The seventh-generation Tensor Processing Unit (TPU) named Ironwood launches Google’s most advanced Gemini models into a new era of AI capability while marking the beginning of what Google calls the “age of inference.” Google’s new chip marks a major change in its AI infrastructure approach, which reflects its purpose-built design to handle difficult computational tasks such as simulated reasoning or what Google refers to as “thinking.” Ironwood serves as an essential infrastructure component that operates in tight coordination with advanced AI models to enhance inference speeds and improve the processing of extensive contextual data.

Google believes that Ironwood will facilitate more powerful “agentic AI” functionalities. This concept defines AI systems that autonomously perform tasks on behalf of users by collecting data and analyzing information to produce useful results without requiring direct instructions. The company’s vision includes AI assistants that can anticipate needs and provide help without waiting for instructions, and Ironwood delivers the necessary computational power to achieve this future.

The Ironwood system marks a major advancement in both performance capabilities and architectural design compared to Google’s past Tensor Processing Units. The system operates through liquid-cooled clusters containing up to 9,216 chips to meet next-generation AI’s substantial computational needs. The system connects these chips through a newly enhanced Inter-Chip Interconnect (ICI) which enables extremely high-speed communication. The advanced interconnect technology plays a vital role in facilitating efficient data exchange and reducing communication bottlenecks in large-scale distributed computing systems.

Google’s scalable architecture will support both its internal research and development operations as well as external developers who use Google Cloud Platform. Google plans to offer Ironwood in two distinct configurations: Google offers Ironwood in two forms: a 256-chip server suitable for smaller deployments and research tasks, as well as a complete 9,216-chip cluster designed to handle the most challenging AI workloads and operate extensive production systems.

Fully configured Ironwood pods deliver immense computational power, which scales up to 42.5 Exaflops of inference. The system enables an unparalleled processing power level suitable for intricate AI applications. Google’s specifications reveal that each Ironwood chip delivers a maximum throughput of 4,614 TFLOPs, which marks a substantial enhancement over former TPU generations.

The memory capacity received a substantial upgrade to meet the demands of increased processing needs. The new Ironwood chip carries 192GB of high-bandwidth memory, which represents a sixfold increase from the older Trillium TPU’s memory capacity. The expansion of on-chip memory provides TPUs with enhanced capabilities to manage larger data sets and model parameters efficiently while minimizing memory transfer requirements, which boosts performance. The memory bandwidth now reaches 7.2 Tbps, which represents a 4.5 times enhancement from the previous standard, thereby facilitating quicker data access and processing operations.

Benchmarking and Contextualizing Ironwood’s Capabilities

Comparing Ironwood to other AI hardware benchmarks is complex because of different evaluation methodologies, but Google offers context for its performance. FP8 precision serves as the main evaluation standard for Ironwood’s performance at the company. The statement about Ironwood “pods” delivering 24 times speed improvement compared to top supercomputers needs careful scrutiny because some supercomputers lack native FP8 precision support in their hardware design. Hardware capability differences impact how accurate and relevant direct performance comparisons can be.

The direct performance comparisons did not contain Google’s TPU v6 (Trillium). Google claims that Ironwood achieves double the power efficiency of its previous TPU v6 model known as Trillium which results in a substantial performance boost per watt. A Google spokesperson stated that Ironwood succeeds the TPU v5p directly while Trillium follows the TPU v5e as its next generation. The peak FP8 performance of Trillium reached about 918 TFLOPS.

Foundation for advanced Agentic AI capabilities.

Google’s AI ecosystem will experience major advancements through Ironwood because of its faster processing speed, larger memory capacity, and better power efficiency, which will enable the development of complex AI applications. Ironwood builds on existing strong infrastructure for advanced AI systems like Gemini 2.5 to enhance agentic AI capabilities.

These AI systems will be able to automatically collect information from multiple sources and use reasoning to create suitable responses or actions for users without requiring detailed instructions. Google considers Ironwood to be the catalyst that will drive this new era of smarter and more independent AI interactions which will lead to major advancements in natural language processing and machine learning as well as the creation of more capable and beneficial AI agents.