Google has expanded its eighth-generation tensor processing unit family with two specialized chips aimed at confidential workloads and larger inference jobs.
TPU 8tx, developed with AMD and based on the TPU 8t, adds hardware protections intended to isolate sensitive data and model weights while they are actively used. Google says a custom inference network also lets the system run reasoning alongside training, which could support distilling smaller models without repeatedly moving proprietary data between devices.
TPU 8i Max combines the resources of two inference-focused TPU 8i chips into one package. The design doubles memory capacity, processing power and bandwidth, allowing a model to handle more information during a query. It extends the 8i product introduced earlier this year rather than replacing Google’s general training hardware.
The additions give Google Cloud customers more alternatives for workloads that might otherwise rely on Nvidia accelerators. Microsoft, Amazon and several AI labs are likewise developing specialized silicon to control cost and supply. The announcement describes the intended capabilities, but it does not provide independent performance, energy-efficiency or price comparisons, so buyers will still need workload-specific benchmarks.