Apple M5 silicon could ditch unified memory for split CPU and GPU designs

[zohaibahd]

Rumor mill: Reports are emerging about some wild changes coming to Apple's next-gen M5 chips. One of these is the lineup ditching the unified architecture that's been a signature of Apple silicon so far. Instead, the CPU and GPU will use separate memory pools.

The claim comes from analyst Ming-Chi Kuo. He says that all M5 variants, including the base M5, M5 Pro, M5 Max, and M5 Ultra, will be fabricated on TSMC's advanced 3nm N3P process node. That's a change from the current M4 and A18 Bionic chips that use the slightly older N3E node.

The potential split CPU/GPU architecture is definitely the biggest surprise, though. Since the M1's debut, Apple silicon has used a unified memory pool shared between the CPU and GPU cores. This unified memory access (UMA) design is widely credited with the incredible performance-per-watt efficiency we've seen on MacBooks. If these rumors hold, that design won't be long for this world. While separating the CPU and GPU memory spaces adds some complexity, it could also unlock performance gains in certain workloads.

Apple M5 series chip

1. The M5 series chips will adopt TSMC's advanced N3P node, which entered the prototype phase a few months ago. M5, M5 Pro/Max, and M5 Ultra mass production is expected in 1H25, 2H25, and 2026, respectively.
2. The M5 Pro, Max, and Ultra will utilize… https://t.co/XIWHx5B2Cy

– 郭明錤 (Ming-Chi Kuo) (@mingchikuo) December 23, 2024

To pull off this change, Apple is reportedly turning to TSMC's server-grade 2.5D packaging tech called System on Integrated Chips-Molding Horizontal (SoIC-mH). SoIC is TSMC's take on 3D stacking and hybrid wafer bonding, which enables ultra-dense connections between two chips. The "mH" variant Apple is using allows them to glue the separate dies horizontally onto the package rather than a vertical 3D stack.

This advanced packaging allows for those split designs while promising better yields and superior thermal performance.

Kuo says that the new packaging approach will particularly benefit Apple's Private Cloud Compute ambitions. This is what the company calls its cloud intelligence system designed specifically for AI processing.

Ultimately, the M5 lineup is shaping up to be Apple's most versatile silicon yet. Kuo says that the base M5 will go into mass production in the first half of 2025, meaning it should probably find its way into updated Macs in late 2025. Meanwhile, the M5 Pro/Max variant is expected to enter production in the second half of next year. Lastly, the M5 Ultra is expected to debut in 2026.

Permalink to story:

Apple M5 silicon could ditch unified memory for split CPU and GPU designs

 

[DanteSmith]

Utilizing MCM does not inherently prevent the design from accessing a shared memory pool between the CPU and GPU...

 

[GodisanAtheist]

Imagine if we end up with Apple dGPUs after all of this...

 

[kiwigraeme]

Imagine if we end up with Apple dGPUs after all of this...

Even if Valve get all PCs games on their version on LInux , with my level of ignorance , not sure how helpful for Apple to use that to get them onto their flavour of Unix , and given Apples past history do they care or want that, as they like to total control

Maybe it would allow a held hand , or super premium Apple TV gaming box
TVs are 16:9 same as older iPhones

But outside of causals no one trusts Apple for serious gaming or more agnostic gaming not tied to their App store

Plus we all know whatever Apple does it's only a few years ahead , so why enslave yourself to them

Also Apple really needed to go after the Nintendo market - fun colourful family games , too many good handhelds coming with emulators to boot

Maybe the Banana in the Room is AI

Apple need AI on a chip that's more powerful Nvidia style . As currently set up works perfectly fine

So I'm calling it this is for AI

 

[rmcrys]

Apple’s chips got too big and complex and that hurts yields. I imagine they want to simulate a unified chip but being two dies glued together, so they don’t depend too much on the unified die getting good yields. As ReBAR exists on the PC, they can give the CPU and GPU memory the same treatment and it’ll be about the same.

Anyway, gaming on Mac is software limited and few will trade their older M series for the new ones as the gaming / 3D possibilities are very scarce…

 

[seeprime]

Even if Valve get all PCs games on their version on LInux , with my level of ignorance , not sure how helpful for Apple to use that to get them onto their flavour of Unix , and given Apples past history do they care or want that, as they like to total control

Maybe it would allow a held hand , or super premium Apple TV gaming box
TVs are 16:9 same as older iPhones

But outside of causals no one trusts Apple for serious gaming or more agnostic gaming not tied to their App store

Plus we all know whatever Apple does it's only a few years ahead , so why enslave yourself to them

Also Apple really needed to go after the Nintendo market - fun colourful family games , too many good handhelds coming with emulators to boot

Maybe the Banana in the Room is AI

Apple need AI on a chip that's more powerful Nvidia style . As currently set up works perfectly fine

So I'm calling it this is for AI

That's a lot of words to say you're calling it for AI.

 

[kiwigraeme]

That's a lot of words to say you're calling it for AI.

Thanks , AI thought came near end of my usual rambling. Cloud AI - is probably expensive, so as much as you can offload back to customer. Think there are few precedents of offloading from Cloud to phones , with translators etc

 

[toooooot]

Apple’s chips got too big and complex and that hurts yields. I imagine they want to simulate a unified chip but being two dies glued together, so they don’t depend too much on the unified die getting good yields. As ReBAR exists on the PC, they can give the CPU and GPU memory the same treatment and it’ll be about the same.

Anyway, gaming on Mac is software limited and few will trade their older M series for the new ones as the gaming / 3D possibilities are very scarce…

QUALCOMM new CPU for PCs shows a promise. It has much better compatibility, I saw people play multiple games on Samsung Book edge 4. It is very similar to Apple in a sense that it is less power hungry, can be very thin and light, but can nevertheless run a lot of Windows apps.
I am sure there are some people who wish to have Apple's features such as long lasting battery on a laptop, while also wishing they could do a few things Windows computers can.
It is much newer though. Which means Apple is most likely better in any other aspect.
But if Qualcomm keeps pushing, they will have plenty of customers for these chips.

 

[poshflamingos]

Utilizing MCM does not inherently prevent the design from accessing a shared memory pool between the CPU and GPU...

You're right, it doesn't. But if you're splitting the CPU and GPU anyway, each with its own memory bus, then might as well give each of them its ideal type of memory (DDR for lower latency and lower bandwidth for CPUs, GDDR for higher bandwidth and higher latency for GPUs) instead of an unified pool that will inevitably sacrifice performance in one of these two aspects, for the CPU (if it's GDDR) or the GPU (if it's DDR).

 

[cyberlizard]

If it’s about profits maybe they should have a singular die with all the trimmings, sure it would reduce their line up but we would have a singular processor across the board that was simply the best they had to offer.

I do wish we had the option of external gpu or gpu cards on apple silicon as well as the ability to upgrade ram or storage after the fact.

 

[rmcrys]

QUALCOMM new CPU for PCs shows a promise. It has much better compatibility, I saw people play multiple games on Samsung Book edge 4. It is very similar to Apple in a sense that it is less power hungry, can be very thin and light, but can nevertheless run a lot of Windows apps.
I am sure there are some people who wish to have Apple's features such as long lasting battery on a laptop, while also wishing they could do a few things Windows computers can.
It is much newer though. Which means Apple is most likely better in any other aspect.
But if Qualcomm keeps pushing, they will have plenty of customers for these chips.

In order to work well for ARM chips, the work has to be both sideways: the hardware (fast emulation, matrix); the software has to support it.

Presently the issue is with the software, specially the DRM / anti piracy software that is mainly x86 only.

 

[maxxcool7421]

This is not a suprise. despite a really good memory controller UMA had bandwidth drawbacks when you are running AI jobs, or running heavy classic core compute tasks and GPU accelerated jobs at the same time needing large pages of gpu ram.

I expect a simple change of making ''some'' dedicated ram and traces to that ram to increase gpu speed when running certain workloads.

 

[bitwarden]

Could this move be due to a hardware insecurity with shared memory, e.g. any CPU secrets potentially accessible to GPUs?
I imagine fixing this can only increase game loading delays, potentially other workloads.