“Google's TPUs use a scratchpad memory architecture where software explicitly issues different instructions to access on-chip versus off-chip memory, unlike a traditional CPU cache which is managed by hardware.”