ARM Mac Pro - WWDC 2023

[leman]

Anyway, my personal wager for the Mac Pro is either

a) M2 Ultra Max (massively overclocked), or
b) some M3 chip (unlikely)

Regardless of the chip technology, I expect a form factor similar to the 2019 model, with full PCI-e expandability. I also expect the SoC to be mounted on a replaceable MPX module, with ability to use multiple SoCs (either up to two or up to four). Maybe even with expandable shared RAM (either as DDR5 slots on the carrier board or a PCIe RAM module).

And when I say ”wager”, that’s just it, a wager. This kind of architecture makes sense to me from technical perspective and it would IMO work well with Apple Silicon. Also, it’s technically feasible and would make a compelling Mac Pro. Doesn’t mean that I have high confidence that this is what they will go for. We also might see a ”boring” version aka. a new trashcan or a soldered-in SoC. But frankly, if that’s what they do they shouldn’t even bother. Would be bad Pr for Mac. Better postpone the MP a year or multiple until they are ready to offer something good.

 

[Sydde]

The chip went into production that late? That was yesterday. This is why I see next year.

But they still could have had time to build some kind of showcase prototype for WWDC, just to hype it. Maybe they will go bespoke for the Pro, installing whatever card functionality the customer orders directly on the main board for them (allowing easy fit in a 1U box).

 

[senttoschool]

Anyway, my personal wager for the Mac Pro is either

a) M2 Ultra Max (massively overclocked), or
b) some M3 chip (unlikely)

Regardless of the chip technology, I expect a form factor similar to the 2019 model, with full PCI-e expandability. I also expect the SoC to be mounted on a replaceable MPX module, with ability to use multiple SoCs (either up to two or up to four). Maybe even with expandable shared RAM (either as DDR5 slots on the carrier board or a PCIe RAM module).

And when I say ”wager”, that’s just it, a wager. This kind of architecture makes sense to me from technical perspective and it would IMO work well with Apple Silicon. Also, it’s technically feasible and would make a compelling Mac Pro. Doesn’t mean that I have high confidence that this is what they will go for. We also might see a ”boring” version aka. a new trashcan or a soldered-in SoC. But frankly, if that’s what they do they shouldn’t even bother. Would be bad Pr for Mac. Better postpone the MP a year or multiple until they are ready to offer something good.

I can't possibly see them go through this much engineering effort and resources to sell an extremely niche Mac. I don't buy that Apple will create SoC modules, new memory system, and rewrite the OS to work with multiple SoCs. The effort and reward is not there. Yes, you can say that the Mac Pro is symbolic and a halo product, but the truth is, anyone who needs that kind of power in a local machine is likely to need an Nvidia GPU. And everyone else can rent something in the cloud.

The market is not there for something like this.

It's like building a bleeding edge Intel Xeon or AMD Epyc chip, hardware platform, and software, but only selling 10,000 of them each year. It's almost 2024 and HPC has moved to the cloud. It's no longer local. Very few HPC use cases are still local. It's the reason why AMD kept delaying Zen3 Threadripper, and eventually launched as OEM only and it's the reason Zen4 Threadripper won't come out until at least one year after Epyc and Ryzen. Intel has also completely deprioritized workstation chips.

Again, I've been harping on this for years, but in order to justify the kind of engineering a Mac Pro needs, Apple needs to create a cloud service where anyone can rent an M3 Extreme in the cloud. This would open up the market for extremely high end Macs to more than just a one-time $20,000 machine. And maybe then, it might make some economic sense to engineer an OS that can handle multiple Extreme SoCs and rent it out to people in the cloud.

For example, maybe you can rent an M3 Extreme SoC with 512GB of unified memory in the cloud in order to do inference for an AI model that requires more RAM than what a single Nvidia GPU can provide but don't want to spend a million dollars on a whole Nvidia DGX system. Or maybe xCloud might be built to integrate with cloud M3 Extremes services so that build times for large projects can be cut down exponentially. IE. You can work with extremely large software projects sitting in a coffee shop coding on a Macbook Air.

The most likely scenario is Gurman's original report on the Jade 4C-Die, which is 4 Maxes glued together and no memory expansion slots. Over time, each Max would increase in speed and memory capacity and that's the limit of the Mac Pro. No additional engineering is needed beyond the Max and the bridge connecting 4 of them and the OS does not need a rewrite since it still sees just one single SoC.

 

[schappim]

Just as a preview I feel.

Unfortunately I think you're right!

 

[leman]

I can't possibly see them go through this much engineering effort and resources to sell an extremely niche Mac. I don't buy that Apple will create SoC modules, new memory system, and rewrite the OS to work with multiple SoCs.

These are all valid points. I still prefer to remain cautiously optimistic for a number of very simple reasons:

- Apple did go though a lot of engineering effort for a niche product in 2019, even designing custom a expansion module standard; so it's not unprecedented (and the cylinder Mac Pro was an unnecessary amount of effort for a niche product as well)
- the Mac Pro needs to offer compelling performance at the minimum to be an attractive product, otherwise it makes more sense to drop that product class altogether and just stick with Mac Studio
- multiple SoC boards on a switch board is arguably the simplest way to achieve both scalable performance and upgradeability (two important concerns for a modern workstation), and the technology required to do this already exists
- finally, SoC modules would be a very Apple-like solution to the problem (they already shipped this for GPUs) and would be an extremely strong PR move, after all, this is bringing datacenter-like technology to the consumer level

The main challenge in supporting multiple SoCs is probably on the software side. NUMA is hard. Still, it's not unsolvable. Apple already supports multiple GPUs in Metal. And they have extensible CPU thread affinity APIs. The foundation for tackling the problem is there.

It's like building a bleeding edge Intel Xeon or AMD Epyc chip, hardware platform, and software, but only selling 10,000 of them each year.

Which is why I am very sceptical about people suggesting that Apple builds their own dGPUs etc. Designing a custom processing chip just for the Mac Pro is not worth it. Horizontal scalability is the path that allows them to keep the economy of scale where it matters.

It's almost 2024 and HPC has moved to the cloud. It's no longer local. Very few HPC use cases are still local.

One can still make a case for local HPC. There is still a niche for it, one that Apple could potentially exploit to their advantage.

The most likely scenario is Gurman's original report on the Jade 4C-Die, which is 4 Maxes glued together and no memory expansion slots.

So far there is no evidence of a quad-die configuration. Maybe in the future? But even 4 Maxes glued together will hardly make a compelling Mac Pro. Especially if you are locked into the config. Another question: what's cheaper at scale — manufacturing a chip capable of 4-way interconnect; or staying with 2-way interconnect and multiple modules?

 

[bobcomer]

Why would they need to change the chip architecture? All Apple Silicon chips come with PCIe 4 out of the box. The MP just needs more of them.

Via TB. TB is slow compared to a full PCIe bus. You can't even get full SSD speed out of thunderbolt, much less an enclosure with more than one PCIe card in it.

 

[leman]

Via TB. TB is slow compared to a full PCIe bus. You can't even get full SSD speed out of thunderbolt, much less an enclosure with more than one PCIe card in it.

Yes, it’s wired to TB. What is the reason to assume that the chip itself wouldnt be able to drive regular PCIe slots? They do it with the SSD for example.

 

[satcomer]

proBack to original thought I tend to think we will see a nMac when the M3 Ultra is released!

 

[bobcomer]

Yes, it’s wired to TB. What is the reason to assume that the chip itself wouldnt be able to drive regular PCIe slots? They do it with the SSD for example.

One would hope that would be the case! No reason other than it's a fairly big change on the hardware and software side, and will they want to do it.

I didn't know the SSD cards are on a PCIe bus (in a Studio), I thought they had some kind of proprietary link to the controller.

 

[leman]

I didn't know the SSD cards are on a PCIe bus (in a Studio), I thought they had some kind of proprietary link to the controller.

If I am not mistaken, all Apple Silicon SSDs are connected via PCIe (whether there are some Apple-specific extensions there or not, who knows).

 

[Boil]

PCIe does not automatically mean slots; TB4 ports, SSD controller, & Ethernet are all running over PCIe...?

 

[bobcomer]

If I am not mistaken, all Apple Silicon SSDs are connected via PCIe (whether there are some Apple-specific extensions there or not, who knows).

Interesting. Well, it is a standard interface type, that would make things easier.

 

[Ethosik]

Assuming for the sake of the argument that Gurman is correct, what new Mac options would make sense? Not MacBook Pro, that just got refreshed. A 15" MacBook Air has been rumoured for a while, so it's a likely candidate. Maybe even with a new M3 chip, although I personally find it less likely. The other Macs in need of the upgrade are the iMacs and the Mac Studio, but they don't need more than a few minutes of the keynote time.

Mac Pro though? That's a coin toss. A Mac Pro with just an M2 Ultra would make no sense for its target users and Apple would literally get booed. If they go Mac Pro, it would likely feature some impressive or at least fancy-looking tech. I really don't know. I suppose it's a 50-50 from me.

Unless they overclock it and pump up the heat. Mac Pro has the power and cooling to support it.

 

[Macintosh IIcx]

I also expect the SoC to be mounted on a replaceable MPX module, with ability to use multiple SoCs (either up to two or up to four). Maybe even with expandable shared RAM (either as DDR5 slots on the carrier board or a PCIe RAM module).

Hmmm, that sounds like a latency nightmare - especially CPU to CPU on different MPX modules. Clunky and hacky in a very non-Apple way. Compute-only (GPU like) on MPX modules, sure … but a whole SoC? I just don’t see that at all, this being Apple.

Having said that, a M2 Ultra and call-it-a-day for the Mac Pro would be a huge disappointment.

 

[bcortens]

Hmmm, that sounds like a latency nightmare - especially CPU to CPU on different MPX modules. Clunky and hacky in a very non-Apple way. Compute-only (GPU like) on MPX modules, sure … but a whole SoC? I just don’t see that at all, this being Apple.

Having said that, a M2 Ultra and call-it-a-day for the Mac Pro would be a huge disappointment.

I think something like a compute node/accelerator based on an M2 Ultra SoC on an MPX module could make sense.

IIRC the MPX slot is an additional 16 lanes of PCIe so that could theoretically be 32 lanes of PCIe5 which would allow pretty good SoC to SoC bandwidth.

Latency isn't why you would be doing this, you'd only have secondary SoCs on MPX modules if you wanted to do something extremely computationally intense that would benefit from a highly parallel distributed compute architecture. Something like large dataset analysis or very large rendering projects could be distributed across multiple nodes within a single case. Essentially you have a computer cluster in a single case.

 

[mi7chy]

Doubt they're going to piss off pros that want to stay on x64 since there's already ARM Mac Studio. More likely Mac Pro will remain x64 and get refreshed with better CPU, GPU, etc.

 

[Mr Rib]

Since this post has become a santa claus wishlist, here's mine: a bare bones 30" non-OLED PRO monitor (without speakers, a toaster, etc) with a lot of local dimming zones and a Mac Pro with a GPU slot accepting 3rd party graphic cards please Oh and since it's a santa claus letter, I'd like that 30" to be a pen display screen with a nice tilt platform so that I can ditch Wacom Cintiq.

 

[bobcomer]

Doubt they're going to piss off pros that want to stay on x64 since there's already ARM Mac Studio. More likely Mac Pro will remain x64 and get refreshed with better CPU, GPU, etc.

Even I doubt that -- Apple's fastest PC can't remain an x64, it would be a big black eye for AS.

It would be fast though, and very capable. And still too expensive for me to buy.

 

[leman]

Unless they overclock it and pump up the heat. Mac Pro has the power and cooling to support it.

They can, but even overclocked to unreasonable limits M2 Ultra won't outperform an enthusiast-level desktop with a 4070 RTX.

Hmmm, that sounds like a latency nightmare - especially CPU to CPU on different MPX modules. Clunky and hacky in a very non-Apple way. Compute-only (GPU like) on MPX modules, sure … but a whole SoC? I just don’t see that at all, this being Apple.

If Nvidia can do it for their 250 kilo-$ datacenter computer, Apple can also do it for their $20k workstation. Latency will be fine. CXL RAM (via PCI-e) has latency in the ballpark of 200-250ns, that's comparable with M1 Ultra's RAM access latency or the CPU-to-CPU communication in the multi-socket systems. Bandwidth could be more of an issue, but problems that benefit from huge number of cores tend to be parallel enough. Of course, software would need to take the NUMA properties of the system into consideration, but this is true for pretty much any large-core system nowadays.

Of course, I'm just speculating. I don't have the expert knowledge to make a definitive statement whether this kind of architecture would be feasible or not.

Having said that, a M2 Ultra and call-it-a-day for the Mac Pro would be a huge disappointment.

Exactly. And since there is no evidence that M2 Extra is a thing, multi-socket (+heavy overclocking) is pretty much the only way to make a Mac Pro viable.

Of course, they might just surprise us and announce some sort of 3nm super-duper-chip, who knows.

Doubt they're going to piss off pros that want to stay on x64 since there's already ARM Mac Studio. More likely Mac Pro will remain x64 and get refreshed with better CPU, GPU, etc.

Who cares about the pros that want to stay on x86? They don't generate any revenue for Apple going forward. X86 is a deprecated target for macOS. Folks who really want to stay on x86 should jump ship.

 

[Ethosik]

They can, but even overclocked to unreasonable limits M2 Ultra won't outperform an enthusiast-level desktop with a 4070 RTX.

Depends on what is being tested. My M1 Ultra beats out my 13900k and 4090 in some scenarios.

 

[Choco Taco]

It will be announced like AirPower and then never come into existence.

 

[avkills]

If it's not at WWDC, we might as well consider the Mac Pro silently cancelled. The thing was supposed to have been announced last year. It is the last Mac that has not been moved over to Intel yet, and a Mac Studio that is a 1/3 of the price is outperforming it in everyway, in such a small form factor no less. The longer we go, the more the Mac Pro becomes redundant.

Honestly they might as well just rebrand the Mac Studio as the Mac Pro, because it's a Mac Pro at this point in everyway except expandability.

Yeah, but no. Apple Silicon is still weak as **** compared to the overall market when it comes to actual GPU processing power. VRAM? Yes it kick ass because of the unified memory but everything else it sucks wind.

If they do cancel it they better announce they are doing so; so the people who need actual GPU processing power can move on, i.e Windows or Linux.

 

[avkills]

Depends on what is being tested. My M1 Ultra beats out my 13900k and 4090 in some scenarios.

The only scenario I can see the Ultra beating that is in 3D sculpting and texturing because the M1 Ultra *might* (depends on config) have more VRAM available than the 4090. Anything else I am calling ********. Well maybe ProRes handling but "duh", the M1 has dedicated hardware for that.

 

[Ethosik]

The only scenario I can see the Ultra beating that is in 3D sculpting and texturing because the M1 Ultra *might* (depends on config) have more VRAM available than the 4090. Anything else I am calling ********. Well maybe ProRes handling but "duh", the M1 has dedicated hardware for that.

That is one of it. The other area is video editing. It's not just about CPU and GPU. The extra media encoders on the M1 Ultra make exporting very large videos so much faster than my Windows PC.

 

[Algr]

Why does anyone want a Mac Pro? Surely anyone whose work can't be done on a Studio should have abandoned Apple by now. On the PC side, pro users are the superstars, and everyone loves to talk about the products aimed at them. Apple treats pros like crap, shipping products with obvious deficiencies, inflexibility, and insultingly overpriced accessories. And then forgetting they exist for 3-6 years at a time!

It doesn't matter what Apple announces. No one who needs PCI slots or a terabyte of ram is going to trust Apple again.