Pure curiosity before Mac Pro shows up later this year. Let's say a new chip's name for Mac Pro is M2 Extreme.
1. M2 Extreme will be a powerful chip but not in GPU because many workstations already can use more than 1 graphic card or can be used up to 4 or more. Mac Pro 2019 supports up to 4 graphic cards. Which means a new Mac Pro needs up to 4x M2 Extreme chips.
This is the unsubstantiated premise that sends the rest off into the road. The Mac Pro 2007-2012 Models had two CPU packages. Going forward past 2012 they only had one. All computatoinal workloads don't scale across GPUs. Many in the common Mac Pro user workloads space do , but some apps do not. For the ones that need one , 'big horsepower' GPU what Apple is doing is not disconnected from the technological trend line.
So there is no huge requirement that they have more than one SoC in the system. So no 2x , 3x , 4x SoC is not mandated by anything. Apple won't be able to sell every workstation configuration to everybody ... but Apple is not in the everything for everbody business ( e.g., dropped out of selling dual CPU package workstations around a decade ago. and are doing just fine (have profitable enough product to make/ship) in the workstation market.
The M1 Mini backslid on some configuration parameters ( total supported video screens out , max RAM capacity , etc. ) The iMac backlisde on parameters ( max RAM , in most cases supported video screens out , not even a direct 27" product (replaced by headless Studio). ) . None of that hurt Apple mac deskstop sales in any significant fashion. Sure there are some 27" iMac fans 'protesting' by not buying anything , but Apple as replaced them over the intermediate term with folks who do want to buy what is in the line up. The same people don't have to be constant.... Apple just needs the unit sales to grow and make more money. Stuff after that is 'gravy'.
There are also some major limitations here. First, without very deep and substantive changes to the kernel, macOS can't do more than 64 threads. So having more than 50 or so CPU cores really is a huge mismatch to where the kernel is going (and hence where macOS is going). So 4 * 40 CPU cores is way , way , way off in the swamp. Extremely likely not going to happen. The current MP tops out at 28 cores. So if they got to 40 , more powerful cores that would be a
large net increases in CPU 'horsepower'. Trying to spin that as a 'fail' is a huge misdirection.
Second, some folks buy a Mac Pro with a 16-24 core and 580X/5550X and it totally fits their workloads. Not every Mac Pro workload has to be GPU bound. There is an relatively opposite problem/issue here is that some folks are being dragged into paying too much money for GPU cores they may not need. Pricing wise that could become a critical issue also once move far enough away from the entry price ( which is likely another demand lowering hurdle that moving up the MP entry price 100% was/is from 2008-2013 era. )
If you are fine with just one M2 Extreme, then that's fine but how about those people who need more than just one graphic card? Are you gonna say you dont need more than 1 graphic card lol. There are reasons why many workstations support multiple GPU and you never know if Mac Pro becomes a beast in 3D market.
Adding an external GPU might be the option but Apple Silicon take advantages from unified memory and SoC so I doubt it.
When Apple introduced the Ultra SoC and Mac Studio they demonstrated that the Ultra configuration often did better than a MP 2019 configuration that had 16 cores and a W5700. If Apple just doubled that then they'd be doing better than two W5700. So they would cover people who needed two W5700.
They also said that their most commonly sold configuration was a 16 core / W5700 . If there is a simple bell curve distribution on configurations that means half of the Mac Pros sold are at the same or under that 16 core / W5700 configuration. And that the 28 core / 2x W6800 duo is way , way , way out of on the fringe edge of Mac Pro's sold.
So again. Is Apple going to cover everybody? No. Are they going to cover enough to ship a viable product? Probably yes ( presuming the entry price is about the same). The current MP 2019 16c / 48GB RAM / W5700 goes for $8,699 . If the new entry Mac Pro is a M2 Ultra / 64GB that is $5,999 then that is more desired configuration at about $2,700 less. If Apple cuts the price that much they will probably sell lots more. That has a very good chance of offsetting the unit sales of the fringe 28 core / 2x W6800 sales. That will sell more Mac Pros, but might not bring in as much profit. But that will be OK because they'd also have "Extremes" to sell (which are not going to come cheap). Similarly, the 2 x W6800 Duo is $9,499 and add a Afterburner card it totals > $11K .. Similar move where Apple can chop the price $5K and they'll be able to get some people to move who have heavy, multiple concurrent stream 8K ProRes RAW workloads.
2. The maximum RAM size is in question. With only one M2 Extreme, it might have up to 256gb or maybe up to 384gb based on M2's memory. Well, that's way lower than Mac Pro's maximum RAM size which is 1.5TB. Others can support up to 4TB. Even if they can use 4x M2 Extreme, it's barely 1.5TB. Yes, that's only IF they use 4 chips. And Yes, they might be able to add a lot of memory chips along with one M2 Extreme but can they really add those memories around the chip?
Yes, but how many folks are out there. The bigger mismatch there though is not capacity , but lack of ECC. Going into the triple digits of GB RAM without ECC is going to get the system tossed from some workloads also (where data integrity is a top priority).
Again though, where is Apple's average/median MP customer configuration at? If the average is 128GB and they cover 256GB is that really going to be a huge problematical issue for product viability? It is a "bragging rights" problem , but bragging rights doesn't pay R&D costs. The real relevant issue is whether they can sell enough M2 Extreme SoC to make a profit or not. If Apple leaves 15K users on the sidelines, swaps in 5K more new users, and on aggregate makes 5% higher profit , then that is a slam dunk.
If Apple can avoid high capacities then they can avoid having to do ECC memory controller work. Pretty likely Apple would have to either roll ECC out of a couple more members of the Mac product line up to get a much bigger user base to amortize the costs over or they will want to skip it and allocate R&D resources to other parts of the SoC.
Apple has said the Mac Pro is in the single digits of Mac product market. Some folks presume that is 5% or higher. Pretty likely it is far closer to 1% . So if have a feature that 10% of 1% are using then basically rounds to 0%.
The 1.5TB stuff for the MP 2019 had an Intel tax on it. There were 24/28 core optons that were $2-3K less than Apple skipped to drop the more expensive versions into the Mac Pro. Apple slaps another Apple tax on top of the Intel tax. So Apple's margins were substantially higher selling the 1.5TB Xeon W 6200's . It is not clear at all if customer need was the dominate factor there at all. With M-series SoCs Apple makes good margins on the RAM costs, so they don't really need to goose the Max RAM capacity with a tax to make healthier margins. They are already are quite high.
Will they loose some customers at 256GB ? Yes. Do they need to completely solve that in the M2 generation? No. Even more so if willing to sell 4 year old MP 2019 models into 2023 and stop around 2024.
3. Will Mac Pro supports PCIe slots? What about upgradability and expandability? Currently, Apple is very hostile toward both area even for Mac desktops. If Apple wants Mac Pro to be another Mac Studio, what's the point? So far, I'm quite skeptical that Apple is willing to support PCIe slots and other upgradable parts.
This is basically vacuous . There are already over 50 cards that work with Apple M-series . "hostile" would mean that number would be a lot smaller.
Apple is hostile to three category of cards:
1. Overly depended upon early UEFI for proper initialization. There is no UEFI anymore in the new systems. So if the world view revolves around UEFI only then that is a bit 'hostile'. Apple isn't big on including others in the early boot process.
2. Storage cards that don't present as standard NVMe or SATA drives. If there is something funky the card wants to do at boot to present as a coherent drive then they are out. ( historically Apple put in hooks in legacy systems for things like SoftRAID to get included in standard Apple bootloaders. ). Once macOS is running if want to do an initialization there on macOS then that is fine. But that could be view as "hostile".
3. 3rd part GPU drivers. (pragmatically at this point dGPU drivers). There is nothing there.
[ macOS even on Intel hasn't been all that friendly to RDMA , very high end Networking cards. That isn't really 'new'. ]
if Apple was extremely hostile to PCI-e slots then they would be hostile to Thunderbolt. Provisioning those is a substantive feature of Thunderbolt. Apple has been a major proponent of it. It is the primary reason why there are >50 card ready already before a M-series Mac Pro even shows up. That has laid lots of ground work.
The bigger issue with the M-series so far is that Apple has been very keen to reduce the number of PCI-e lanes out of the SoC. They have taken the viewpoint that with PCI-e v4 that x4 lanes can do the same amount of bandwidth as x8 v3 lanes... so less lanes is OK.
That is the bigger issue. So instead of getting x64 top end lanes out, the number is going to dwindle back to x32 or x28 lanes that will be coupled to a PLEX PCI-e switch to be multiplied out. The laptop Mx Series SoC seem have x4 PCI-e v4 amount of lane provisioning. ( the Ultra gets a side-effect allocation of x8 and an over allocation of a TB controller ).
Apple does need more of a desktop focused building block. Doing "everything" with laptop opitmized large dies is going to be a mismatch.
I have no idea how AS Mac Pro will be look like but it shouldn't be another Mac Studio or Mac Pro 2013. If not, then what's the point of having Mac Pro? Mac Pro 2013 was a huge failure anyway.
The Mac Pro 2013 was being bought by the pallet right up until the last day it was sold. ( It wasn't the classic Mac Pro customers. It was places like MacStadium/Colocation customers that used it as a "Mac via a cloud service" system. It is/was a good fit for that in terms of compute density. ) .
The huge next Mac Pro issue likely does not hinge on the SoC as much as whether Apple wants to make it a literal desktop system ( has to fit on a desktop as opposed to beneath or beside it). Apple seems to have a "must fit in a 7" x 7" square" requirement for literal desktops. Mini, taller Mini (Studio) , foot of the iMac/XDR/Studio display stand . The Mac Pro saddled with that largely is what gets you the MP 2013.
Some rumors have been that Apple is doing something "half sized". At half the volume , there is a decent chance that literal desktop restriction is not in play. It will have lower shipping costs, but you'll still be able to add those $100/wheel option.
There was also a rumor of a one-slot-wonder that was an M1 SoC that probably is likley not to see light of day. Whether that is a prototype or Apple doing the barest minimum slot allocation, it is still indicative they are not trying to diretly revivie the MP 2013. That doesn't make any rational sense. The evolution for the MP 2013 is largely
MP 2013 --> iMac Pro ---> Mac Studio.
They already have something that largely matches that role. Why would they want another direct overlap with Mac Studio? Pretty likely though that the M2 "Utlra class" SoC for the Studio could/would overlap with the entry Mac Pro . Apple likely has to get the SoC run rates higher to fork off of just using MBP 14/16" SoC in a mash-up as the basic building block.
We'll see after Apple announces a whole new Mac Pro...
I suspect one of the other flaws here is that the expectation that the M2 Utlra is going to look almost exactly like the M1 Ultra. Using a the "Max"-class die as a chiplet is somewhat goofy. It makes sense used as a single die in a laptop, but once start moving it into the multiple chip packages it doesn't really scale. There is zero reason why M2 , M3 , M4 has to rigidly stick to the same size dies used in the same size patterns once new packaging technology and fab processes arrive.
There are two paths that would give a wider balance of Mac Pro (and Mac Studio) configurations.
1. Go to two more desktop oriented building block largish dies.
20 CPU cores + NPU cores + Pro class GPU cluster + x16 PCI-e v4 provisioning ( or x8 PCI-e v4 and 1-2 block TB controllers) + 2+ UltraFusion connectors.
bigger than Max GPU ( e.g., 40 GPU) cluster + display controllers + NPU cores + PresRes de/encode units , TB controllers + 2+ UltraFusion connectors
So would be able to do a variety of quads.
20 CPU + 40 GPU + 40 GPU + 40 GPU ( customers that are GPU core workload limited )
20 CPU + 20 CPU + 20 CPU + 40 GPU ( customers that are CPU core workload limited )
20 CPU + 20 CPU + 40 GPU + 40 GPU ( customers that need balance mix of both )
Could probably sell some of the CPU weighted dies as an alternative "Max" in some MBP 14/16" configurations. Or another desktop ( if bring back iMac large screen ). Is Apple going to create 7900/4090 'killer' GPUs that way? Nope. Will it be more completive that using laptop optimized building blocks? Yes. (going to cover more future upper mid-range. And that future upper midrange is going to be be 'good enough' for a wider set of folks. )
2. Even more modular. CPU + GPU + I/O in their own chiplets and fuse them together with ultra fusion. Could make "Max" class SoC out of chiplets rather than being medium-large monolithic die. That would keep the "Max"-class SoC sales volume for amortiziation of R&D and give a wider set of options for multiple CPU/GPU die packages .
Also breaks out the I/O into a separate chiplet so that can have a 'laptop' almost not PCI-e lane ones and 'desktop' more PCI-e lane one if that makes more sense.
If Apple went TSMC N3 and saved some power and then 'lost' small amount of power efficiency by changing Max monolithic die to chiplets then it could be an even 'wash' trade-off. It wouldn't be a battery win for the MBP's but it wouldn't be a battery loss either.
The other part is largely just software. If they carve out a Metal compute only subclass to the PCI-e driver framework classes then can weave back in GPGPU compute GPUs. I doubt they'd be provisioned directly in the Mac Pro case. but a PCI-e exapnader box was used by some back in Mac Pro 2009-2012 days to do 3-4 GPUs. Worked before ... could work again if just have a free x16 PCI-e v4 slot. ( v5 would be better but I don't expect Apple to go there in M2 generation. ) .
