Google introduced its new seventh-generation Tensor Processing Unit (TPU), known as Ironwood, a custom chip designed to enhance its artificial intelligence capabilities significantly. The new architecture represents a strategic leap that directly addresses the changing needs of Google’s most advanced Gemini models instead of being just an incremental improvement. Ironwood’s engineering targets simulated reasoning tasks, which Google describes as “thinking.”

The company strongly emphasizes how their advanced AI models function together with their tailored infrastructure. The Ironwood platform embodies this design philosophy by delivering greatly improved inference speeds while enabling larger context windows for powerful models. Google maintains that Ironwood stands as its most scalable and powerful TPU to date while enabling an AI future that functions autonomously to gather data and produce results on behalf of users. Google’s vision for “agentic AI” centers on proactive user engagement, which Ironwood powers as the core of this inference-based future.

Performance Unleashed: Ironwood’s Impressive Specs

Ironwood delivers significantly higher performance throughput when compared to prior Google TPUs. The firm intends to deploy a large-scale network of clusters that will feature up to 9,216 liquid-cooled Ironwood chips operating together. The newly enhanced Inter-Chip Interconnect (ICI) allows the massive arrays to achieve seamless communication throughout the system while delivering high-bandwidth and low-latency data exchange.

Google’s internal teams, along with cloud developers, will gain access to this powerful processing capability. Ironwood will be available in two configurations: Ironwood technology will support small requirements through a 256-chip server and enable complex AI tasks with a powerful 9,216-chip cluster.

The sheer computational power of a full Ironwood pod is staggering: 42.5 Exaflops of inference computing. Google reports that individual Ironwood chips reach 4,614 TFLOPs of peak throughput showing a significant leap beyond earlier chip generations. The memory capacity of each chip has expanded dramatically to 192GB which represents a sixfold enhancement from the Trillium TPU. The system benefits from a 4.5 times increase in memory bandwidth now hitting 7.2 Tbps.

Contextualizing the Power: Ironwood’s Place in the AI Landscape

The assessment of AI chip performance proves difficult because multiple measurement methods are used across the industry. The standard benchmark for Ironwood performance evaluation by Google uses FP8 precision. The company asserts Ironwood “pods” surpass comparable segments from top supercomputers by 24 times, but users should be careful when interpreting these numbers because not all leading supercomputers have FP8 hardware support.

Google did not include TPU v6 (Trillium) in their direct performance comparison results. The company has indicated that Ironwood delivers double the performance per watt in comparison to v6. According to Google, a spokesperson, Ironwood succeeds TPU v5p and Trillium follows the TPU v5e. Trillium reached a maximum performance level of 918 TFLOPS when operating at FP8 precision.

The Road Ahead: Ironwood and the Future of AI

Despite the complexities of benchmarking, the message is clear: Google’s AI infrastructure experiences a major advancement with the introduction of Ironwood. Enhanced speed and efficiency in Ironwood extend the strong foundation that enabled rapid progress in Gemini 2.5 models, which run on previous TPU technology.

Google predicts that Ironwood’s advanced inference abilities and performance improvements will enable groundbreaking AI innovations throughout the following year. Ironwood delivers essential processing power for advanced models and agent-driven functions, which positions it as the central driver of Google’s “age of inference” vision that integrates AI deeper into our digital existence.