Google reworks TPU silicon into specialized systems for agents and security
What happened
Google has introduced two new versions of its eighth-generation TPU hardware: TPU 8tx and TPU 8i Max. TPU 8tx is designed for secure, hardware-isolated inference, targeting organizations in regulated industries and government. TPU 8i Max combines two 8i chips in one package to provide higher compute and memory bandwidth for agent-related workloads that require longer context. Both systems are based on the existing TPU 8t and 8i silicon and are intended to speed up AI deployment for enterprise and government users.
Key details
- TPU 8tx uses a 5th Gen AMD EPYC CPU and Boardfly network instead of the Arm-based Axion hosts found in TPU 8t and 8i.
- TPU 8i Max doubles the compute, SRAM, and HBM bandwidth compared to the original TPU 8i.
- TPU 8tx is aimed at confidential inference tasks for regulated sectors.
- TPU 8i Max is designed for extended-context agentic workloads.
- Google reports nearly 100 TPU customers, including Meta.
Why it matters
Enterprise and government customers now have access to specialized TPU systems that address specific needs: secure, isolated inference for compliance and higher performance for complex agent tasks. This allows organizations in regulated industries to deploy AI solutions more quickly and with greater assurance of data security. The new hardware options can reduce deployment times from years to months, giving customers more flexibility and speed in adopting AI technologies.