Google Announces 8th-Gen TPU ‘8t’ and ‘8i’ — Maximizing Efficiency by Separating Training and Inference

Google unveiled its 8th-generation TPUs, splitting the line into TPU 8t for training and TPU 8i for inference.

Both chips are set to ship later this year, with 8t targeting better performance and cost efficiency than the previous generation.

TPU 8i focuses on lower latency and more efficient inference, while both models improve power efficiency.

The move is aimed at strengthening Google Cloud’s AI infrastructure and improving operational efficiency for customers in the AI agent era.