Google unveils eighth-generation TPUs, split into TPU 8t for training and TPU 8i for inference

Google DeepMind/Tensor Processing Unit (TPU)

Google announced two specialised eighth-generation TPU chips designed for the agentic era. TPU 8t focuses on training with nearly 3x the compute per pod of the previous generation. TPU 8i emphasises inference with 288 GB of high-bandwidth memory and doubled physical hosts per server, both featuring liquid cooling and 2x better performance-per-watt.

• TPU 8t (training): scales to 9,600 chips with 2 petabytes shared memory, delivers 121 ExaFlops, achieves >97% goodput and near-linear scaling to million chips
• TPU 8i (inference): 288 GB high-bandwidth memory plus 384 MB on-chip SRAM, doubled interconnect bandwidth for Mixture of Expert models, 80% better performance-per-dollar
• Both feature custom Axion ARM-based CPUs, liquid cooling, JAX and PyTorch support, and 2x better performance-per-watt versus prior generation
• Designed for agentic systems reasoning through problems and executing multi-step workflows; general availability planned later in 2026