Google unveils eighth-generation TPU chips for AI training and inference

Google has unveiled the eighth generation of its Tensor Processing Units, introducing two dedicated chips: the TPU 8t for training and the TPU 8i for inference. The company said the hardware was developed in partnership with Google DeepMind and features purpose-built architectures for model training, agent development, and inference workloads. Amin Vahdat, Google's senior vice president and chief technologist for AI and infrastructure, described the TPU 8t as a training powerhouse built to reduce frontier model development cycles. A single TPU 8t superpod can scale to 9,600 chips with two peta The TPU 8t introduces a new networking architecture called Virgo Network, which Google said supports a fourfold increase in data center bandwidth. The company also introduced TPUDirect RDMA and TPU Direct Storage to enable direct data transfers that The TPU 8i is designed for latency-sensitive inference workloads, particularly the collaborative work of specialized agents. It scales to 1,152 chips per pod and delivers 11.6 exaflops of FP8 compute performance. Google said the TPU 8i delivers 80 percent better performance-per-dollar compared to Ironwood, with up to twice the performance-per-watt. Both chips run on Google's custom Axion Arm-based CPUs, support liquid cooling, and will be generally available later this year.