The TPU 8t and TPU 8i represent a strategic shift for Google, which previously experimented with separate variants in the fifth generation but later adopted a single-design approach with Trillium and Ironwood. According to Phil Fersht of HFS Research, the split designs aim to improve utilization and cost efficiency in production environments by tailoring hardware to specific AI tasks.
Google stated that the TPU 8t scales up to 121 exaflops over 9,600 chips, with double the bidirectional scaling bandwidth and quadruple the network bandwidth compared to its predecessor Ironwood. Omdia analyst Alexander Harrowell noted that the increased performance and inter-rack bandwidth will support training even larger models with shorter run times.
These two chips are designed to power our custom-built supercomputers, to drive everything from cutting-edge model training and agent development, to massive inference workloads. TPUs have been powering leading foundation models, including Gemini, for years. These 8th generation TPUs together will deliver scale, efficiency and capabilities across training, serving and agentic workloads.
The inference-focused TPU 8i features at least three times more memory than Ironwood, including 288 GB of high-bandwidth memory and 384 MB of on-chip SRAM. Harrowell said this brings TPUs closer to the memory footprint of leading GPUs, while the expanded SRAM reduces latency for large models. Sopko of Hyperframe Research added that the architectural changes reflect the industry's shift toward Mixture of Experts and long-context models.
Google plans to use the new TPUs for its Gemini models and also sells the chips to other parties, hoping to compete with Nvidia's dominant GPUs. In a statement, Google said: "These two chips are designed to power our custom-built supercomputers, to drive everything from cutting-edge model training and agent development, to massive inference workloads." Availability dates and pricing have not been disclosed.