I genuinely do not understand keeping the same form factor if they also aren't planning some form of GPU support alongside it. Why have all that extra expansion ability for PCIe otherwise?
1. It is cheaper.
Intel Mini -- M1 Mini exact same case. ( even though M1 mini logic board takes up about 1/2 the space)
Intel MBA -- M1 MBA exact same case.
Intel MBP 13" --- M1
and M2 exact same case .
R&D dollars spent on the chassis for the transitions system. $0.0
If Apple intends to keep selling the 2019 Mac Pro along side the Apple Silicon version for 12-14 months. It is much , much cheaper to just let both products just share a completely 'paid for' case. More volume (economies of scale). etc. That makes the margins higher for both and Apple getting to fill the Scrooge McDuck money pit even deeper.
Even If Apple chopped the number of PCI-e slots from 8 to 5 (or 6) it still would probably be a cheaper path just to reuse the exact same container ( and just leave three or two 'extra' slot fillers on the chassis with no real slot behind them. )
2. If Apple dropped the M2 quad-chiplet version ("Extreme") they are even less able to drop the 2019 Model. Even if they had planned to decommission it at the end of 2023, they are probably going to limp along with it longer now.
Lower priced W6000 series MPX cards or the wished for by some folks a AMD 7900 upgrade card ( doubtful it would be MPX. But something with an official Apple stamp of approval on it. ) would help the old MP 2019 infrastructure limp along into 2024 not looking quite as bad. ( as peak crypto mania pricing on GPUs and years old hardware at PCI-e v3).
3. They don't have to do GPU support for macOS. The two major virtual machine software packages that sit on top of their hypervisor framework have a virtual UEFI environment. Those hosted operating systems could use the GPU. Apple's virtualization framework that is mainly aimed at Linux ... same thing. Could host some AI/ML training workloads that run better in Linux anyway.
For example, there are folks with one or two Nvidia card stuffed into current MP 2019 systems. Do those cards work in macOS? No!. Do those folks BootCamp over to Windows to get work down durning part of their day? Yes.
Add an IOMMU pass through to the VM framework so that card is only assigned to the virtual machine and let it initialize and drive the GPU card that macOS doesn't want to work with. There wouldn't even be much performance overhead at all.
These also don't have to be 'display' GPU cards either. What Apple is deeply missing with "Extreme" model is peak TFLOPs. They need compute more than they need something to drive a monitor with. A 'compute only' GPGPU card would work. An AMD MI210 would work and it has no video out. It isn't a 'dirt cheap' retail card that end users can buy at Newegg/Microcenter/discount retail vendor du-jour. But it would add 10's of TFLOP processing to the system.
4. Having one , and only one, internal storage drive in a workstation is not tractable. Apple even admitted that back in April 2017 as one of the 'didn't work out so well' factors of the Mac Pro 2013. Leaning too hard on Thunderbolt doesn't work . Minimally if want to have 3-5 internal SSDs, then you need some general purpose slots. (e.g, a four x4 M.2 SSD carrier card. Apple overtly showed a picture of one at the Mac Pro 2019 introduction. ).
There are currently over 50 PCI-e cards that currently work with M-series systems via Thunderbolt enclosures. All of those would work in a new Mac Pro with slots. So the question more so is why would Apple walk away from all of that investment that has already been put in? That! would actually be the bigger head scratcher.
There is a subset of those cards that a x16 and x8 PCI-e v3/4 now that do work quite suboptimally in Thunderbolt expansion boxes now. Again it is the "lean to hard on Thunderbolt' problem that Apple has already admitted to.
Two 10GbE network sockets may not be enough. 8K video capture... not really conducive to Thunderbolt. etc. etc.
There can be other accelerators that take more than just the 75W bus power cap to provide a performance boost.
As much as I love my M1 Max Laptop, I'd love to be able to use my eGPU with it for some added power doing noise reduction in Resolve.
again some more 'compute' GPGPU addition might be more likely than a full 'display and compute' driver update.
Maybe they're planning PCIe add in video cards with their own GPU design but even that doesn't make sense given the efficiencies of tying them in with the unified memory and having them on-chip with the CPUs.
All the implicit assumptions in the apps highly tuned to Apple silicon presume unified memory. Add in cards that bust unified memory would not work any more smoothy than 3rd party ones did. If going to put in extra work for display GPU support ... a GPU 'off the shelf' from a 3rd party probably works just as well.
And if the M2 "extreme" got cancelled because low volume buys.... the Apple GPU would largely be in the exact same boat. Apple isn't going to sell it into the Windows market. So the number of potential buyers is relatively small.
Folks can try to tap dance and say that the "Extreme" was too big ... what difference is a relatively small Apple GPU going to bring to the more compute horsepower story? If just as small (or smaller ) as the Max GPU is it any better than an Ultra ? No. It isn't going to be some 4070-4090 'killer'.
Intel took their over decade experience of making an iGPU and rolled out a discrete GPU drivers that were up to their eyeballs in problems. There is no indications that Apple would be immune from that kind of problem. And probably not seeking it either.
Apple's who GPU architecture and software is deeply assumes being able to do a substantively large amount of very localized internal package network passing of data.
