Google's eighth-generation TPU splits into two architectures for the first time. STT breaks down how TPU 8t removes the training I/O bottleneck, and how TPU 8i uses large on-chip SRAM and OCS optical circuit switching to cut MoE inference network hops by 56%.