Overview
- Ironwood, Google’s seventh‑generation TPU, will be available to customers in the coming weeks and is built for large‑scale model training, reinforcement learning, and low‑latency inference.
- Ironwood pods scale to 9,216 chips with 9.6 Tbps inter‑chip bandwidth and access to shared high‑bandwidth memory, using optical circuit switching and torus fabrics for reliability and flexible sizing.
- Google cites performance gains of roughly 10× over TPU v5 and about 4× over TPU v6/Trillium, positioning Ironwood for both training throughput and high‑volume serving.
- Alongside the TPU rollout, Google introduced Axion Arm‑based instances, with C4A generally available and N4A and C4A Metal in preview to support general workloads and specialized deployments.
- Anthropic plans to use up to one million TPUs for Claude as Google targets Nvidia’s dominance in AI infrastructure and lifts 2025 capex guidance to about $93 billion to meet demand.