Google adds two specialized eighth-generation TPU variants

Google has announced two new variants of its eighth-generation Tensor Processing Units, designed for confidential computing and artificial intelligence workloads that need long context windows. The TPU 8tx and TPU 8i Max expand the company’s lineup of custom AI processors, which also includes the TPU 8t and TPU 8i.

Google says the 8tx is intended for sensitive workloads at government agencies and businesses in regulated industries. It pairs TPU 8t-based silicon with custom central processing unit hosts from Advanced Micro Devices and adds hardware-level confidential computing. Google says the design keeps sensitive data and model weights protected while processing takes place. Amin Vahdat, Google’s chief technologist for AI infrastructure, said the company reconfigured existing chips to address tasks customers had requested.

The 8i Max is designed for applications that need to process large amounts of information within a single request, including work involving long documents, code or other data. Google says it combines two TPU 8i chips in one package, providing twice the compute and memory resources of the 8i. The company has not confirmed the 8i Max’s high-bandwidth memory...