Not sure, was hoping to get some insight here. I know NUMA has many significant challenges but so does adding more cpu/gpu cores and on package memory to compete with current desktop/workstation levels on the ram and gpu sides.

Just seems elegant to do 4 M1's and get 64gb total memory and 4x gpu/cpu and seems hard to scale up a single chip to match current multi chip / discrete gpu systems.

Maybe even dimms for main memory and current on package memory acting just as a very large level of cache.