Our newest TPUs: 8tx and 8i Max, built for the agentic era (blog.google)

Google has announced the expansion of its eighth-generation Tensor Processing Units (TPUs) with the introduction of two new systems, TPU 8i Max and TPU 8tx, designed to address emerging needs in the AI/ML landscape. These systems enhance Google's existing TPU lineup, which includes TPU 8t for large-scale training and TPU 8i for real-time reasoning. The TPU 8tx focuses on secure and confidential inference by integrating hardware-level encryption to protect sensitive information during processing, catering particularly to regulated industries. Meanwhile, the TPU 8i Max is engineered to tackle memory bandwidth limitations, delivering double the compute and memory bandwidth of its predecessor, thereby optimizing performance for large-scale AI workloads. This significant development underscores the growing demand for specialized hardware that can support autonomous agents and handle increasingly complex AI tasks. By allowing customers to deploy TPUs directly into their own data centers and providing low-latency performance with a focus on ease of use, Google is making these advanced accelerators more accessible to a wider audience. The new TPUs facilitate seamless integration with popular platforms like PyTorch and enhance operational efficiency through features like the Agent Compute Engine, which minimizes idle compute waste, thereby fostering innovation and scaling across the rapidly evolving AI landscape.