Google detailed its eighth-generation custom accelerators: TPU 8t for training and TPU 8i for inference.

Google presented its eighth-generation Tensor Processing Unit family at Hot Chips 2026. Tensor Processing Units are Google-designed accelerators for machine-learning workloads rather than general-purpose central processing units.

The family separates its primary workload roles: TPU 8t is designed for model training, while TPU 8i is designed for inference. Training covers the computation used to develop or update models, whereas inference runs trained models to produce outputs.

Google is one of the hyperscale cloud operators developing its own training hardware. The supplied source identifies the two accelerator variants but does not provide memory capacity, power consumption, benchmark results, system scale or general-availability timing.