Google rolled out its eighth‑generation Tensor Processing Units, TPU 8t for large‑scale pre‑training and TPU 8i for real‑time serving, as part of the new AI Hypercomputer.

TPU 8t introduces a SparseCore accelerator, native FP4 precision, and the Virgo Network, delivering up to 4× more data‑center bandwidth and near‑linear scaling across 1 million chips. Direct RDMA and storage cut data‑bottleneck latency, allowing line‑rate ingestion of petabyte datasets.

TPU 8i doubles on‑chip SRAM, adds a Collectives Acceleration Engine, and adopts the Boardfly topology to accelerate sampling and multi‑turn reasoning. Together, the two chips let Google train and serve world‑model workloads like DeepMind’s Genie 3 with unprecedented efficiency.