SemiAnalysis says moving TPU routing to switches could free chip space for Anthropic
SemiAnalysis describes potential room for more compute or memory if routing cores could be dropped. It also points to a unified network architecture with other accelerators, such as Trainium or GPUs.
TLDR
SemiAnalysis says Anthropic could gain more chip space for compute or memory if switches took over routing, allowing routing cores to be removed. It also describes a potential benefit in unifying the network architecture with other accelerators, such as Trainium or GPUs. The configuration it discusses would leave 64 ports for cross-plane or intra-plane dragonfly interconnects, with a 4:1 oversubscription ratio and a maximum cluster size of up to 1,280 interconnected TPUs.
Combined views
117.5K
4 Sources, first seen 18d ago