In a significant stride towards the future of artificial intelligence, Google has unveiled its latest custom-designed processor: The Ironwood TPU represents the seventh iteration of Google’s Tensor Processing Unit architecture. The cutting-edge chip design targets the increasing computational needs of Google’s advanced Gemini models, which require complex reasoning capabilities referred to by Google as “thinking.”

Google highlights the essential synergy between its advanced AI models and their carefully designed infrastructure. Ironwood stands as a critical element in this architecture by delivering rapid inference speed improvements and broadening AI systems’ contextual comprehension capabilities. The company believes that Ironwood stands as its most scalable and powerful TPU yet, which enables AI systems to interact with users directly and independently gather information to produce relevant results. Google envisions “agentic AI” through this active user-focused framework, which finds its driving force in Ironwood as the engine powering the “age of inference.”

Ironwood: Powering the Next Generation of AI

Ironwood delivers significantly higher throughput performance relative to earlier generations of TPUs. Google’s deployment strategy aims to build enormous liquid-cooled clusters that integrate up to 9,216 Ironwood chips. The upgraded Inter-Chip Interconnect (ICI) enables these massive computational arrays to exchange data rapidly and with high bandwidth throughout the complete system.

Google’s internal research teams and cloud developers will both benefit from the access to this vast processing power. Ironwood will be available in two distinct configurations: The Ironwood system delivers two separate options, including a 256-chip server suited for moderate AI workloads and a large 9,216-chip cluster designed to execute the most intensive AI tasks.

The Ironwood pod achieves a stunning level of processing power and can deliver 42.5 Exaflops of inference computation when fully configured. The Ironwood chip achieves a peak performance rate of 4,614 TFLOPs, which represents a considerable improvement over older TPU models, according to Google’s reports. The memory architecture of Ironwood chips has been upgraded significantly to 192GB of memory, which represents a sixfold enhancement over the Trillium TPU’s memory capacity. The memory bandwidth improvement of 4.5 times enables the system to achieve 7.2 Tbps performance.

Decoding the Performance Metrics

Assessing AI chip performance through direct comparison proves complex because different benchmarking methodologies create variations. Ironwood’s performance evaluations at Google are based primarily on FP8 precision measurements. Although Ironwood “pods” reportedly deliver performance enhancements that are 24 times faster than existing segments of top supercomputers, the company has released this data with caution because many supercomputers do not support native FP8 hardware.

Google’s performance comparisons failed to include its TPU v6 (Trillium) hardware as a comparison point. Google claimed that Ironwood achieves double the performance for each watt compared to its previous version TPU v6. Ironwood serves as the next generation TPU after v5p according to the company spokesperson while Trillium succeeded the less powerful TPU v5e. Trillium reached peak performance levels of around 918 TFLOPS while operating at FP8 precision.

The Implications for the Future of AI

Despite the inherent complexities in benchmarking AI hardware, the underlying message is clear: Ironwood delivers a major advancement to Google’s AI infrastructure capabilities. Ironwood delivers substantial improvements in speed and efficiency, which support the strong base that enabled quick progress in advanced models such as Gemini 2.5 that run on older TPU technology.

Google expects Ironwood’s improved inference performance and efficiency to lead to groundbreaking advancements in artificial intelligence during the next year. Ironwood delivers essential computational power for highly complex models and true agentic capabilities, which makes it a critical component of Google’s “age of inference” vision where AI integrates proactively and intelligently into our digital existence.