Google TPU 8t Superpod

Google's eighth-generation training superpod: 9,600 TPU 8t chips on a 3D torus, with 47 petabits per second of non-blocking bisectional bandwidth, and a Virgo network able to link over 134,000 chips across superpods. At Google's published 12.6 PFLOPS of FP4 per chip that is 120.96 EFLOPS across the pod, with 2.07 PB of HBM3e and 62.7 PB/s of aggregate memory bandwidth. Google publishes no pod-level compute figure for this generation; the total here is chip count times its published per-chip peak, which is Google's own arithmetic — on the two generations where it publishes both, v4 and Ironwood, the pod figure equals exactly that product. FP4 is the only precision Google states, and it publishes no sparse figure for any TPU, so this is dense. Announced at Google Cloud Next in April 2026 and not yet generally available.