Overview
- Google, which announced the split chips Wednesday at Cloud Next in Las Vegas, said TPU 8t and TPU 8i will be available to Google Cloud customers later this year.
- The TPU 8t training superpod scales to 9,600 chips with 2 petabytes of shared high‑bandwidth memory to deliver nearly three times Ironwood’s training compute.
- Google targets over 97% goodput on TPU 8t by automatically rerouting around failed inter‑chip links and reconfiguring hardware without human intervention.
- The TPU 8i for inference adds 384 MB on‑chip SRAM and a new Boardfly interconnect to cut latency, with Google claiming up to 80% better performance per dollar than Ironwood.
- Both chips were co‑developed with Broadcom and engineered with Google DeepMind, and Google Cloud will also offer Nvidia’s Vera Rubin GPUs as it pushes TPUs to improve AI performance, energy use, and cost at scale.