todays only way to HPC

think stacking chips, 3D memory, MLC memory etc – you will NOT get to the highest performance since you are still left with interconnects that go thru several solder points, traces on PCBs, fiber optics etc. that will be the biggest bottleneck you have – transmitting the data to an adjacent device

consider putting all CPUs, cache, memory on one 12 inch wafer. that’s a few hundred billion transistors with todays tech. then interconnect two wafers with a PCB in the middle like a sandwich, that reaches out on the sides to offer space for network and power, and cool the whole thing with a liquid – you will need to use liquid since you have a great power density

with sophisticated algorithms you need to lock out (and disconnect power) from CPUs and memory that is damaged, or run those with lower frequencies or capacities, most likely the outer diameters of the wafer (since the target angle here during UV processes is unfavorable). so just group slow dynamic RAM on the outside, and all CPU in the center, surrounded by cache and interconnects. leave the “dead” components there but with no power – their surface will help cool down the adjacent working devices anyways.

it will not get more dense than that – unless you use via’s thru the wafer and sandwich dozens of layers, but how will you cool that? small holes or channels for liquid in between the layers … there are some solutions but for today the sandwich with 2 wafers is the fastest way to get HPC right now