The seventh generation Tensor Processing Unit from Google, Ironwood by code name, marks a major development in specialized AI hardware that will drive future artificial intelligence applications. The new chip has been purpose-built to handle the sophisticated computational needs of Google’s leading Gemini models with a focus on simulated reasoning or “thinking” capabilities.

Google has made the combination of tailored hardware with cutting-edge AI models fundamental to its AI development approach. Ironwood serves as a critical component to boost inference speeds and extend context windows within these high-performance AI models. Google presents Ironwood as its top scalable and potent TPU ever built, while seeing it as the essential component to enable advanced “agentic AI” capabilities, which marks a new era called the “age of inference” where AI takes proactive actions for its users.

The Architecture and Performance of Ironwood

Throughput performance of Ironwood has improved substantially over its earlier versions. Google plans to implement these chips within massive liquid-cooled systems containing up to 9,216 units. A new advanced Inter-Chip Interconnect (ICI) enables direct communication between interconnected chips to achieve fast and effective data transfer throughout the system.

The durable infrastructure supports Google’s internal artificial intelligence projects along with external developers who use Google Cloud. Ironwood will be available in two configurations: Ironwood will be offered as a 256-chip server for limited environments alongside a complete 9,216-chip cluster designed to handle intense AI applications.

The Ironwood pod achieves tremendous computational power with its peak capability of 42.5 Exaflops in inference computing. Google’s specifications reveal that each Ironwood chip reaches 4,614 TFLOPs peak throughput, which represents a major advancement from earlier TPU models. Each chip now features a memory capacity of 192GB, which is six times more than what the Trillium TPU offers. The memory bandwidth now stands at 7.2 Tbps, which demonstrates a 4.5 times enhancement.

Understanding the Benchmarks

Direct comparisons of multiple AI chips remain difficult because benchmarking methods differ significantly. Google has chosen FP8 precision as the performance standard for its latest TPU. The company’s claim that Ironwood “pods” are 24 times faster than similar parts of the world’s top supercomputers requires careful analysis because certain supercomputing systems lack native FP8 hardware support.

The direct performance comparison between AI chips omitted Google’s TPU v6 (Trillium) model. According to Google, Ironwood delivers double the energy-efficient performance of TPU v6. Google’s company spokesperson noted that the Ironwood model succeeds TPU v5p while Trillium was built after the less powerful TPU v5e. Trillium achieved peak performance of around 918 TFLOPS when processing FP8 precision tasks.

The Implications for the Future of AI

Despite the complexities inherent in benchmarking AI hardware, the underlying message is clear: Ironwood marks a major advancement in Google’s AI infrastructure. The previous generation of TPUs enabled rapid developments in models like Gemini 2.5 through its strong foundation, which the enhanced speed and efficiency of Ironwood now build upon.

Google expects Ironwood’s improved inference performance and efficiency to lead to major advancements in AI throughout the upcoming year. Ironwood delivers essential processing power to support advanced model development and genuine agentic capabilities, making it key to Google’s vision for the “age of inference,” where AI becomes an active and essential element of our digital existence.