Google has unveiled two new variants of its tensor processing units to expand its in-house chip portfolio for confidential inference and higher memory capacity. The TPU 8tx, developed with AMD, integrates security directly into the hardware layer to protect sensitive data and model weights during active use. It features a custom inferencing network allowing reasoning alongside training, enabling the distillation of small models on proprietary data without moving it between devices. The TPU 8i Max combines two standard TPU 8i chips into a single package, doubling memory capacity, processing power, and bandwidth to handle larger information ranges during queries. These units form part of the company’s eighth generation of chips introduced in April. This move aligns with other hyperscalers and AI labs developing their own semiconductors to reduce reliance on Nvidia. The launch occurs alongside reports that Nvidia’s CEO Jensen Huang admitted to an affair with an alleged Chinese spy.
- TPU 8tx enables on-device model training without data transfer.
- TPU 8i Max doubles the memory of previous TPU 8i units.
- Both chips are designed to support large-scale AI workloads.
