3 things to know about Ironwood, our latest TPU (blog.google)

Google announced Ironwood, its seventh-generation TPU, now available to Cloud customers. Purpose-built for the current era of high-volume, low-latency inference and model serving, Ironwood delivers more than 4× better per-chip performance for both training and inference compared with the prior generation while improving energy efficiency. The chip minimizes internal data movement to accelerate complex calculations, making interactive, large-scale AI services run faster and smoother across Google Cloud. Ironwood is a core part of Google’s AI Hypercomputer: TPUs are grouped into pods and can scale up to a 9,216-chip superpod linked by a 9.6 Tb/s Inter-Chip Interconnect (ICI), providing access to 1.77 PB of shared High Bandwidth Memory (HBM). That massive connectivity tackles data bottlenecks, cutting compute-hours and power for training and serving demanding models. The chip also reflects a tight hardware–research feedback loop—DeepMind collaborates directly with TPU engineers, and Google uses reinforcement-learning-based “AlphaChip” layout tools that helped design the last three TPU generations, including Ironwood—so hardware and models evolve together, speeding deployment and lowering cost for the AI/ML community.