Google introduces TPU 8t and TPU 8i

Google has unveiled its eighth-generation Tensor Processing Units (TPUs), the TPU v8t and TPU v8i, representing a decade of custom silicon development aimed at powering the next wave of AI. The new chips are explicitly designed for what Google calls the 'agentic era,' where AI models evolve from simple chatbots into autonomous agents capable of planning, reasoning, and executing complex, multi-step tasks. The TPU v8t is optimized for training massive models, while the TPU v8i focuses on high-efficiency inference.

Key technical advancements include a 2.5x improvement in training performance per chip and a 1.7x boost in inference efficiency compared to the TPU v5e. The architecture is built to scale efficiently, supporting the development and deployment of trillion-parameter models. This performance leap is critical for running the complex reasoning and tool-use required by advanced AI agents. The chips will be available through Google Cloud, powering services like Gemini and Vertex AI, and offer a more performant and potentially cost-effective alternative to NVIDIA's GPUs for large-scale AI workloads.