[1]https://www.intel.com/content/www/us/en/foundry/emib.html
Launch early, assess market fit and plan accordingly. Intel did differently with Phi and it didn't quite work the way they (and I) expected.
Yeah but their target isn't to compete with a 48 core monolithic CPU it was to compete with AMDs EPYC CPUs which have even less uniform connectivity between sockets (3 per used for in socket interconnect and only 1 used to connect to the other socket). Unfortunately for Intel AMD is still going to be a step ahead this generation with the chiplet + IO module architecture IMO.
My belief is that it behaves similarly to a 4-way 12-core machine using less real estate for the CPU. Even though memory will occupy the same amount of space, I suspect we may see it on high-end workstation-class machines by the likes of Supermicro and Gigabyte.
>*"that leverages one of the three UltraPath Interconnect (UPI) links to hook the two physical chips together to create what essentially is a 48-core socket with 12 memory controllers"
If that's correct and it's 3 UPI, then presumably it'd be the square config so unlike a normal 4P system each processor does not necessarily have a direct link to all others, they may have to go through multiple hops. It definitely would not be "essentially a 48-core" at all since it'd complicate memory access patterns significantly wouldn't it? If that's the layout the potential for benchmark cherry picking and specific optimization that isn't applicable to other workloads seems even higher then normal.
AMD got a leap on Intel last with x86-64 because they didn't want to do that on the desktop, maybe MCM will turn out to be something a little like that (though less dramatic) as well given the realities of increasing difficulties in large single chip fabrication. It'd be interesting if internally Intel is rethinking that but this doesn't seem like "doubling down" so much as a first scramble.
----
1: https://news.ycombinator.com/item?id=18389373
2: https://www.anandtech.com/show/13535/intel-goes-for-48cores-...