Google has introduced two new Tensor Processing Units, the TPU 8t for training and TPU 8i for inference, targeting what the company calls the agentic era of AI. These eighth-generation chips follow the Ironwood TPU from 2025 and promise faster, more efficient AI development. The hardware aims to cut training times for large models from months to weeks.
Google announced the TPU 8t and TPU 8i on Tuesday, positioning them as specialized accelerators for different stages of AI model lifecycles. The TPU 8t focuses on training frontier models, with updated server clusters called pods housing 9,600 chips and two petabytes of shared high-bandwidth memory. Google states these pods deliver 121 FP4 EFlops of compute, nearly three times higher than the previous Ironwood generation, and can scale linearly to a million chips in a single cluster. The company claims a 97 percent 'goodpute' rate, thanks to improved memory handling, automatic fault management, and real-time telemetry across chips, reducing wasted time and effort. Training times for massive AI models are expected to drop from months to weeks, Google says. The TPU 8i handles inference, the phase where trained models generate responses. These chips operate in larger pods of 1,152 units, providing 11.6 EFlops per pod. Each TPU 8i features triple the on-chip SRAM at 384 MB, enabling larger key-value caches for models with extended context windows. For the first time, the chips pair exclusively with Google's custom Axion ARM CPUs, using one CPU per two TPUs, which Google says boosts overall efficiency compared to the prior x86 setup servicing four TPUs. Efficiency gains extend to power and cooling. The new TPUs offer twice the performance per watt of Ironwood, while data center designs integrating networking and compute have increased computing power per electricity unit sixfold. Liquid cooling now uses actively controlled valves to match water flow to workloads. These chips will support Google's Gemini-based agents and third-party developers via frameworks like JAX, MaxText, PyTorch, SGLang, and vLLM. Nvidia's stock dipped 1.5 percent briefly after the news but recovered.
