Apple Developing 32-Core ARM CPUs and 128-Core GPUs to Replace AMD Graphics, Report

News
By Paul Alcorn
last updated

The ARM revolution continues

Apple M1

(Image credit: Apple)

A report from Bloomberg, citing anonymous sources with direct knowledge of the matter, claims that Apple's next step to bringing ARM chips to the masses will consist of new CPUs that span up to 32 cores to supplant Intel's Xeon processors in the company's Mac Pro desktops. That move follows Apple's decision to stop using Intel processors in its future products. Apple is also reportedly developing two GPUs with 64 and 128 dedicated cores, a move obviously designed to replace AMD's discrete GPUs in both its mobile and desktop PCs.

Apple's inaugural ARM-powered M1 chips come with four high-performance and four lower-performance efficiency cores, which allowed the company to replace Intel processors in its MacBook Air, MacBook Pro, and Mac Mini lineups. With a stated plan to switch its entire lineup to its own silicon, Apple still has the task of replacing the Intel Xeon W processors it currently uses in its higher-end MacPro desktops. Bloomberg reports that Apple is developing a new 32-core model with high-performance cores for just that task, but it is unclear if those cores will be accompanied by efficiency cores as well. These chips could roll out later in 2021.

Apple is also said to be developing next-gen chips with 16 high-performance cores paired with four efficiency cores (20 total cores). Additionally, the company is developing CPUs with either 8 or 12 high-performance cores accompanied by four efficiency cores, signalling that the company plans for a full product stack of ARM processors with multiple options. These processors could begin rolling out as soon as spring 2021 for the MacBook Pro and iMac.

The new 32-core high-end chips will replace Intel's (up to) 28-core Xeon W chips that currently power Apple's workstations, but will still far short of AMD's beastly 64-core Threadripper models. Although Apple hasn't used AMD's Threadripper chips in its workstations, Apple may still trail AMD in performance while matching or exceeding Intel's current crop of chips. 

Apple's M1 chips also heralded the arrival of the company's own seven- and eight-core GPUs, but it is also reportedly developing 16- and 32-core GPUs for its high-end laptops and mid-range desktops. 

Additionally, Apple is also developing 64- and 128-core GPU models for its highest-end PCs. These new GPUs are said to be several times faster than the AMD GPUs Apple uses in its current products, portending a severing of ties with AMD for future Apple PCs. 

The new 64- and 128-core graphics chips are expected to debut "later in 2021 or potentially 2022."

Paul Alcorn
Paul Alcorn
Managing Editor: News and Emerging Tech

Paul Alcorn is the Managing Editor: News and Emerging Tech for Tom's Hardware US. He also writes news and reviews on CPUs, storage, and enterprise hardware.

See more CPUs News
28 Comments Comment from the forums
  • nofanneeded
    Apple will have a hard time convincing people to pay $10000 for a MAC PRO with ARM CPU and Apple made GPU ...

    Also 32 cores ARM is like 5GHZ 16 cores /32 threads Ryzen in performance .

    I dont think that Apple could clock 32 cores higher than 2.5 Ghz each ..
    Reply
  • alquix
    nofanneeded said:
    Apple will have a hard time convincing people to pay $10000 for...
    Thats what I've thought about most of their major moves, but here we are. Their track record is pretty impressive
    Reply
  • GoatGuy
    nofanneeded said:
    Apple will have a hard time convincing people to pay $10000 for a MAC PRO with ARM CPU and Apple made GPU ...

    Also 32 cores ARM is like 5GHZ 16 cores /32 threads Ryzen in performance .

    I dont think that Apple could clock 32 cores higher than 2.5 Ghz each ..
    I don't think so.

    Ryzen (as an example) for a single core, single thread, is about 25% slower than M1, single core/thread. Turn on 2 threads on that same single core, and M1 and Ryzen 3 at the same GHz are about the same, except when very wide X86-extensions are used. X86 beats M1, CORE-for-CORE then. Takes 2 threads per core to do it, but it does.

    However, under all but the most contrived circumstances, SMT only adds between 25% and 40% more per-core performance, per GHz. Certainly no-where near 100% better throughput. Nowhere near.

    My thinking is that first RYZEN then belatedly (but rather impressively), Intel will both leap full-belly into the 5 nanometer EUV world, stepping up layer counts (at 5 NM) for some truly spectacular performance increases in the next 2 years. And the cores-per-chip will also rise. Cores and cache; for Ryzen, also with increasingly competent on-chip GPU coprocessor services. Like the M1, in a way. But bigger physical pinout package, for dedicated GPU memory not-affiliated with 'unified' memory of the M1. Non Von-Neuman machine.

    Basically, we are looking at the breakthru to 6 GHz (or higher!) computing, at sub 0.8 V logic. 5 nm chip structures, very heavily weighted toward big.SMALL.gpu architecture. Also note that with little penalty, even the X86 architecture can be big.SMALL reconfigured. I could easily see AMD popping forth an 8s + 16B chip, single chiplet. 5 GHz before 'turbo'.

    That'd give Apple a real run for the money. Never, ever underestimate the ability fo both AMD and INTEN to seriously invest-and-step-up to compete with the 'new guys'. Putting 12 chiplets ona single Threadripper package ... for 192 cores, maybe 300 threads ... seriously changes what a single-package compute device means.

    GoatGuy
    Reply
  • syadnom
    nofanneeded said:
    Apple will have a hard time convincing people to pay $10000 for a MAC PRO with ARM CPU and Apple made GPU ...

    Also 32 cores ARM is like 5GHZ 16 cores /32 threads Ryzen in performance .

    I dont think that Apple could clock 32 cores higher than 2.5 Ghz each ..

    I have clients with multiple Mac Pros, each costing well over $10k. If it performs, they don't care what's inside.

    I happen to own an M1 mac mini and an M1 Macbook Pro. They are fantastic. Smokes my previous Macbook Pro handily without the case getting warm. It runs x86 software faster than my previous macbook pro, including fusion 360 which isn't ported to M1 yet and runs smoother.

    I have no doubt apple can put 32 cores together and maintain high speeds because they have the thermal headroom to do it. In a Mac Pro they can put a big heat sink on and crank the volts up.

    And an 8 core chip with 4 performance cores (which is all that's being used in benchmarking) is already keeping up with 8 core Ryzens. I don't think your claim that 32 M1 cores is equivalent to 16 ryzen cores holds water at all, if we are measuring M1 performance cores to Ryzen cores they are very similar overall.

    I would expect that a 32 core M1 (32 performance + 4 efficient?) will absolutely smash a 16c/32t ryzen and every single core benchmark out there on the M1 supports this.

    For the sake of argument, call an M1 performance core roughly equivalent to a Ryzen or intel core in performance.
    Reply
  • Makaveli
    GoatGuy said:
    I don't think so.

    Ryzen (as an example) for a single core, single thread, is about 25% slower than M1, single core/thread. Turn on 2 threads on that same single core, and M1 and Ryzen 3 at the same GHz are about the same, except when very wide X86-extensions are used. X86 beats M1, CORE-for-CORE then. Takes 2 threads per core to do it, but it does.

    However, under all but the most contrived circumstances, SMT only adds between 25% and 40% more per-core performance, per GHz. Certainly no-where near 100% better throughput. Nowhere near.

    My thinking is that first RYZEN then belatedly (but rather impressively), Intel will both leap full-belly into the 5 nanometer EUV world, stepping up layer counts (at 5 NM) for some truly spectacular performance increases in the next 2 years. And the cores-per-chip will also rise. Cores and cache; for Ryzen, also with increasingly competent on-chip GPU coprocessor services. Like the M1, in a way. But bigger physical pinout package, for dedicated GPU memory not-affiliated with 'unified' memory of the M1. Non Von-Neuman machine.

    Basically, we are looking at the breakthru to 6 GHz (or higher!) computing, at sub 0.8 V logic. 5 nm chip structures, very heavily weighted toward big.SMALL.gpu architecture. Also note that with little penalty, even the X86 architecture can be big.SMALL reconfigured. I could easily see AMD popping forth an 8s + 16B chip, single chiplet. 5 GHz before 'turbo'.

    That'd give Apple a real run for the money. Never, ever underestimate the ability fo both AMD and INTEN to seriously invest-and-step-up to compete with the 'new guys'. Putting 12 chiplets ona single Threadripper package ... for 192 cores, maybe 300 threads ... seriously changes what a single-package compute device means.

    GoatGuy

    Why is your example for Ryzen using Gen 1?

    What is the single core performance for Zen 3 vs M1?
    Reply
  • jkflipflop98
    I wonder how much Apple has to pay Nvidia on licensing fees for using ARM.
    Reply
  • Makaveli
    jkflipflop98 said:
    I wonder how much Apple has to pay Nvidia on licensing fees for using ARM.

    Those deals would have already been in place way before Nvidia bought ARM.

    I highly doubt they would go back and try to change the licensing fee's on a signed contract. So it should stay the same as they were paying before.
    Reply
  • nofanneeded
    GoatGuy said:
    I don't think so.

    Ryzen (as an example) for a single core, single thread, is about 25% slower than M1, single core/thread.

    GoatGuy

    lolz really? Actually M1 is 8% slower than Ryzen 3 in single core

    https://www.cpu-monkey.com/en/compare_cpu-apple_m1-1804-vs-amd_ryzen_9_5950x-1749
    you are 1000% Apple fanboy
    Reply
  • jkflipflop98
    Makaveli said:
    Those deals would have already been in place way before Nvidia bought ARM.

    I highly doubt they would go back and try to change the licensing fee's on a signed contract. So it should stay the same as they were paying before.

    Maybe. However there is this . . .
    Reply
  • nofanneeded
    syadnom said:
    I happen to own an M1 mac mini and an M1 Macbook Pro. They are fantastic. Smokes my previous Macbook Pro handily without the case getting warm. It runs x86 software faster than my previous macbook pro, including fusion 360 which isn't ported to M1 yet and runs smoother.

    Your claims are false. unless your old macbooks are 3gens old.
    Reply