The interesting thing is William Ma said Apple was working on an "M2 Duo" which would effectively be two M2s - 8 performance cores, 8 efficiency cores and 18 or 20 GPU cores. That sounds superfluous in terms of efficiency cores.
M2 only Duo? That doesn't make much sense. It would actually be more perf/watt efficient just to make that one die. What is Apple suppose to be 'buying' there by tossing a substantial edge of a modest size die at a fractional UltraFusion connector? If wanted to disaggregate the Pro/Max up into smaller more reusable block they would remove CPU and GPU cores from the same chiplet. Really wouldn't have a "M2" chunks anymore.
The M1 Pro has. 8P/2E and 16GPU cores. Going up 6E and 2-4GPU cores doesn't require a chiplet/tile solution at all. Especially if getting 10+% die shrink with the upgrade. Even if using TSMC N5P and no density improvement... just make a larger die . The M1 Pro is 251mm^2 die. Going to 280-90mm^2 wouldn't be onerous at all.
At the 350-450mm^2 it is wishy-washy of whether chopping up really brings any improvements.
450-550mm^2 lots more traction.
550-700 mm^ even more traction.
> 800 mm^2 don't really have much of a choice anymore.
but down in the 100-200 range the overhead adding doesn't make much sense for general purpose computation.
And then they make a version with 12P cores, 4E cores and 36/40 GPU cores and call it the M2 Max (to replace M1 Max)?
If All M2 is A15 cores and N5P then there is very little 'shrink' going on here with this update. N5P is geared to saving a little bit of power or nudging clocks up but there are not substantive area/density improvements at all. Doing a two more E cores would be the lowest area penalty. It is easiest thing to do if highly limited with N5P as a foundation.
A M2 Pro with 8P+4E would be this "M2 12 core" being rumored and may not even touch the Max. ( If Apple is forking off the M2 Pro only to the Mini). M2 and M2 Pro would allow updates for : MBA , MBP 13" two port , revised shrunken Mini. and done in 2022. [ NOTE: all of those are small capacity enclosures so running more on E cores has upsides. ] They could do a non-UltraFusion Max that just took two E cores, but that probably doesn't make much sense if M3 Max really are not that relatively far way in 2023.
If the M2 is new cores and N4/N4P then still wasn't much of an area win and if Apple wanted to allocate two more GPU cores then again the 'left overs' for the CPU core usage would again be E core sized. Could also get an 'easy' Max bump with two more E cores here and some thrown at a couple of GPU cores (or other function unit uplifts). [ again if the Max was going to get tossed into a smaller/thinner container than previous generation system that would be helpful. ]
The other issue is if Apple think's Intel's push to get into the "core count" wars with Gen 13 and 14 ( Raptor and Meteor Lake) will have traction with general market consumer. That Dell XPS has 24 cores and the Mini only has 12 . 8 E cores sounds like someone mapping Intel's marketing moves onto what Apple feels compelled to follow. That rings hollow. Intel has problems. That's why trying to turn it into a "core count" war.
12P+4E+36 ( +2 , +4 , +4 ) for a Max with N5P ... probably not going to happen. You are basically growing the chip bigger, and the memory bandwidth isn't changing that much. Even with N4/NP there really isn't a "grow everything" density improvement to support that (and again probably not getting a memory bandwidth bump). It is already a "big" chip (relatively to Apple's previous norms. ). I highly doubt Apple will be looking to
decrease their margins with a bigger die (with higher costs). The "add to everything" is far more plausible with a more sizable node shrink and that isn't coming until TSMC N3 (and 2023 ).
