Timeline

Google adds two chips to its eighth-generation TPU line

TPU 8tx isolates data and weights in hardware for confidential inference; 8i Max joins two 8i chips for double the compute, SRAM and memory bandwidth. Google cites nearly 100 customers.

  • Compute & infrastructure
  • Minor

Google announced two additions to its eighth-generation TPU line, TPU 8tx and TPU 8i Max. They join TPU 8t, for training, and TPU 8i, for reasoning workloads, which the company introduced in April. Amin Vahdat, Google’s senior vice president and chief technologist for AI and infrastructure, wrote that it is the first time Google has introduced four TPU systems in a single year.

TPU 8tx is aimed at regulated enterprises and governments that run proprietary models and agents on sensitive data. It uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC host processors, and builds security into the hardware layer to keep data and model weights isolated while in use. It also uses the low-latency Boardfly network from TPU 8i; Google claimed up to twice the latency improvement over TPU 8t with a near-zero performance penalty. Google said teams could fine-tune, distil and run reinforcement learning on private data on one system without moving it between clusters.

TPU 8i Max puts two TPU 8i chips on one package. Google said that gives double the compute, SRAM and HBM bandwidth of TPU 8i, in the same rack footprint, and described the target as agents working over large codebases or long video. The post names no context-window size and gives no availability dates or pricing, and TechCrunch also reported none.

Google restated its April figures for the first two eighth-generation chips: up to 2.7 times better training performance for 8t and 80% better inference price-performance for 8i. It said TPUs now have nearly 100 customers, including frontier labs, trading firms, retailers and universities. All of these figures are Google’s, without independent benchmarks or comparison with other vendors’ chips. TechCrunch framed the launch as part of hyperscalers’ push to reduce reliance on Nvidia, an interpretation that is TechCrunch’s and not stated in Google’s post.

The announcement follows Google and Meta’s release of a PyTorch backend for TPUs in October, which targeted the software barrier to adopting the chips.