Google unveils TPU 8tx and TPU 8i Max AI systems

The TPU 8i Max combines two TPU 8i chips into a single module. According to Google, compared to the standard TPU 8i, the new system offers twice the computing power and SRAM memory, while HBM bandwidth is also doubled. With identical physical dimensions, it contains twice as much silicon, which the company says increases data centre density per unit area and per watt.

The TPU 8tx is aimed at government agencies and companies in strictly regulated industries. In it, the TPU 8t operates alongside fifth-generation AMD EPYC processors, with data and model weights protected by hardware even during processing. Google claims the system's latency is up to twice as low as that of the TPU 8t. It can be used to fine-tune models on confidential data without transferring it between clusters.

The company also showcased software tools for the accelerators, including the Agent Compute Engine, TPU Sync, and the open-source XLA Serving Engine. According to Google, chip component placement layouts are now generated in hours instead of weeks with the help of Gemini.

The company revealed the TPU 8t and TPU 8i in April at the Google Cloud Next conference. As iXBT reports, Google's Chief Technologist for AI and Infrastructure, Amin Vahdat, cited the number of TPU ecosystem clients at nearly 100. Delivery dates and pricing for the new systems have not been disclosed.