The seventh-generation Tensor Processing Unit (TPU) called Ironwood shows Google’s continued advancement in artificial intelligence. The custom-designed chip represents Google’s strategic hardware advancement by delivering comprehensive enhancements to meet the complex needs of its top-tier Gemini models. Ironwood has been purposefully designed to dominate simulated reasoning tasks identified by Google as “thinking” and will establish the foundation for a new age in artificial intelligence.

Ironwood achieves its capabilities through substantial advancements in performance paired with architectural design enhancements. Ironwood achieves much higher throughput than its prior TPU models while operating in vast liquid-cooled cluster environments. A newly improved Inter-Chip Interconnect (ICI) connects 9,216 individual chips within these clusters to enable rapid and effective communication and data transfer. The scalable architecture enables Google’s internal R&D teams and external Google Cloud developers to utilize systems that vary from servers with 256 chips up to clusters containing 9,216 chips.

Google anticipates that Ironwood’s superior speed, memory capacity, and power efficiency will create transformative effects throughout its AI ecosystem to enable substantial advancements. Ironwood delivers an essential computational platform for complex AI models, which will enable significant discoveries in natural language processing, machine learning, and the creation of agentic AI systems. The upcoming AI generation will function proactively while collecting data independently and reasoning about information to perform user tasks with little direct instruction. Through its pioneering AI advancements, Google achieves success with Ironwood acting as a crucial facilitator for this transformation.

The Driving Force Behind Ironwood

Google’s development of Ironwood demonstrates the company’s deep belief that purpose-built infrastructure is essential to the advancement of cutting-edge AI models. Google describes Ironwood as a fundamental component of its strategy, which aims to boost inference speeds while enlarging AI model context windows to achieve the full capabilities of “agentic AI.” Google refers to this transformative period as the “age of inference” because it expects AI systems to take proactive actions on behalf of users.

The fundamental specifications of Ironwood demonstrate its computational power. The fully configured Ironwood pod reaches an extraordinary peak of 42.5 Exaflops in inference computing performance. Individual Ironwood chips achieve a peak throughput of 4,614 TFLOPs, which demonstrates notable advancement beyond earlier TPU generations. Ironwood includes a substantially enhanced memory architecture to accommodate its improved processing abilities. The new chip possesses 192GB of high-bandwidth memory, which represents a six times increase from the capabilities of the Trillium TPU. Memory bandwidth reached 7.2 Tbps, which represents a 4.5x improvement.

Google released benchmarks to contextualize Ironwood’s performance metrics through FP8 precision testing. The company claims that Ironwood “pods” deliver 24 times faster performance than equivalent parts of the world’s top supercomputers, but this statement requires careful interpretation. Google recognizes that certain supercomputing systems lack native FP8 precision support, which affects the outcome of performance comparisons. The benchmark results did not feature direct performance comparisons between Ironwood and Google’s TPU v6 (Trillium). Google claims that Ironwood delivers double the performance per watt compared to Trillium, which demonstrates its superior energy efficiency. Ironwood replaces TPU v5p as Google’s next-generation supercomputer while Trillium follows TPU v5e. The highest FP8 performance level for Trillium reached around 918 TFLOPS.