Apple's Mac Pro Refresh: What We Know So Far

[iPadified]

GPU is not automatically superior in compute compared to CPU. For instance, CPU render software are superior in terms of functions than the corresponding GPU based. The reason GPU got popular was because they where cheap compared to CPU. Go back 10 years and look at the GPU/CPU landscape and you understand. If you had gotten 40 core CPU cores 10 years ago at a competitive price, GPU would not be nearly as large market as it is today. . GPU has by tradition provided cheap compute capabilities while CPU has not. Unclear why.

Still I am curious how Apple is going to compete in the general Mac Pro market and not only as a video cutting machine (which MP 2013 also was).

2013 MBP was before its time. Thunderbolt 1.0, no breakout boxes, no good upgrade path from Intel/AMD/NVIDIA so it died on the vine. Today we have USB-C as standard, nice breakout PCI boxes and high performing chips at low wattages. Quite a different starting point.

 

[AlexMaximus]

If the Apple Silicon version of a Mac Pro will be honest, it will need to be modular and allow third party cards for special interest groups such as Audio folks, as well as expansion options for sonnet, owc, and many other solutions. I have the feeling that this will not happen. In order to do that, Apple would need to talk to any of those third parties and let them have unique third party hardware solutions for their ASMP. I don’t think Apple is willing to take on that hassle. However Apple is in a double bind for exactly this reason. If the Intel MP disappears, they will be forced to go that route. And this is exactly the reason why the real modular intel MP will stay even longer. The new side dish Apple SOC MP will there for be NOT an intel MP replacement. I feel it will be a fork in the road for that product segment, meaning the Apple SOC will not be an intel replacement. I just hope it will not be another excuse again for an all in one product that cant be really expanded, just as the trash can debacle.

 

[theluggage]

Adding more cores only does so much depending on what is being run on the computer, how the OS is managing the processes, and how the chip is managing the cores.

...true, but true for the competition as well - the higher-end Core i, Xeon-W, Ryzen etc. models are mainly differentiated by having more-of-the-same cores, and that partly applies to high-end GPUs as well, so it's not as if Apple are going out on a limb with multi-core. Squeezing an extra 20% single-thread performance out of a core is hard c.f. doubling the number of cores. Optimising software to exploit those multiple cores can also be hard but the payoff is that using double the number of cores can nearly double the performance on suitable tasks. Audio, video, scientific computing etc. workloads tend to be amenable to multi-core processing and represent a huge market - I'm sure there's a 'niche' for super-powerful individual cores but it probably is a niche.

A few people here still talking about M-series power per watt being a consideration. Sad truth is away from mobile systems like laptops and iPads/iPhones literally no one cares about that.

I agree that power efficiency is nothing like the night-and-day advantage you see in mobile/laptop applications.

However, it still is a consideration - look at how much space and expensive engineering there is in the Mac Pro just to handle power supply and cooling, mainly for the CPU and graphics cards. Electricity bills for data centres can be a big deal too - you pay for the electricity to run the computers, the computers turn the electricity into heat, then you pay again for the electricity to run the air conditioning to get rid of that heat...

Problem is the 2019 Mac Pro concept sits pretty much at the point where power consumption is least important: a stand-alone workstation that has to be big enough to physically hold 8 full-size PCIe cards on one hand, and isn't really suitable for data centre or high-density computing use on the other. A like-for-like replacement of the current Mac Pro never going to be the best showcase for Apple Silicon.

What we know so far is M-series devices have a very walled-off architecture with virtually no opportunity to upgrade literally anything, and even if Apple put a gate in the wall, currently nothing out there is compatible. All that is going to have to change.

The upgrade path for the M1 range is to add more M1s - either on-chip (M1 Pro/Max) on-package (the rumoured x2 and x4 versions) or just by adding more packages. Also, unlike some of Apple's Intel offerings, ir's not just locked down for the sake of it, it's reaping performance benefits by combining everything in a single package.

On top of that, the x2 and x4 are likely to have a shedload of TB4 ports.

Plus, the M1 does appear to have PCIe (the 24" iMac uses a PCIe-to-ethernet chip, for starters) which presumably means the number of available lanes gets scaled up with the larger versions. I doubt that it can match the Xeon-W for PCIe lanes but then, if PCIe GPUs aren't going to be supported, it won't be using 16 or more lanes just for a GPU. So it is down to Apple whether or not to have PCIe slots.

However, making a bad 2019 Mac Pro clone probably wouldn't be a good use of Apple Silicon. Instead, they could make a good trashcan replacement (ideal for all those YouTubers producing prosumer videos) and/or some sort of blade system that could take multiple M1-based compute modules. Meanwhile - unlike in 2012 - they've got a fairly up-to-date 'conventional' Intel Mac Pro for the customers that really need it.

Only Apple know how well the 2019 Mac Pro has been selling, and how much cost/effort is justified in keeping it going. The cost of supporting the Intel Mac Pro for another few years, while simultaneously exploring more radical options using multiple M1s could be a better option than sinking a lot of money into trying to make an Apple Silicon-based Xeon killer chip that would only ever have a tiny market.

 

[AndiG]

Fair enough, but if your ML workflow is based on CUDA, you have not been using a Mac / macOS for a decade or longer so what a Mac Pro has for ML support is irrelevant to you.

Who it will be relevant to is those using Apple's ML language and tools (Core ML / Create ML).

Thats why a lot of my worflow is based on Linux or even Windows. The point about CoreML ist that CorML only exists in Apple universe. So for Apps- CoreML may be fine.

Remains to be seen how many libraries and tools will support Apple silicon.

 

[CWallace]

Remains to be seen how many libraries and tools will support Apple silicon.

Hopefully it will prove over time to be powerful enough to be an incentive for other libraries and tools to support it, but as you noted, we shall see.

 

[kaisdaddy]

Would the upcoming Pro models still use M1 based chips or would it make more sense to jump straight to whatever the next version is? At this point it has been a minute since the M1 was released and by the time the new Mac Pro ships it will be quite long in the tooth.

 

[CWallace]

Would the upcoming Pro models still use M1 based chips or would it make more sense to jump straight to whatever the next version is? At this point it has been a minute since the M1 was released and by the time the new Mac Pro ships it will be quite long in the tooth.

The presumption is they will use multi-SoC M1 Max (2x and 4x) based on rumors / claims / reports.

 

[CWallace]

I get the "once burned, twice shy" feelings about the first-generation Apple Silicon Mac Pro (MacPro8,1), but I honestly do not see Apple intentionally handicapping the machine in terms of capabilities just because they have a "vision" that every Mac must be a sealed box.

Apple specifically held an apology meeting in 2017 to address the problems with the 2013 Mac Pro - and Apple knew they went beyond just not updating the CPUs and GPUs due to the "thermal corner". They knew removing all the internal expansion and depending solely on external Thunderbolt devices was a mistake that hobbled the capabilities and performance. This is why they strongly hinted that MacPro 7,1 would be a return to the foundation of Mac Pro5,1 and not just MacPro6,1 with better cooling.

So as some have noted, it makes absolutely no sense for Apple to go back to the "sealed box" of MacPro6,1. Apple (eventually) recognized it was a mistake, publicly apologized for it, and then engaged the people whom MP5,1 worked for and MP6,1 did not to develop MP7,1 so that it did work for them again.

We know Apple is not going to support Intel for just one model, so MP7,1 will be done within the next five years or so and the rumored CPU/systemboard refresh will likely be the only one it gets. After that, it is only Apple Silicon and if the ASi Mac Pro does not offer much of the flexibility of the Intel Mac Pro, then there is no reason to release it.

Now all the above being said, MP8,1 will likely not be a clone of the Intel model with an ASi SoC instead of an Intel Xeon (as MacPro1,1 was of the PowerMac G5). And therefore it will likely not have the full expansion flexibility of the 2019 model. But I expect it will have PCIe slots and I believe at least one of them will be full-length (not for video cards, but for other types of full-length cards like audio).

MacPro8,1 will show that Apple Silicon has a future in the macOS workstation market, even if it may not match MacPro7,1 in terms of CPU/GPU benchmarking or RAM capacity or internal expansion ability (hopefully whatever SoC MacPro9,1 uses will have closed the gap).

 

[George Dawes]

I'd love the new mini mac pro one , after my experiences with overheating issues on many mac minis , this is definitely the way i'd go forwards

 

[maxoakland]

Gonna be insane amounts of power per watt, however, how does Apple address the GPU given it’s all integrated? On a laptop, makes sense, but in a desktop?

Hopefully it’s not all integrated because that would be an issue. Maybe it’s something they need to keep working on which is why there’s still going to be an Intel version

 

[Mr.PT]

I get the "once burned, twice shy" feelings about the first-generation Apple Silicon Mac Pro (MacPro8,1), but I honestly do not see Apple intentionally handicapping the machine in terms of capabilities just because they have a "vision" that every Mac must be a sealed box.

Apple specifically held an apology meeting in 2017 to address the problems with the 2013 Mac Pro - and Apple knew they went beyond just not updating the CPUs and GPUs due to the "thermal corner". They knew removing all the internal expansion and depending solely on external Thunderbolt devices was a mistake that hobbled the capabilities and performance. This is why they strongly hinted that MacPro 7,1 would be a return to the foundation of Mac Pro5,1 and not just MacPro6,1 with better cooling.

True, although a “mini” MP could follow that “seal box” path: a step up, performance wise, to MBP, iMac and Mac Mini within the same conceptual design. This would make more justifiable having two MP offers which wouldn’t compete.

So as some have noted, it makes absolutely no sense for Apple to go back to the "sealed box" of MacPro6,1. Apple (eventually) recognized it was a mistake, publicly apologized for it, and then engaged the people whom MP5,1 worked for and MP6,1 did not to develop MP7,1 so that it did work for them again.

We know Apple is not going to support Intel for just one model, so MP7,1 will be done within the next five years or so and the rumored CPU/systemboard refresh will likely be the only one it gets. After that, it is only Apple Silicon and if the ASi Mac Pro does not offer much of the flexibility of the Intel Mac Pro, then there is no reason to release it.

Now all the above being said, MP8,1 will likely not be a clone of the Intel model with an ASi SoC instead of an Intel Xeon (as MacPro1,1 was of the PowerMac G5). And therefore it will likely not have the full expansion flexibility of the 2019 model. But I expect it will have PCIe slots and I believe at least one of them will be full-length (not for video cards, but for other types of full-length cards like audio).

Although I’m curiously waiting for other form of expansion and upgrade that can “sustain” the advantage of unified architecture, the coming MP, will likely offer at least 2 PCIe (v5?) x 16. But the “capping” inability to support 3rd party graphics, namely the CUDA Nvidia, will keep the niche MP an UltraNiche machine preventing a broader adoption by a significant portion of it’s potencial professional users.

MacPro8,1 will show that Apple Silicon has a future in the macOS workstation market, even if it may not match MacPro7,1 in terms of CPU/GPU benchmarking or RAM capacity or internal expansion ability (hopefully whatever SoC MacPro9,1 uses will have closed the gap).

New MacPro SoC must at least match the Intel benchmark, otherwise “there’s no reason to release it”. Besides the embarrassment, it would send a deadly wrong message about AS capability.

 

[Mr.PT]

Thats why a lot of my worflow is based on Linux or even Windows. The point about CoreML ist that CorML only exists in Apple universe. So for Apps- CoreML may be fine.

Remains to be seen how many libraries and tools will support Apple silicon.

Hopefully it will prove over time to be powerful enough to be an incentive for other libraries and tools to support it, but as you noted, we shall see.

Almost 18 months into transition and AS optimization hasn’t been impressive. Apple really needs to push this and start to engage more proactively in the 3rd party software AS adoption. The Blender initiative may be a hopeful sign. As it is there are already some “overpowered” machines which are “stalling” or “capped” by lacking software capability.
The usual passive attitude of throwing hardware coupled with some “generic” benchmarks, is not enough to convince “deeply embedded in pro software” customers. So either Apple reaches out to 3rd parties or stars to produce offers that simultaneously caters those professional needs while demonstrate superior performance of AS. The last is very unlikely although I’d love a Final Cut’s “twin brother” for 3D graphics. Instant buy if supported/coupled with hardware ray tracing.

 

[Boil]

With TB4/USB4 and not needing discrete GPUs, do we really need PCIe slots...? ;^p

Six TB4/USB4 (USB-C) ports, four USB 3.2 (USB-A) ports, and dual 10Gb Ethernet (RJ45) ports on the ASi Mac Pro Cube...

Imagination Tech PowerVR Photon-based hardware ray tracing with M3...?

An Apple DCC software suite would be, well, sweet...!

 

[seek3r]

I'm not sure that there is some all-encompassing hardware standard for ARM-based systems to compare to the 'x86 PC compatible' de-facto standard

There is, and the fact that you have that embedded in your comment kinda invalidates the rest

The M1 Chips are ARMv8.4-A with some added extensions. Apple has their own extensions in Intel machines too via the T2, etc. I’m fairly confident that those can be reverse engineered. The biggest problem is graphics drivers, but that’s more of a problem in the “run other things on Apple hardware” direction than the other way around. Linux is already running bare metal minus accelerated graphics on M1s, plenty of people are exploring the AS chips. By the time Intel machines are unsupported entirely there will be a wealth of work done in getting both other OSes running on AS chips and getting MacOS running on other ARM chips

 

[theluggage]

The M1 Chips are ARMv8.4-A with some added extensions.

That's just the instruction set. There's a lot more to a working computer system than that - firmware/bootloader, graphics, disc controllers...

The biggest problem is graphics drivers, but that’s more of a problem in the “run other things on Apple hardware” direction than the other way around.

We're talking about running MacOS on non-Apple-Silicon ARM chips. You'd need MacOS drivers for the various GPUs found in other ARM systems. AFAIK MacOS for Apple Silicon doesn't support anything other than Apple Silicon GPUs. Even the x86 version only supports AMD GPUs since Apple fell out with NVIDIA.

Linux is already running bare metal minus accelerated graphics on M1s

That's a big "minus". No accelerated graphics = pretty much useless for modern desktop apps.

Anyway, running Linux for ARM on Apple Silicon is a totally different problem to getting MacOs for Apple Silicon running on non-Apple Silicon. Linux is highly portable and even the ARM versions of Linux are written to run on the widest variety of ARM processors - and you can get the source to patch. Apple have no reason to make MacOS for Apple Silicon compatible with anything but Apple Silicon and you don't get to see the source of most of it.

Even if you do get MacOS for ARM running on a non-Apple ARM chip it will lack a lot of Apple Silicon features used by MacOS - neural engine, hardware acceleration for Apple codecs, security features, optimisation for Rosetta 2, unified RAM (maybe), a GPU designed for Metal.

Probably you could get the Darwin kernel and some basic functionality working on generic ARM. Maybe it would even run the MacOS desktop with a laundry-list of "ifs" and "buts" and sub-Apple Silicon performance but the appeal of Intel Hackintoshes is that they're not only very close to being functionally complete and completely usable but they're often faster than any hardware being offered by Apple at a comparative price.

 

[spetznatz]

"As for the smaller ‌Mac Pro‌, it's been described as looking similar to the existing ‌Mac Pro‌, but with a more compact enclosure that's half the size."

You guys may wish to brush up on your math. The image you've got shows a Mac Pro alongside something a quarter its size.

I once nearly came to blows with some eejit in a copy shop who was berating the employee for typing in a reduction factor of 71% when the eejit wanted his image “half the size”. Seventy-one percent is basically 1/sqrt(2)…

 

[Snahbah]

The presumption is they will use multi-SoC M1 Max (2x and 4x) based on rumors / claims / reports.

OMG! This.

What if Apple create a desktop version of AS? One that can accept PCI cards.

I've been at pains trying to figure out how Apple will deal with discrete graphics (I am a 3D pro and need multiple GPUs). I was thinking that Apple might introduce their own at one point, but you may have something here. Workstations (no matter what people say) can use as much power as they need if they are cooled properly. Compared to my day-rate, electricity is negligible and as long as my legs don't melt I'd want as much power as possible and an enclosure I can ram as many GPUs in as can fit.

 

[Realityck]

OMG! This.

What if Apple create a desktop version of AS? One that can accept PCI cards.

I've been at pains trying to figure out how Apple will deal with discrete graphics (I am a 3D pro and need multiple GPUs). I was thinking that Apple might introduce their own at one point, but you may have something here. Workstations (no matter what people say) can use as much power as they need if they are cooled properly. Compared to my day-rate, electricity is negligible and as long as my legs don't melt I'd want as much power as possible and an enclosure I can ram as many GPUs in as can fit.

Years ago you had industrial STEs where many card slots were available to insert/populate various computer processors and specialized functions cards, very common to see additional processors and communication CCAs added. Apple should be at such a advanced stage that you could use something similar but way more cutting edge technical where you could install additional processors on a card along with other cards.

 

[Mr.PT]

OMG! This.

What if Apple create a desktop version of AS? One that can accept PCI cards.

I've been at pains trying to figure out how Apple will deal with discrete graphics (I am a 3D pro and need multiple GPUs). I was thinking that Apple might introduce their own at one point, but you may have something here. Workstations (no matter what people say) can use as much power as they need if they are cooled properly. Compared to my day-rate, electricity is negligible and as long as my legs don't melt I'd want as much power as possible and an enclosure I can ram as many GPUs in as can fit.

Apple needs to come up with something that can match CUDA and hardware ray tracing.

Years ago you had industrial STEs where many card slots were available to insert/populate various computer processors and specialized functions cards, very common to see additional processors and communication CCAs added. Apple should be at such a advanced stage that you could use something similar but way more cutting edge technical where you could install additional processors on a card along with other cards.

And that could mean “indiscreet” computing architecture: a “neural engine” processing unit that would assign computing power according to momentum workloads, regardless of task typology. So p.ex. one M1 Max Soc would be “indiscrete” 42cores rather that 10+32.😎
There’s still the need to come up with a connection other than PCI, that can offer bandwidth similar to “internal” 400GB/s.

 

[Snahbah]

not needing discrete GPUs, do we really need PCIe slots...?

Sorry, but you are ill informed here. 3D/VFX pros need multiple GPUs not just a single powerful GPU. A 7,1 ladened with GPUs will destroy a AS Ultra and the MP 7,1 still stands peerless in the Mac line up for my sector.

 

[gabrieljsmith1]

Nice post thanks Boil!

Y'all even bothering to proofread the articles before publishing them anymore...?!? ;^p

• LPDDR5X RAM
• Pin-compatible with LPDDR5 RAM
• 20% less power usage
• 33% faster
• 64GB maximum chip density
• Single M1 Max SoC = 256GB RAM / 500GB/s UMA
• Dual M1 Max SoCs = 512GB RAM / 1TB/s UMA
• Quad M1 Max SoCs = 1TB RAM / 2TB/s UMA

View attachment 1964122

 

[CWallace]

So with WWDC on Monday, we will hopefully see the Apple Silicon Mac Pro announced (not released).

So looks like Apple did at least develop an "M1 Extreme" which is effectively four M1 Max using the UltraFusion Chip Interconnect. Code-named "Redfern", this is the 40 CPU core (32P/8E) + 128 GPU core model first leaked by Mark Gurman as "Jade4C-Die".

Now even though Apple did develop it, there is some belief that it might never see the light of day because Apple said M1 Ultra was the last M1 SoC, so there can be no "M1 Extreme".

Some believe it will be a new generation SoC with 40/128 cores based on M2/A15 (so Avalanche/Blizzard). Sami Fathi said it won't be called "M" at all, but instead something new like X1 or Z1.

 

[deconstruct60]

So with WWDC on Monday, we will hopefully see the Apple Silicon Mac Pro announced (not released).

So looks like Apple did at least develop an "M1 Extreme" which is effectively four M1 Max using the UltraFusion Chip Interconnect. Code-named "Redfern", this is the 40 CPU core (32P/8E) + 128 GPU core model first leaked by Mark Gurman as "Jade4C-Die".

Now even though Apple did develop it, there is some belief that it might never see the light of day because Apple said M1 Ultra was the last M1 SoC, so there can be no "M1 Extreme".

One contributing reason Apple really doesn't do roadmaps is because even when the do tell stuff in advance ( e.g., "This is the last M1" ) folks go off and believe whatever they want to believe anyway. So why bother.

The M1 Max doesn't have full interrupt stack support for four dies (as the Pro die didn't have full support to 2 dies). UltraFusion is a bit oversized pin count ( 10,000 paths) for just two dies , but four M1 Max dies with "next to no NUMA" access and provision some PCI-e and/or external DDR5 slots seems doubtful. Yes, Apple photoshopped out he UltraFusion connector on the Max and sandbagged on what the Max was. But enough I/O to get back to a 4+ die set up would be a stretch to submarine there.

Apple did "Jade" and Jade2C" with just one die implementation. Stretching that out into one die covering all three (add Jade4C) is a bit of a leap. The Max die just isn't laid out that way.

A quad die set up probably would require more die edge space sacrifice than what the M1 Max die is doing. Quad dies is going to stretch the memory timings. ( e.g,. can't dense pack along the edge that has the memory packages. ) . The TSMC InFO_LSI packaging tech barely stretches to do the two package Ultra. It won't stretch to four dies at 5nm scale. Possibly can put two InFO_LSI packages on top of another substrate , but the UltraFusion connections between the two pairs won't be at the same small pad , lower power connectivity as between the two dies. TSMC CoWos-LSI packaging could be used to scale up to that big of a footprint ( 4 * 432mm^2 = 1,736 mm^2 ), but Max die probably isn't tuned for that.

What Apple needs is a die that can be better shared across the "Utlra" and "Extreme" set ups. Either a monolithic Ultra (and Extreme is just a "duo" of that on CoWos-LSI) or a set of different dies (with scalable I/O. ) that are not optimized for a MBP 14" chassis. ( could be shared across upper Studio , iMac Pro (if comes back) , and partial Mac Pro for scale. )

Some believe it will be a new generation SoC with 40/128 cores based on M2/A15 (so Avalanche/Blizzard). Sami Fathi said it won't be called "M" at all, but instead something new like X1 or Z1.

The "won't be called "M" at all" seems likely to be some kind of death spiral path. If completely decouple the Mac Pro SoC from the rest of the Mac SoCs there is relatively no volume for long term support. Pretty likely to so something priced at least as high as the current Mac Pro ( which has already burned off part of the legacy Mac Pro customer base. ). Minimally some soldered on RAM (if not all RAM) and integrated GPU will likely burn off incrementally more of the legacy customer base.

M2 is at least as decently likely hooked to A16 as it is to A15 at this point. M2 is probably off track for release but if it was originally targeted to Feb-April 2022 then coupling it to A16 would have made just as much sense as the A15. if the M2 was originally targeted for June 2022 for the two year anniversary dog-and-pony show then coupling the M2 to the A15 doesn't make much sense at all.

Apple doesn't "have to" use every A-series generation core in a M-series chip. The Watch is on A13/ 7nm tech. The iPad Pro A--X sequence skipped non node shrink A-series cores for several generations . Even inside the M-series when the "extreme"/"Ultra double duo" ships out every Mac SoC probably isn't going to catch every 'plain' Mx core produced ( The SoCs that are shared across the iPad line up) .

Apple's "hit the snooze button" , Rip-van-Winkle pace on the Mac Pro has been consistent for over a decade. Periodically, Apple wakes up, does something and then goes back to 'snooze'. Effectively that gives the Mac Pro models longer support because they aren't being superseded as quickly. Highly forked off SoC for Mac Pro would only likely reinforce that more.

Unless Apple is creating a "desktop series" and deoupling the Studio non entry iMac/Mini from the rest then maybe not a dead ender. But if it is only for the Mac Pro (and a sub 100-200K run rate ) then that's probably bad long term.

 

[MikeELL]

One contributing reason Apple really doesn't do roadmaps is because even when the do tell stuff in advance ( e.g., "This is the last M1" ) folks go off and believe whatever they want to believe anyway. So why bother.

The M1 Max doesn't have full interrupt stack support for four dies (as the Pro die didn't have full support to 2 dies). UltraFusion is a bit oversized pin count ( 10,000 paths) for just two dies , but four M1 Max dies with "next to no NUMA" access and provision some PCI-e and/or external DDR5 slots seems doubtful. Yes, Apple photoshopped out he UltraFusion connector on the Max and sandbagged on what the Max was. But enough I/O to get back to a 4+ die set up would be a stretch to submarine there.

With Mac Studio it seemed to me that the UltraFusion Interconnect was used in that case to combine 2x M1 Maxes into the Ultra, but the actual power of UFI is giving customers ability to customise what kind of compute power the machine will be focussed on, even within the SoC architecture, rather continuing to double everything with "extreme".

So say every Mac Pro was build to order with an M1 Ultra as the base, but the UFI is extensible in a way that allows you to fill up to (say) 8 extra slots with chips that are some combination of dedicated CPU, GPU, RAM or neural engine modules, all connected to the M1 Ultra via UFI. If the performance is there, there might not be too much grumbling about lack of support for third party components.

 

[Joe The Dragon]

So say every Mac Pro was build to order with an M1 Ultra as the base, but the UFI is extensible in a way that allows you to fill up to (say) 8 extra slots with chips that are some combination of dedicated CPU, GPU, RAM or neural engine modules, all connected to the M1 Ultra via UFI. If the performance is there, there might not be too much grumbling about lack of support for third party components.

still need pci-e for pci-e cards and m.2 slots are nice vs apples storage that is one disk only at big mark up.