There is the rub, spread the cost of desktop-class Mn Max SoCs by using in the ASi Mac Pro & the Mac Studio, but then the Mac Studio has the unused PCIe lanes just sitting there...?
So just sitting there. As long as not a costly "space waste" it isn't a show stopper. There are extra Secure Element , SSD controller , x4 PCI-e v4 , and Thunderbolt Controller in an Ultra . ( Studio doesn't have more than one SSD, Ethernet Port or 8 Thunderbolt ports. ]
The iMac Pro has a Xeon W-2100 series processor in it provisioning 48 lanes and doesn't use more than 32. About 33% on the sidelines. Also not a show stopper. Cross 50% unused prehaps start of an issue. It is better trade off on the 18 core version (because using lots more of the 'other stuff') with the same socket and logicboard.
The desktop class Mn 'Max sizsed" building block doesn't have t have PCI-e on the die. Disaggregate the PCI-e and desktop I/O ( perhaps some SATA and extra USB ) onto a separate die then don't have to 'duplicate' in every single building block die for CPU and GPU cores. If pull away from the constraints of a full monolithic laptop die then have options.
Look at Intel Gen 14 ( Meteor Lake )
At Hot Chips 34, Intel shared details on its upcoming 2023 Meteor Lake processors that will be composed of disaggregated tiles
www.servethehome.com
Intel has a CPU , Graphics , SoC ( memory , display controller , security , switch-intra-network 'glue' , etc) , I/O ( PCI-e , thunderbolt ) tiles. Apple probably won't disaggregate that extensively ( Apple not going to take as high of a perf/watt hit as Intel is. Nor do they have to tap dance across a wide range of fab providers. ) . However, splitting off the "desktop" I/O would make lots of sense if not going to put any these in a laptop ( or a desktop that the thinness politburo had complete control over ... iMac == iPad on a stick).
so if there was a single x16 PCI-e v4 + two 2x PCI-v4 for Ethernet/or something else , 2 SATA + 2 USB 3 I/O tile chiplet. (maybe push a subset of TB controllers out here also).
Then attach one to a "Ultra" plus two base CPU+GPU+NPU dies . Likewise two to a quad CPU+GPU package set up.
Feed either one or two x16 PCI-e v4 into a PLEX switch chip and regardless provision the 6 slots on the logic board from that ( thereby isolating rest of the logic board form the SoC pacakge differences on PCI-e ). Get thinner PCI-e aggregate bandwidth with higher provisioning with the Utlra and better backhaul on the "Extreme".
When throw the single or dual into the Studio than only one (or zero ) of the desktiop I/O chiplet. So relatively little waste ( other than connector to I/O chiplet ) . [ Studio's could get mildly defective I/O chiplets if not using some part of the chiplet. ]
Similarly could have two building blocks tiles that were largely monolithic. Largely the same except swap out a x16 PCI-e v4 complex for the 4 TB controllers on one version. One/Two tiles with TB for Studios and and a mix (TB and PCIe ) for Mac Pro.
Same thing. "expand" the 1-2 x16 PCI-e v4 feed out into 6 slots.... and basically done with relatively little "waste" of space on the dies.
The major problem with the primarily monolithic laptop dies as building blocks is that they are too chunky for a 'chiplet'. Do not make 'good' chipsets because they 'optimized' too much toward being monolithic and about zero notion of scale past two.
So maybe Apple just goes ahead with bespoke "building block" SoCs just for the ASi Mac Pro...?
That makes no economic sense at all. Shrinking the highly custom SoC onto the Mac Pro unit volume only makes the Mac Pro significantly more expensive. It has already gone though a 100% rise in entry price. Another 20-30% isn't going to help much keep unit volumes higher.
(Possible) PCIe slot specs (and what they might be used for):
- Gen4 - x4 - Apple I/O card (USB/TB/3.5mm headphone jack)
- Gen4 - x4 - third-party audio DSP card
- Gen4 - x8 - third-party 8K video I/O card
- Gen4 - x16 - third-party 64TB M.2 NMVe SSD RAID card
- Gen5 - x16 - ASi MPX GPGPU card
- Gen5 - x16 - ASi MPX GPGPU card
Pretty skeptical that going to get x48 PCI-e v4 worth of backhaul bandwidth out of anything that Apple ships in a Ultra/Extreame SoC. Especially if trying to shoot for maximum RAM capacity they can cobble toegether wtih their semi-custom LPDDR5 RAM packages.
The catch 22 is that the quad tile configuration are going to soak up even more edge die space as the single sided M1 Max UltraFusion connector did. (more UltraFusion connectors on die means less space for RAM and I/O functions) That is going to get in the way of keep up with memory bandwidth. Apple isn't likely going to sacrifice that to because have to keep > 100 GPU cores fed with data. Which means that general I/O bandwidth is highly likely going to get the short end of the stick.
Also why relatively dubious outcome of Apple GPU cores sitting on the other side of a relatively slow PCI-e bus from the SoC is going to see any kind of effort.
There also is quite unlikely going to be TB on a "apple i/o card". The TB controller is in the SoC. Why Apple would 'undo' that is seriouly unmotivated. It is working and essentially needed in the rest of the line up. So Mac Pro probably gets exactly the same thing to save money. the Mac Pro's only issue is that they need to cap the number of TB controllers at around 6 (maybe 7). So building block should reflect that scaling issue.
Apple has already done work on replaceable USB4/TB4 ports and Audio modules on other Macs without cobbling that onto a half sized PCI-e I/O card. No good reason why Mac Pro can't use same 'fix it' modules.
I would expect ( 'old slot' via the MP 2019 )
x8 (old slot 7 )
x8 (old slot 6 )
x16 (old slot 5 )
x16 (old slot 4 )
x4 (old slot 8 . prehaps an electrical x8 if they are feeling generous. )
x16 (old slot 2 )
Same thing as on MP 2019 in that those are all provisioned out of a PLEX switch. So may have x16 electrical but not necessarily getting x16 worth of bandwidth. The "old slot 8" shifted down 'old slot 2" in case have to "throw away" the slot on a 2 wide card. ( if going to a only single wide line up on all slots. Otherwise, if a double wide at the bottom could move that x4 up to the top. ). Similar set of "knobs" on a PCI-e slot manager app to dole out the limited bandwidth the way the user wants it for the set of cards they are using (but some automagically set splits by the Mac Pro if don't bother to set explicitly).
What loosing are the 'direct to CPU package" slots.
If there are six single width slots then perhaps Apple is also a bit less rigid about M.2 blade bay(s) on the logic board. Similar to tribute of putting a couple of SATA ports on the board but not doing anything in the Apple store BTO options with it. ( just Apple acknowledging that folks probably have more than one SSD in these systems on average).
With zero support for GPU cards in macOS there is likely no huge PCI-e backhaul here. Apple has already laid that groundwork.
