[ITmedia News] Google adds “8tx” and “8i Max” to 8th generation TPUs, specialized for sensitive data processing and large-scale AI agents

Google announced on November 9th (local time) that it will add two new systems, "TPU 8tx" and "TPU 8i Max," to the 8th generation of its proprietary AI processor "TPU."

This follows the "TPU 8t" for large-scale learning and "TPU 8i" for real-time inference announced in April, and is the first time that four types of systems of the same generation have been announced in one year.

The two new systems are designed to address the different challenges of protecting sensitive data and memory performance for processing large amounts of information at once, against the backdrop of the proliferation of AI agents that autonomously execute multiple tasks.

The TPU 8tx is designed for businesses and government agencies that handle highly sensitive data. By adopting the same semiconductor as the TPU 8t and combining it with AMD's 5th generation EPYC processor, it protects the sensitive data being processed and the weight of the AI ​​model at the hardware level. By adopting the low-latency network ``Boardfly,'' the latency has been reduced by up to half compared to TPU 8t. Fine-tune models and perform inference without moving sensitive data to another system.

TPU 8i Max is specialized for processing that handles long contexts. It integrates two TPU 8i chips into one package, doubling computing performance, doubling on-chip memory (SRAM) capacity, and doubling high-bandwidth memory (HBM) bandwidth. It is intended for use with multiple AI agents handling large code bases and long videos.

Google also announced software enhancements for TPUs at the same time. "TorchTPU", which was made open source in October, allows PyTorch code and Hugging Face models to be run on a TPU without having to be rewritten in principle. The newly released "TPU Sync" is said to reduce data transfer delays required for AI processing by more than 90%. Additionally, Agent Compute Engine reduces wasted computing resources while the AI ​​agent is stopped waiting for responses from external tools.

Speaking at X, CEO Sundar Pichai emphasized that co-designing between Google DeepMind and hardware engineers will help speed innovation for customers, rather than waiting years for traditional semiconductor development cycles.

Google is also expanding its external supply of TPUs. In April, Anthropic announced plans to expand its partnership with Google and other companies to utilize next-generation TPUs. In October, Oracle also announced plans to provide Google's TPUs on Oracle Cloud starting in 2027.

Please note that this announcement does not reveal the specific launch date or pricing for the new system.