Overview
- Google announced on Monday two new eighth‑generation TPU system variants that reconfigure the previously unveiled TPU 8t and TPU 8i designs to target confidential computing and large‑context inference.
- The TPU 8tx pairs TPU 8t silicon with custom AMD EPYC hosts and hardware‑level encryption to isolate data and model weights during active processing, and Google says it can cut latency by up to half versus TPU 8t with minimal performance cost.
- The TPU 8i Max combines two TPU 8i chips in a single package to double on‑chip SRAM, HBM bandwidth and compute density within the same rack footprint so models can hold far larger context windows without frequent data transfers.
- Google credited co‑design with DeepMind and use of Gemini to speed chip layout and system work, said its TPU business serves nearly 100 external customers, and has not disclosed pricing or firm availability dates for the two new systems.
- By offering a security‑focused option and a memory‑dense option, Google is expanding choices for regulated enterprises and long‑context AI workloads and positioning TPUs as an alternative to incumbent GPU stacks while independent benchmarks and commercial terms remain pending.