What I've heard is that the M1 Ultra has to be all contiguous to exist. The interconnact cannot be assembled later. All those M1 Ultra chips are indeed 64 cores and the IO on the same die. https://patents.google.com/patent/US20210217702A1/en?oq=2021...
Sure, it's not interconnected as in multiple sockets on the motherboard talking over QPI, but it's a silicon interconnect. The M1 Ultra is two M1 Max processors on an interconnect with some DRAM.
Here's the Pro, with 8 or 10 CPU cores and 14 or 16 GPU cores, depending on die yield issues where they might have to shut off a couple cores:
Yeah but those two Max chips that make one Ultra have to be next to one another already on the wafer. That leaves no room for error or picking up a chip from somewhere else.
There are a lot of ways to win Tetris with 1x2 blocks!
If you make an M1 Max wafer there are a lot of ways to spin out 8/24-core Maxes, 10/32-core Maxes, and adjacent 10/32 core Maxes into an Ultra. Nothing in the patent seems to indicate they have to decide which ones to use before testing the wafer...
