Google adds confidential computing and large-context systems to its TPU range
Google has announced two systems built around its existing eighth-generation AI processors. TPU 8tx targets work with sensitive information, adding hardware protections while data is being processed. TPU 8i Max combines two inference chips for large-context tasks such as working through codebases or long videos. The announcement expands system designs rather than introducing a new chip generation.
Artificial Intelligence··Midday
Two systems build on the existing processor generation
Google has added TPU 8tx and TPU 8i Max to its eighth-generation Tensor Processing Unit family, processors designed for artificial intelligence workloads. The November 9 announcement concerns systems built around existing silicon. The two additions address sensitive-data processing and applications handling large contexts, the material a model works through during a task, such as a codebase or long video.[1], [2]
The company introduced TPU 8t for large-scale training and TPU 8i for inference at its Cloud Next conference in April. Training shapes a model’s behaviour; inference runs the trained model to produce an answer. The new systems reorganize computing, networking, memory and security within that existing generation, rather than marking the launch of a newly numbered chip family.[1]
Confidential computing extends protection into processing
TPU 8tx is designed for regulated businesses and governments working with sensitive information. It pairs the same silicon used in TPU 8t with fifth-generation AMD EPYC host processors. Its confidential-computing design seeks to protect data while it is actively processed, adding that protection to measures for information held in storage or travelling over a network.[1]
The system uses Google’s Boardfly network design for low-latency communication. The company lists private fine-tuning, model distillation and reinforcement learning among its intended uses. Those processes adapt a model, transfer capabilities into another model or shape behaviour through rewards. These are Google’s descriptions of the architecture and supported tasks; the announcement contains no independent assessment of security across customer deployments.[1]
A combined package targets large-context work
TPU 8i Max places two TPU 8i chips in one package, expanding computing resources, on-chip memory and high-bandwidth memory connections together. Google positions it for tasks involving entire codebases and lengthy videos, where moving information through memory can constrain processing. The system brings those resources together around the same inference-chip design already in the eighth-generation portfolio.[1]
The announcement also refers to earlier software work, including TorchTPU, which Google open-sourced the previous month. That provides context for the hardware systems rather than another new release in this announcement. Google has not set out a complete regional general-availability timetable or every deployment condition for the two additions.[1]
