sell those chips elsewhere, meaning that they are not as pressured to make them "cheap" in terms of area for example. Since they're going to recoup the profit from elsewhere in the product.

This is probably they key difference; Apple can afford the larger design that is more power efficient for the same performance. Whereas AMD/Intel have to aim for the smaller design that is less power efficient for the same performance.

 

[sam_dean]

Why would software care? Apples MCM simply stitches together multiple chips into one big chip. From the software perspective, it doesn’t make any difference how many actual dies are there. Of course, the system needs to be aware of the overall topology for scheduling etc. purposes, but that’s also the case for something like Max.

Thank you for answering your own question.

Again, it boils down to if the SoC will benefit >80% of user's use case.

I read one Mathematica user complaining about Ultra parts having more GPU cores they'd care for.

How many Mathematica users are there relative to video editors who get a dedicated Media Engine?

Within the first 3 months of the M1 Ultra's release the GPU cores were commercially comparable (not identical) to a RTX 3090 relative to the whole system's cost.

At the time of their release a one die M2 Ultra and two die M2 Ultra would be comparable and even exceed the RTX 40xx.

 

[DWHH1]

It's about time that the M2 Ultra finally makes its way into the MacBook Pro 20-inch M2 Ultra.

The M1 Ultra has: 16x and 4x
Hardly anyone will buy an M2 Ultra that only has 16x and 8x.
The competition has significantly increased the number of CPU cores as well as the number of performance cores.

For example, Stockfish achieves about the same performance with 2 high-efficiency cores as with 1 high-performance core.
M1 Ultra: 16x + 4x = 18
M2 Ultra 16x + 8x = 20
This corresponds to a performance increase of only about 10%.

If Apple had introduced an M2 Ultra with 20x + 4x = 22 instead, we would now have an increase of 20%!!!

The GPU only has 12 cores more.
That's only about +18%.

Only the RAM has become noticeably larger.
The RAM unfortunately does not even reach the 256 GB, let alone 512 GB RAM.

Unfortunately, the memory bandwidth has remained the same and the number of cores in the Neural Engine has not increased either.

16 TB SSD are still not available!!
Raytracing cores are still not available.
Tensor cores are still not available.
AVX-512 (VNNI) is still not available!!!
...

Reluctantly I have to suggest that Apple Silicon has not been the wonder Apple would like us to believe. The M2 Pro and M2 Max were delayed at the end of last year so the MacBook Pro updates moved from last Oct/Nov to Feb23 . The alleged nudge/nudge/wink/wink M1/M2 Extreme has apparently been cancelled. There are no suggestions of a Mac Studio Max upgrade to M2 Max chips despite M2 Max clearly being available. The Mac Pro replacement has drifted right with a suggestion it might be equipped with a M2 Ultra instead of an M2 Extreme although an M3 something chip is yet to be leaked by the 'Friends of Apple' clique that Apple leaks through.
Is it possible that Apple is experiencing chip yield problems, thermal management issues it can't get around, slow transition to and availability of M3 affecting its entire product line as well as yet more delays to the AR/VR headset?
On the positive side though we should not underestimate the difficulty of the transition Apple is attempting nor the impressive performance achieved so far with low power that we have seen.
I wish Apple would not leak 'unofficial official' announcements that they will not stand with when Apple would sue any of its staff for leaking the real truth and cause of Apple products delays.

There is a large family of us who would not dream of buying anything but an Apple pruduct so why not be honest with us?

 

[mcnallym]

Well they aren’t going to launch the studio until the M2 Ultra is ready to go AND available in quantity.
as they can get two max for 1 ultra and just launched the MBP then makes more sense to get those out the door then the studio.

once MBP demand satisfied then can look at m2 Ultra and studio, which is what they did with the M1 generation of MBP and Studio.

MBP when just checked about a 2 week time for delivery, with a BTO.

everyone is suffering supply chain issues at the moment, let alone new products.

company ordered a new ms125-48lp meraki switch and told is a 24 week lead time, or basically 1/2 a year.

other network vendors same sort of timescales.