Google released its latest Tensor Processing Unit (TPU) seventh-generation model, known as Ironwood, which represents a major progression in its customized AI hardware. Ironwood represents a significant advance, enabling advanced “agentic AI” abilities through simulated reasoning capabilities that Google calls “thinking” for its leading Gemini models in what the company terms the “age of inference.”

The company maintains that its Gemini models depend fundamentally on its infrastructure capabilities while its custom AI hardware significantly enhances inference speeds and context window sizes. Ironwood stands as Google’s most advanced TPU designed for scalability, enabling proactive AI operations that gather data and produce outputs independently which aligns with Google’s agentic AI vision.

The performance of Ironwood significantly surpasses that of its earlier models in terms of throughput capabilities. Google intends to implement these chips through very large liquid-cooled clusters which could comprise up to 9,216 units. A newly improved Inter-Chip Interconnect (ICI) will enable these chips to communicate directly for high-speed data transfer throughout the whole system.

The powerful design will serve multiple purposes beyond Google’s internal operations. Developers looking to run demanding AI projects in the cloud will also be able to leverage Ironwood through two distinct configurations: Ironwood offers both a 256-chip server configuration and an extensive 9,216-chip cluster option.

Google’s Ironwood pods reach an incredible peak of 42.5 Exaflops when configured at their maximum capacity. Google reports that each Ironwood chip delivers a peak throughput of 4,614 TFLOPs, which reflects a substantial advancement from previous models. Google’s new TPUs feature a substantial memory upgrade as each chip now comes with 192GB, which amounts to six times the memory capacity of the previous Trillium TPU generation. Memory bandwidth expanded significantly to achieve a new peak of 7.2 Tbps, which represents a 4.5 times enhancement.

Google performs benchmark tests on Ironwood using FP8 precision because direct hardware comparisons between AI systems remain complex because of different measurement methods. The assertion that Ironwood “pods” operate at a speed 24 times faster than segments of the world’s top supercomputer must be approached carefully because FP8 support is missing in some of those systems. Google’s direct comparison excludes its Trillium TPU v6 hardware. Google claims that Ironwood delivers double the performance per watt compared to v6. A company spokesperson indicated that Ironwood represents the successor to TPU v5p and that Trillium followed the less powerful TPU v5e. Trillium achieved computational performance levels of about 918 TFLOPS when operating at FP8 precision.

Ironwood stands out as a major development within Google’s AI framework despite the challenges associated with benchmarking. Ironwood delivers superior speed and efficiency compared to past TPUs, capitalizing on Google’s established strong foundation that has driven swift advancements in large language models and simulated reasoning. The advanced Gemini 2.5 model by Google continues to operate using earlier generation Tensor Processing Units (TPUs). Ironwood’s improved inference speed and efficiency will enable major advancements in AI abilities during the upcoming year, which marks the beginning of the “age of inference” and the development of more advanced agentic AI systems.