Google Packs Two AI Chips Into One Package and Adds a Separate System for Confidential Computing

Google's answer to AI's memory problem is to put two chips in the same package. On November 9, it introduced TPU 8i Max alongside TPU 8tx, a separate system for running AI on sensitive data. Both reconfigure existing eighth-generation silicon, shortening the wait for specialized hardware.

More memory in the same rack

Google says 8i Max doubles compute, on-chip memory and high-bandwidth memory bandwidth relative to TPU 8i. The target is agents working across entire codebases or hours of video, where keeping data close to the processor becomes the constraint.

That is a different problem from making a model smarter. A processor waiting for memory cannot spend that time generating useful answers, however impressive its theoretical speed.

Keeping data protected during use

TPU 8tx pairs TPU 8t silicon with fifth-generation AMD EPYC hosts to protect data and model weights during processing. Google pitches it to governments and regulated businesses, with faster networking for multi-step agents.

The announcement expands Google's range without waiting for another chip generation. These are Google's design and performance claims; the post does not provide customer pricing. How much buyers save will depend on the workload and the bill.

Some offers on this page may be paid placements or contain affiliate links.