Google Presents Ironwood, a TPU Optimized for AI Inference

At its Cloud Next 2025 conference, Google introduced ‘Ironwood’, the seventh generation of its Tensor Processing Unit (TPU) designed specifically for AI inference tasks. Unlike previous TPUs, which were often optimized for training models, Ironwood focuses on running models efficiently at scale. The chip is built for Google Cloud and will be available later this year in two configurations: a 256-chip cluster and a much larger 9,216-chip cluster. According to Google, Ironwood is the company’s most powerful and energy-efficient TPU to date.

Each Ironwood chip boasts 4,614 TFLOPs of peak computing power, 192GB of dedicated RAM, and a data bandwidth of up to 7.4 Tbps. It includes a specialized core, SparseCore, tailored for recommendation engines and ranking algorithms. Google claims the architecture reduces data movement and latency, significantly improving power efficiency.

Ironwood will soon be integrated into Google’s AI Hypercomputer, a modular infrastructure for high-performance computing in the cloud. With growing competition from Nvidia, Amazon, and Microsoft in the AI hardware space, Ironwood marks a strategic advancement in Google’s AI infrastructure offerings.