Google has unveiled its seventh-generation Tensor Processing Unit (TPU), dubbed Ironwood, at Hot Chips 2025, building on its initial announcement at Google Cloud Next ’25 in April, marking a significant advancement in AI computing.
The Ironwood TPU is specifically designed for large-scale inference workloads, a shift from previous generations that focused on training. Each Ironwood chip incorporates two compute dies, delivering 4,614 TFLOPs of FP8 performance. The chip features eight stacks of HBM3e, providing 192 GB of memory per chip with a 7.3 TB/s bandwidth. The system architecture is designed to scale up to 9,216 chips per pod, facilitated by 1.2 TB/s of I/O bandwidth, eliminating the need for glue logic and achieving a total of 42.5 exaflops of performance.
A notable highlight of Ironwood is its extensive memory capacity. A single pod provides 1.77 PB of directly addressable HBM, which Google claims is a new world record for shared memory supercomputers. This is made possible by optical circuit switches that link racks together, enabling seamless communication between the chips.
The Ironwood TPU also prioritizes reliability and resilience. The hardware is equipped with the ability to automatically reconfigure around failed nodes and restore workloads from checkpoints. Additional features include an on-chip root of trust, built-in self-test functions, silent data corruption mitigation, and logic repair functions to improve manufacturing yield. According to Google, a strong emphasis on RAS (reliability, availability, and serviceability) is evident throughout the architecture.
Cooling for the Ironwood TPU is handled by a cold-plate solution integrated with Google’s third-generation liquid-cooling infrastructure. Google claims that Ironwood achieves a twofold improvement in performance per watt compared to its predecessor, Trillium. Dynamic voltage and frequency scaling further enhance efficiency during varied workloads, ensuring optimal performance.
The design of Ironwood also leveraged AI techniques to optimize ALU circuits and floor plans. A fourth-generation SparseCore has been added to accelerate embeddings and collective operations, supporting workloads such as recommendation engines. This integration of AI in the design process underscores Google’s commitment to pushing the boundaries of AI computing.
Ironwood deployment is currently underway at hyperscale within Google Cloud data centers. However, the TPU remains an internal platform and is not directly available to Google Cloud customers. The development and deployment of Ironwood reflect Google’s long-term investment in AI compute infrastructure.
Ryan Smith of ServeTheHome commented on Google’s presentation at Hot Chips 2025, stating, “This was an awesome presentation. Google saw the need to create high‑end AI compute many generations ago. Now the company is innovating at every level from the chips, to the interconnects, and to the physical infrastructure. Even as the last Hot Chips 2025 presentation this had the audience glued to the stage at what Google was showing.”
