Google advances artificial intelligence through its latest seventh-generation Tensor Processing Unit (TPU), Ironwood. Google’s new custom chip represents a transformative step in its hardware development strategy by addressing the complex requirements of its advanced Gemini models. Ironwood has been specially designed to master simulated reasoning tasks referred to as “thinking” by Google and marks the beginning of a transformative period for AI.

The performance and design improvements drive the capabilities of Ironwood. Ironwood demonstrates much higher throughput capabilities than previous TPUs while being built to function within extensive liquid-cooled cluster systems. The newly enhanced Inter-Chip Interconnect (ICI) enables high-speed data transfer between up to 9,216 chips in each cluster. Google’s scalable architecture enables both its internal research and development operations alongside external Google Cloud developers to use system configurations that vary from servers with 256 chips to full clusters consisting of 9,216 chips.

Google projects that Ironwood’s improvements in speed and memory capacity, combined with power efficiency, will transform its AI ecosystem and establish new milestones for progress. Ironwood provides a strong computational base necessary for advanced AI models, which should lead to important discoveries across natural language processing, machine learning, and agentic AI development domains. The upcoming AI generation will function with increased proactivity by independently collecting data to analyze information and execute tasks for users while needing only minimal instructions. Ironwood operates as a crucial enabler for Google’s ongoing expansion of AI frontiers.

The Driving Force Behind Ironwood

Google’s creation of Ironwood demonstrates its strong belief in the essential connection between sophisticated AI algorithms and dedicated infrastructure systems. Google states that Ironwood serves as a foundational element for its strategy aimed at enhancing inference speeds while broadening AI model context windows to fully realize its “agentic AI” vision. Google refers to its new paradigm shift as the “age of inference” because it foresees AI systems that will actively perform tasks for users.

The core specifications of Ironwood reveal its impressive computational capabilities. The maximum performance of a completely set-up Ironwood pod reaches 42.5 Exaflops in inference computation. The computing capacity of each Ironwood chip reaches 4,614 TFLOPs, which represents a significant advancement compared to earlier TPU generations. The upgraded memory architecture in Ironwood provides essential support for its advanced processing capabilities. The Ironwood TPU chips contain 192GB of high-bandwidth memory, which represents six times more capacity than the Trillium TPU. The memory bandwidth now stands at 7.2 Tbps, which shows a 4.5 times improvement.

Google has published performance benchmarks for Ironwood using FP8 precision as the main measurement reference. The company claims Ironwood “pods” achieve 24-times faster performance than similar parts of the world’s top supercomputers, but this statement requires careful analysis. Google recognizes that not all supercomputing systems have built-in support for FP8 precision, which affects their comparative performance. The analysis did not contain direct performance comparisons between Ironwood and Google’s TPU v6 (Trillium). According to Google, Ironwood delivers double the performance per watt than Trillium showcases improved energy efficiency. According to a Google spokesperson, Ironwood succeeds the TPU v5p and Trillium comes after the TPU v5e. The highest FP8 performance achieved by Trillium reached around 918 TFLOPS.