Chiplets are unlikely to 'save' the Mac Pro in a core count "war" with AMD ( and/or Intel ) if Apple stays 100% grounded to it being a SoC in scope ( some minimal GPU + NPU + etc etc. apple custom fixed function logic ) as first priority more so than raw focus on application host processor core count .
Apple's constant rigging the bell on "Unified Memory" is insanely create points to substantially different design priorities. More likely to get a multi chip module build of a CPU chip with a GPU chip sharing a common memory pool than something that was more monomaniacally focused. Apple doing that just to get better yields on bigger blocks of the chip rather than even more expensive monolithic die. Or using RAM system cache chiplets to free up space on the consolidated logic chips for logic to get to a much bigger system pool.
Apple probably isn't going to get into a core count war' ( past 64 cores at all.) Pretty likely will stay in similar count zone that the Mac is currently in. ( 32 cores range).
If the nominal Apple Silicon SoC could feed the nominal USD4/TB4 ( or incremental future of that line) ports with video then the entry level Mac Pro ( and rack model) could free up another PCI-e slot for other stuff. The apps that were optimized for uniform memory GPU workload would run "OK" and the workload could cover something that wasn't so GPU heavy ( audio work. virtual host server work . etc. )
In non Mac Pro SoC perhaps the cell modem is on a chiplet so don't have to completely coupled those but lower the power consumption of interconnect because so close ( short). Save an iMac or MBP 16" SoC where have higher core target but budget individual die sizes.
I know they used the word SoC constantly during the keynote but I really don't think they'll use an SoC in the manner you're assuming. EPYC is technically an SoC. It has no seperate chipset for USB, SATA and all that while Threadripper and Ryzen do.
I think Apple will go with that approach. There will not be a GPU accessible to users inside the "SoC" that Apple provides for the Mac Pro is my opinion. But it will contain everything else like all the PCIe controllers, Thunderbolt, USB. Possibly all the networking too (WiFi, Bluetooth, 10Gb Ethernet possibly).
And that central SoC die would then be surrounded by ARM execution units as chiplets. Similar to EPYC like in this photo below.
In the above photo the big central die is the I/O die which does all I mentioned before but also has the memory interface. The smaller 4 dies on either side are the 8 core chiplets which talk only to the central die for all their memory access and other things.
I think Apple will likely do something similar to this for all Macs. You may have noticed in the keynote they said they would be making brand new ARM chips specifically for Macs. Not using for instance the iPad or iPhone CPU's in Macs.
It is my feeling and opinion that this will mean the MacBook Air will likely get an I/O die + single chiplet. That chiplet will contain 8 cores but have 4 disabled. The MacBook Pro 13" will probably get the same thing but 6 cores enabled, the MacBook Pro 16" will likely get all 8 cores enabled.
That will allow Apple to bin the chips so the ones with defects can still be utilised. Similarly the MacBook Air only needs a single or dual Thunderbolt ports while the 16" needs four ports. These too can be part of this binning process.
When it comes to the iMac I'd expect a single I/O die combined with dual 8-core chiplets so they can offer configurations from 8 to 16 cores. For the iMac Pro I'd expect four chiplets taking that system up-to 32 cores and the I/O die will likely contain a quad-channel memory controller which can be cut down to dual-channel when used in MacBook Pro's and Airs.
For the big Mac Pro I'd expect 8 chiplets so they can offer 16 (2 cores per die active) to 64 (all cores active) core configurations. But again this is all just conjecture but the fact is they have access to the same nodes AMD do and so the density is going to be relatively identical. Whether they do that, I'm just making educated guessing.
If they go with this approach it will result in lower prices for them and extremely high yield rates of all the dies they have manufactured and the Intel approach of just making ever increasingly dense monolithic dies is one of the reasons they can't compete right now and are being rejected by Apple. Intels yield issues with 10nm are highly exasperated by the large size and density of the dies they are trying to fabricate, an issue AMD simply don't face and Apple would be fully aware of this.
www.statista.com
www.marketwatch.com
