TLDR on my views on AI Memory: -Custom memory either 3D + custom base die (maybe mix of both) is the future. If you aren’t working closely with your memory suppliers to do this right now you will die. -Base die will be on leading edge logic as compute moves near memory. I can see this being on N4 & N3 in no time. -Therefore, I expect to see deep collaboration between logic & memory foundries going forward -Commodity DRAM market share per AI inference server will fall by ~40% & I expect more despec of ai servers. -Flash memory eats the reduction in DRAM market share & flash over scale out network becomes standard. -Scale up domain is the new bottleneck & expect scale up to go to over 128 nodes and some will do 500+. Will lean on NPO, VCSEL, copper, PCB etc. -SRAM or in better words on chip weights becomes real beneficial as we see things move more from memory bound to compute bound. Sources: my imagination + experience + reading papers
