Google and AMD Detail Confidential AI Computing on 8th-Gen TPU 8tx and 5th-Gen EPYC
Google and AMD shared technical details on pairing 5th-Gen AMD EPYC host processors with Google's 8th-generation TPU 8tx to run hardware-enforced confidential AI inference.

By Marcus Vance, Semiconductors & Hardware
1 min read
Google and AMD have detailed how 5th-Gen AMD EPYC processors serve as the confidential host layer for Google's 8th-generation TPU 8tx accelerators, extending hardware-isolated execution to enterprise AI agents.
Securing model weights and prompts in hardware
Following Google's rollout of the TPU 8tx for confidential inference alongside the dual-die TPU 8i Max with 384 GB of HBM, the joint architecture encrypts data in memory between the EPYC host CPU and the TPU accelerator. Regulated enterprises in healthcare, finance, and government can verify cryptographic attestation before sending sensitive prompts or proprietary weights.
Competing on trust for agentic workloads
As cloud providers court regulated industries for autonomous agent deployments, hardware-enforced confidential computing is becoming a standard procurement requirement alongside raw token throughput.