M4+ Chip Generation - Speculation Megathread [MERGED]

[leman]

Any rumors/truth to this ?

M4 Mac mini Is A Capable Machine For Running Older AAA Games Like Death Stranding; Base Configuration Pushes The 60+FPS At Very High Quality Settings

M1-M5 is one (first) generation ending with 3nm fabrication. M6-M10 is second generation starting with 2nm EUV Lithography. Everything in between is an iteration of the former where Asahi Linux clarifies that even further. You're almost at M5, all that experience from 1st generation would be...

disq.us

I have difficulty parsing what this person is saying. It they mean that M6 is likely to be the first Apple chip on M2, yeah, that sounds likely.

 

[tenthousandthings]

I have difficulty parsing what this person is saying. It they mean that M6 is likely to be the first Apple chip on M2, yeah, that sounds likely.

He’s replying to someone considering replacing an M1 iMac with an M4 mini and his advice is to be patient and then “reward yourself” with an M6 Mac Studio, so I don’t think we’re talking about deep insights into industry trends, if you catch my drift.

I think he’s saying that Apple silicon is here to stay because Qualcomm is competing with it on ARM and the Asahi Linux project exists. He’s implying that if the industry hadn’t responded in this way, then Apple silicon would have been a dead end. I doubt anyone here agrees with that!

I do think it’s possible TSMC 2nm will be a watershed, because of the new transistor architecture in the first generation and the advent of back side power delivery in the second and/or next generation, along with the continued evolution of advanced packaging now that it is in the wild, from Apple’s Ultra initiative to Nvidia’s Grace Blackwell. So the comment in question is right that a lot of valuable lessons have been learned by TSMC and its customers from N5 to N3P, but the pace of change is more incremental than he suggests.

 

[Xiao_Xi]

An Apple engineer posted last month:

The Apple M4 line of cores includes the Scalable Matrix Extension (SME) feature. The M4s do not implement Scalable Vector Extension (SVE), although the processor is in Streaming SVE Mode when the SME is being used. The most obvious side effects of being in SSVE Mode are that (on the M4 cores) NEON instructions cannot be used, and watchpoints may get false positives, the address comparisons are done at a lowered granularity.

When SSVE mode is enabled, the kernel will provide the Streaming Vector Length register, which is a maximum of 64 bytes with the M4. Also provided are SVCR (with bits indicating if SSVE mode and SME mode are enabled), TPIDR2, SVL. Then the SVE registers Z0..31 (SVL bytes long), P0..15 (SVL/8 bytes), the ZA matrix register (SVL*SVL bytes), and the M4 supports SME2, so the ZT0 register (64 bytes).

When SSVE/SME are disabled, none of these registers are provided by the kernel - reads and writes of them will fail.

[lldb][debugserver] Read/write SME registers on arm64 by jasonmolenda · Pull Request #119171 · llvm/llvm-project

Note: The register reading and writing depends on new register flavor support in thread_get_state/thread_set_state in the kernel, which will be first available in macOS 15.4. The Apple M4 line of c...

github.com

It reminded me of last year's discussion about SME and SVE.

 

[krell100]

I don’t see apple releasing an M5 Max/ultra first. That would be a complete 180 on their schedule to this point. I suppose anything is possible but I’d say that’s unlikely. IMHO of course.

 

[thenewperson]

I don’t see apple releasing an M5 Max/ultra first. That would be a complete 180 on their schedule to this point. I suppose anything is possible but I’d say that’s unlikely. IMHO of course.

They could preview it though. Preview at WWDC and promise and end of year (or next year) release. They’ve done this for previous Mac Pros.

 

[novagamer]

They could preview it though. Preview at WWDC and promise and end of year (or next year) release. They’ve done this for previous Mac Pros.

This assumes no M4 Ultra and no Mac Studio. I would be gobsmacked if they previewed both machines 6 months ahead of time, and releasing an M4 Studio with an M5 fully previewed on the horizon would kneecap that machine's sales

I think a Hidra surprise chip and/or M4 Ultra are the two options, at least until we get die shots disproving the interconnect (why are they taking so long?!)

 

[DrWojtek]

I have to agree with that guy who said he’s less likely to buy now that development is going much faster than with Intel.

It’s like deflation: the longer I wait, the more my money is worth.

This of course assumes I want to buy partly because of enjoyment, not 100% need.

Or maybe I’ve just gotten older.

 

[thenewperson]

This assumes no M4 Ultra and no Mac Studio. I would be gobsmacked if they previewed both machines 6 months ahead of time, and releasing an M4 Studio with an M5 fully previewed on the horizon would kneecap that machine's sales

I think a Hidra surprise chip and/or M4 Ultra are the two options, at least until we get die shots disproving the interconnect (why are they taking so long?!)

Well, the way I reconcile all the rumours is this:

- A Mac Studio with M4 Ultra will be released. IIRC there was space in the identifiers list this.

- A Mac Pro with [unspecified newer cores] is previewed, with a promised release date for December 2025 or ‘next year’. By the time this machine is released most (or all) of the M5 gen would have been released. These would be the Mac17 identifiers in the list.

 

[innerproduct]

Remember that a m4ultra would be a very powerful system for many many tasks if it is a 2x m4max. Very similar in perf to a threadripper pro 32 core + a rtx 4090 but with much higher mem perf and shared mem up to 256 GB. Yes, not a server grade machine or top end ws. For comparison, Out of desperation i just configured a threadripper + an old a100 (available through my reseller) and that costs me about 35000$. I really need the mem so this is the cheapest option on the nvidia side. H100 starts at 60k. Blackwell i dont even know.

 

[tenthousandthings]

I don’t see apple releasing an M5 Max/ultra first. That would be a complete 180 on their schedule to this point. I suppose anything is possible but I’d say that’s unlikely. IMHO of course.

The simplest explanation is Ultra will no longer use Max.

I agree it’s unlikely M5 Max will launch anywhere other than in the 2025 MacBook Pro.

The most likely scenario, given what we think we know, reconciling the identifiers list with Gurman’s timeline predictions and the Hidra rumor, is as follows:

Mac16,9 = Mac Studio M4 Max

The next Ultra is announced at WWDC. It is on N3P and it uses A19/M5 core designs. It does not use Max. It still uses UltraFusion (perhaps 2.0 but not necessarily), but instead of using two Max it is something like Nvidia’s fusion of Grace CPU with Blackwell GPU. The two components are different, and there are configurations available for the Mac Pro that go beyond the Mac Studio:

Mac17,1 = Mac Studio M5 Ultra
Mac17,2 = Mac Pro M5 Ultra

Another feature of the WWDC announcement will be about Apple Intelligence servers, and how the new, improved Ultra fits into that context.

 

[Antony Newman]

Are the community that are buying the Mac Pros only doing because they need to utilise 3rd party PCI cards?
Is the market for hardware acceleration shrinking as Apples Mac Studio become more powerful?

Given the choice - would that community rather have an external PCIe Gen5 / TB5 chassis that could connect to a Mac Studio using a couple of TB5 connectors? This would allow Apple to charge for a PCIe one-off ‘business tax’ - and allow those companies to more frequently upgrade their Mac Studios.

If Apple did intend to bifurcate the architecture for its Mac Pro’s on performance basis - it would require the Mac Pros to have the same release cadence as the Mac Studio going forwards (or the next Studio could overtake the Pro) - AND for there to either be an architectural difference between every Pro and every Ultra going forwards (eg a 4 chip version / workstation class SoC) - or that the Studio is thermally constrained.

If Apple do have a workstation class SoC - Surely they would want to sell it in larger quantities (ie have it available in the Mac Studio)? If it did need 50% more heat dissipation - and a 15% larger Mac Studio be sufficient - would this not be a preferred option to the masses that Don’t need the PCIe tower, but do want the most performant Mac available?

 

[tenthousandthings]

Are the community that are buying the Mac Pros only doing because they need to utilise 3rd party PCI cards?
Is the market for hardware acceleration shrinking as Apples Mac Studio become more powerful?

Given the choice - would that community rather have an external PCIe Gen5 / TB5 chassis that could connect to a Mac Studio using a couple of TB5 connectors? This would allow Apple to charge for a PCIe one-off ‘business tax’ - and allow those companies to more frequently upgrade their Mac Studios.

If Apple did intend to bifurcate the architecture for its Mac Pro’s on performance - that would require the Mac Pros to have the same release cadence as the Mac Studio going forwards (or the Studio could overtake the Pro) - AND for there to either be an architectural difference between the Pro and the Ultra (eg a 4 chip version / workstation class SoC) - or that the Studio is thermally constrained.

If Apple do have a workstation class SoC - Surely they would want to sell it in larger quantities (ie have it available in the Mac Studio format)? If it did need 50% more heat dissipation - and a 15% larger Mac Studio be sufficient - would this not be a preferred option to the masses that Don’t need the PCIe tower, but do want the most performant Mac available?

I think there’s a reason the Mac Pro chassis is still in production. I think the 2019 design always envisioned the current architecture, with everything on a single large board.

I could be wrong, but I think kits are coming that will allow a Mac Pro M2 Ultra to be upgraded to a Mac Pro M4/M5 Ultra (whatever that turns out to be). It would be a simple board swap, while also switching out the Apple I/O PCIe card and carrying over storage if desired.

I don’t think Apple is interested in what you’re talking about. The community that invests in the Ultra to save time (= money) wants to be able to quickly and easily swap out hardware. Trading in a Mac Studio for a new one is even easier than upgrading your Mac Pro, with both those moves being similar in cost.

 

[crazy dave]

The simplest explanation is Ultra will no longer use Max.

I agree it’s unlikely M5 Max will launch anywhere other than in the 2025 MacBook Pro.

The most likely scenario, given what we think we know, reconciling the identifiers list with Gurman’s timeline predictions and the Hidra rumor, is as follows:

Mac16,9 = Mac Studio M4 Max

The next Ultra is announced at WWDC. It is on N3P and it uses A19/M5 core designs. It does not use Max. It still uses UltraFusion (perhaps 2.0 but not necessarily), but instead of using two Max it is something like Nvidia’s fusion of Grace CPU with Blackwell GPU. The two components are different, and there are configurations available for the Mac Pro that go beyond the Mac Studio:

Mac17,1 = Mac Studio M5 Ultra
Mac17,2 = Mac Pro M5 Ultra

Another feature of the WWDC announcement will be about Apple Intelligence servers, and how the new, improved Ultra fits into that context.

This is my thinking as well. Gurman and Kuo are consistently saying that everything is M4 this generation and the next will be M5 so maybe with the caveat that the Hidra/Ultra chip might be "M4+" or something - enough to justify a new identifying device generation, but not enough to call it M5? If I had to guess, depending on which node it'll be on and given its purported other purpose for Apple (its Apple Intelligence servers), the SOC could have M4 CPU cores but with additional AI stuff, maybe in the GPU. However, just because those two oracles have been consistent, doesn't mean that they're right ... but I'm also trying to make sense of why 17,1 and 17,2 appeared in firmware with the 16-generation of devices BUT no other 17's appears. In other words maybe 17,1 and 17,2 will be unique and the M5/M6 whatever is actually next will be 18s. Of course the rest of the 17s may simply be added later. It's just a little odd.

I think there’s a reason the Mac Pro chassis is still in production. I think the 2019 design always envisioned the current architecture, with everything on a single large board.

I could be wrong, but I think kits are coming that will allow a Mac Pro M2 Ultra to be upgraded to a Mac Pro M4/M5 Ultra (whatever that turns out to be). It would be a simple board swap, while also switching out the Apple I/O PCIe card and carrying over storage if desired.

That would be great (probably still need a DFU restore but that wouldn't be hard). Is there a particular reason you think such a kit is coming or just that it makes sense to you? (to be clear: makes sense to me too, but ... as always ... what makes sense to Apple ...)

I don’t think Apple is interested in what you’re talking about. The community that invests in the Ultra to save time (= money) wants to be able to quickly and easily swap out hardware. Trading in a Mac Studio for a new one is even easier than upgrading your Mac Pro, with both those moves being similar in cost.

 

[wmy5]

Still awaiting the M4 Max die shot.

 

[tenthousandthings]

This is my thinking as well. Gurman and Kuo are consistently saying that everything is M4 this generation and the next will be M5 so maybe with the caveat that the Hidra/Ultra chip might be "M4+" or something - enough to justify a new identifying device generation, but not enough to call it M5? If I had to guess, depending on which node it'll be on and given its purported other purpose for Apple (its Apple Intelligence servers), the SOC could have M4 CPU cores but with additional AI stuff, maybe in the GPU. However, just because those two oracles have been consistent, doesn't mean that they're right ... but I'm also trying to make sense of why 17,1 and 17,2 appeared in firmware with the 16-generation of devices BUT no other 17's appears. In other words maybe 17,1 and 17,2 will be unique and the M5/M6 whatever is actually next will be 18s. Of course the rest of the 17s may simply be added later. It's just a little odd.



That would be great (probably still need a DFU restore but that wouldn't be hard). Is there a particular reason you think such a kit is coming or just that it makes sense to you? (to be clear: makes sense to me too, but ... as always ... what makes sense to Apple ...)

My understanding of the identifiers list is those are the models that will launch with Sequoia in them. So everything above Mac17,2 would not be listed, because the October-launch models will have the next macOS in them, not Sequoia.

On the idea of Mac Pro upgrade kits, that’s just me. It seems like such low-hanging fruit. Like you said, it makes sense. An easy way to build goodwill. But I can’t really say I understand the behemoth that is Apple today and why it does what it does.

 

[crazy dave]

My understanding of the identifiers list is those are the models that will launch with Sequoia in them. So everything above Mac17,2 would not be listed, because the October-launch models will have the next macOS in them, not Sequoia.

So after WWDC we should see more identifiers in the next macOS beta? Makes sense.

On the idea of Mac Pro upgrade kits, that’s just me. It seems like such low-hanging fruit. Like you said, it makes sense. An easy way to build goodwill. But I can’t really say I understand the behemoth that is Apple today and why it does what it does.

Aye.

 

[tenthousandthings]

So after WWDC we should see more identifiers in the next macOS beta? Makes sense.

I think so. The new, upcoming hardware needs to be tested with every build of the macOS that will ship with it.

If I understand correctly, it’s a list of supported identifiers, so it also includes older, known hardware — @Jamie I is just highlighting the portion of the list with new, unknown hardware.

Going forward, I think we’re probably okay as long as the list doesn’t get mainstream coverage.

 

[Xiao_Xi]

Not related to Apple, but this article may be interesting for those who want to know more about fault tolerance in SoCs.

100x Defect Tolerance: How Cerebras Solved the Yield Problem - Cerebras

cerebras.ai

 

[Xiao_Xi]

It appears that Apple is interested in UALink.

Ultra Accelerator Link™ Consortium (UALink™) Welcomes Alibaba, Apple and Synopsys to Board of Directors

Ultra Accelerator Link™ Consortium (UALink™) has announced the expansion of its Board of Directors with the election of Alibaba Cloud Computing Ltd., Apple I...

www.businesswire.com

Does Apple joining the UALink Consortium make it more plausible that it is developing its own data center-class processors?

 

[name99]

It appears that Apple is interested in UALink.

Ultra Accelerator Link™ Consortium (UALink™) Welcomes Alibaba, Apple and Synopsys to Board of Directors

Ultra Accelerator Link™ Consortium (UALink™) has announced the expansion of its Board of Directors with the election of Alibaba Cloud Computing Ltd., Apple I...

www.businesswire.com

Does Apple joining the UALink Consortium make it more plausible that it is developing its own data center-class processors?

That's my interpretation. Like nV acquiring Mellanox, there is some sort of grand plan here that's not clear, but which suggests ambitions beyond the traditional Apple markets.

 

[Xiao_Xi]

Anticipated for release during the first quarter of 2025, the UALink 1.0 Specification will enable up to 200Gbps per lane scale-up connection for up to 1024 accelerators within an AI pod.

According to that press release, the UALink 1.0 specification is coming soon. So, if Apple were to use it, how long will we have to wait to hear about it?

 

[DrWojtek]

That's my interpretation. Like nV acquiring Mellanox, there is some sort of grand plan here that's not clear, but which suggests ambitions beyond the traditional Apple markets.

Wouldn’t surprise me in the slightest if they, in the long run, aim to take on Google and OpenAI. Create an Apple AI engine that is both internet search and Answer Everything, with ”Apple Privacy TM”. There is something to call Apple Intelligence. Not these **** ass news summarizes and Siri junk.

 

[name99]

According to that press release, the UALink 1.0 specification is coming soon. So, if Apple were to use it, how long will we have to wait to hear about it?

This is not a technical question, it's a business strategy question.
Apple may well have working UALink IP in a year. That doesn't mean that in a year they have all the rest of their strategy in place...

 

[Xiao_Xi]

ARM has released their specification for the Chiplet System Architecture.

Arm Chiplet System Architecture Makes New Strides in Accelerating the Evolution of Silicon

First public specification for the Arm Chiplet System Architecture is now available, with over 60 companies engaged.

newsroom.arm.com

Documentation – Arm Developer

developer.arm.com

Unfortunately, I have not found any information on whether Apple was involved or even interested in it.

 

[crazy dave]

ARM has released their specification for the Chiplet System Architecture.

Arm Chiplet System Architecture Makes New Strides in Accelerating the Evolution of Silicon

First public specification for the Arm Chiplet System Architecture is now available, with over 60 companies engaged.

newsroom.arm.com

Documentation – Arm Developer

developer.arm.com

Unfortunately, I have not found any information on whether Apple was involved or even interested in it.

My guess would be no since this is meant to build an ARM ecosystem of chiplets and Apple will probably be developing their own systems of chiplets for their own use but you never know.