Google TPU 8i Pod

Google's eighth-generation inference pod, built on a Boardfly topology rather than a torus, with a maximum seven-hop ICI network diameter. Google describes joining up to 1,152 TPU 8i chips together with up to 1,024 of them active; the figures here use the 1,024 active count, because crediting the pod with throughput from chips Google does not call active would overstate it. At Google's published 10.1 PFLOPS of FP4 per chip that is 10.34 EFLOPS, with 294.9 TB of HBM and 8.8 PB/s of aggregate memory bandwidth, plus 393 GB of on-chip Vmem SRAM. Google publishes no pod-level compute figure; the total is chip count times its published per-chip peak, the same convention its own v4 and Ironwood pod figures confirm. FP4 is the only precision stated and no sparse figure exists. Announced at Google Cloud Next in April 2026 and not yet generally available.