The one thing I totally don't understand about this architecture is the meager 40GB of RAM. I understand it's SRAM and the high memory bandwidth allows for reduced replication, but almost every workload I can imagine this addressing is still going to have a larger working set than that. I guess they still haven't figured out how to stack HBM on top of the wafer, and they're just carefully avoiding calling attention to this extreme (factor of 1000) weakness.