Yesterday, Google introduced two new eighth-generation TPU chips, the 8tx and 8i Max, to boost confidential AI inference and massive context processing.
The 8tx is built for secure, confidential AI inference, featuring hardware-level isolation that protects data and model weights during use. It pairs with fifth-generation AMD EPYC CPUs and offers low-latency networking, delivering up to twice the speed of the earlier 8t model with minimal performance loss. The chip also supports small-scale training, fine-tuning AI models with private data, and reinforcement learning for AI agents.
The 8i Max targets massive context windows and the highest compute density per rack, doubling compute, SRAM bandwidth, and HBM bandwidth. It fits twice the silicon into existing racks, allowing data centers to deploy more computing power per square foot and per watt. Both chips were co-designed with Google DeepMind, with input from regulated enterprise and cloud customers.