Despite the complexities of benchmarking, the underlying message is clear: Ironwood stands as a major breakthrough for Google’s AI systems infrastructure. The improved performance and operational efficiency of Ironwood is built on the strong base that facilitated swift advancements in advanced models such as Gemini 2.5 which runs on older TPUs.

Google expects Ironwood’s improved inference capabilities and efficiency to enable groundbreaking developments in artificial intelligence throughout the upcoming year. Ironwood delivers essential computational power to support advanced models and agentic capabilities, which positions it as essential to Google’s “age of inference” vision where AI becomes a proactive and intelligent digital partner capable of true thought processes.

Decoding the Numbers: Ironwood’s Performance Context

Performance evaluations of various AI chips become challenging because benchmarking methodologies differ between systems. Google sets FP8 precision as the main benchmark standard for evaluating Ironwood. The claim by the company that Ironwood “pods” deliver 24 times the speed of segments from top supercomputers requires careful evaluation because many of these supercomputing systems lack native FP8 support.

Google excluded their TPU v6 (Trillium) from direct performance comparisons. Google asserts that Ironwood achieves double the performance per watt when compared to v6. According to a Google spokesperson, Ironwood serves as the TPU v5p successor while Trillium followed the less capable TPU v5e. Trillium reached its maximum performance capability of 918 TFLOPS with FP8 precision.

Inside Ironwood: A Performance Powerhouse

The processing capabilities of Ironwood show a significant improvement compared to earlier Google TPUs. The deployment plan proposes building large clusters that use liquid cooling to support networks with up to 9,216 separate Ironwood chips. A newly improved Inter-Chip Interconnect (ICI) will enable seamless communication between these extensive computational resources to maintain high-speed and efficient data transfer throughout the entire system.

Google’s internal AI research and development teams along with developers who use Google Cloud services will have access to this tremendous processing power. Ironwood will be offered in two configurations: The Ironwood system will come with a 256-chip server designed to handle moderate AI workloads as well as a 9,216-chip cluster designed to process the most demanding AI tasks.

A fully configured Ironwood pod achieves an astonishing computational capacity that reaches 42.5 Exaflops for inference computing. The latest generation of Google’s Ironwood TPU provides individual chips with a peak throughput of 4,614 TFLOPs, which represents a major advancement over prior TPU models. Each Ironwood chip packs 192GB of memory, which represents a sixfold expansion from what was available in the Trillium TPU. The memory bandwidth expansion by 4.5 times now enables a reach of 7.2 Tbps.

Google has just unveiled its latest innovation in custom silicon: Google showcases Ironwood as the seventh version of its Tensor Processing Unit (TPU) architecture. The new chip supports fast processing and is specifically engineered to meet the complex requirements of Google’s powerful Gemini models, which perform simulated reasoning referred to as “thinking” by Google.

The company routinely emphasizes the essential collaboration between its sophisticated AI models and its custom-built infrastructure. Ironwood stands as an essential element of Google’s strategy while delivering major boosts in inference speeds and expanded contextual information processing capabilities for these advanced models. Google presents Ironwood as its current leading TPU for scalability and power, which will enable AI to help users through autonomous data collection and output generation, forming Google’s “agentic AI” vision known as the “age of inference.”