In a significant stride towards the future of artificial intelligence, Google has unveiled its latest custom-designed processor: Google’s new Ironwood TPU marks the introduction of the seventh generation within its series of Tensor Processing Unit architectures. Google’s advanced Gemini models require computational power to handle complex reasoning tasks, which the company refers to as “thinking,” and this new chip provides the necessary support.

Google regularly emphasizes how its advanced AI models interact with its highly specialized infrastructure. The Ironwood system becomes essential to this ecosystem as it offers significant enhancements to inference speeds while increasing AI systems’ contextual comprehension capabilities. The company claims Ironwood as its premier scalable and powerful TPU, which enables an age where AI can independently interact with users by sourcing information and producing relevant results. Google’s “agentic AI” design revolves around proactive user-centric principles, and Ironwood powers this new era of AI inference.

Ironwood: Powering the Next Generation of AI

Ironwood represents a major advancement in throughput performance beyond previous TPU iterations. Google has developed an ambitious deployment plan that includes building enormous liquid-cooled clusters made up of as many as 9,216 Ironwood chips each. These large computing clusters will exchange data rapidly and with high capacity through a newly improved Inter-Chip Interconnect (ICI), which enables seamless communication throughout the system.

The massive processing power that Ironwood provides will serve Google’s internal R&D needs and developers who use cloud-based systems. Ironwood will be available in two distinct configurations: The Ironwood system features two configurations, which include a smaller 256-chip server for moderate AI workloads and a larger 9,216-chip cluster designed for high-intensity AI tasks.

The full-capacity Ironwood pod demonstrates impressive processing power with its 42.5 Exaflops of inference computing capabilities. By Google’s standards, Ironwood achieves a peak throughput of 4,614 TFLOPs on each chip, which represents a substantial performance advancement over previous TPU generations. The memory architecture received a massive enhancement since each Ironwood chip now features 192GB of memory, which represents six times more memory than the Trillium TPU. Memory bandwidth has improved by 4.5 times from previous levels to reach an impressive 7.2 Tbps.

Decoding the Performance Metrics

The evaluation of AI chip performance through direct comparisons becomes complicated because of the different benchmarking methods used. The primary metric Google uses to evaluate Ironwood’s performance is FP8 precision. The company states Ironwood “pods” achieve 24 times faster processing compared to similar parts of the top global supercomputers, but experts should consider this claim carefully because certain supercomputing systems do not include native FP8 hardware support.

Google’s direct performance comparisons excluded its TPU v6 (Trillium) hardware. According to Google, Ironwood provides double the performance per watt of the v6 predecessor. According to a company spokesperson, Ironwood has been positioned to replace the TPU v5p, but Trillium followed from the less capable TPU v5e. The peak performance of Trillium reached around 918 TFLOPS during operations at FP8 precision.

The Implications for the Future of AI

Despite the inherent complexities in benchmarking AI hardware, the underlying message is clear: Ironwood marks a major advancement in Google’s AI infrastructure technology. The new speed and efficiency improvements enhance the strong base, which enabled advanced models like Gemini 2.5 to develop quickly through previous TPU generations.

Google predicts that Ironwood’s advanced inference capabilities along with its improved efficiency will trigger transformative AI breakthroughs throughout the next year. Ironwood will enable Google’s “age of inference” vision by delivering essential computational power for advanced models and authentic agentic abilities that will make AI a more proactive and integrated element of our digital existence.