Google adds two new eighth-gen TPU chips
Google introduced two new chips to its eighth-generation Tensor Processing Unit family. The TPU 8i Max doubles the compute, SRAM, and high-bandwidth memory of the existing TPU 8i, fitting twice the silicon in the same rack footprint to support ultra-large context windows, including the 1 million-token standard used by Gemini. The TPU 8tx is built for confidential inference, embedding a secure enclave at the hardware level to keep sensitive data and model weights protected. It targets regulated enterprises and governments needing to fine-tune models or train agents on private data without migrating it between clusters.
