Google advances its artificial intelligence capabilities with the introduction of Ironwood, its seventh-generation Tensor Processing Unit (TPU). Google’s custom chip marks a significant progression in its hardware approach by creating solutions that exceed basic improvements and address the complex requirements of the latest Gemini models. Google designed Ironwood specifically to perform well in simulated reasoning tasks they call “thinking” and intends for it to initiate a groundbreaking phase in artificial intelligence.

Ironwood achieves its capabilities through significant improvements in performance together with revolutionary architectural design. Ironwood demonstrates superior throughput performance compared to previous TPU models while operating in expansive clusters that utilize liquid cooling. The newly enhanced Inter-Chip Interconnect (ICI) creates a network between clusters that house up to 9,216 individual chips to support rapid and efficient communication and data transfer. Google’s scalable architecture serves both its internal research teams and external Google Cloud developers through configurations that range from 256-chip servers to full 9,216-chip clusters.

Google expects that Ironwood’s superior speed, together with expanded memory capability and improved power efficiency, will create substantial changes within its AI ecosystem and enable major advancements. Ironwood will establish solid computational support for advanced AI models, which will result in innovation across multiple domains such as natural language processing and machine learning, and agentic AI creation. The upcoming AI systems will function proactively to collect data and analyze information before taking independent actions to serve users with very little direct instruction. Google pushes AI boundaries forward with Ironwood as a fundamental facilitator for this transformative progress.

The Driving Force Behind Ironwood

Through the creation of Ironwood, Google demonstrates its belief in the essential link between sophisticated AI models and specialized infrastructure. Google asserts that Ironwood extends beyond its speed capabilities because it forms a crucial element of its strategy to enhance inference performance and widen AI model context windows while tapping into “agentic AI” capabilities. Google’s introduction of “the age of inference” represents a transformation where AI systems will take proactive actions to assist users.

Core specifications demonstrate the computational power of Ironwood. The maximum performance of a completely set-up Ironwood pod reaches an extraordinary 42.5 Exaflops for inference calculations. The peak performance of each Ironwood chip reaches 4,614 TFLOPs, which represents a substantial advancement beyond previous TPU generations. Ironwood benefits from improved processing abilities through its greatly enhanced memory architecture. The memory capacity of each chip stands at 192GB of high-bandwidth memory, which represents a six times larger quantity than found on the Trillium TPU. The memory bandwidth has significantly improved to reach 7.2 Tbps, which represents an increase of 4.5 times the previous standard.

Google has shared performance benchmarks for Ironwood that use FP8 precision as the main evaluation metric. The company claims Ironwood “pods” provide 24 times faster performance than segments from the world’s most powerful supercomputers, but this statement requires careful consideration for context. Google recognizes that a number of these supercomputing systems lack inherent FP8 precision support, which impacts comparative analysis. Direct performance comparisons against Google’s TPU v6 (Trillium) were omitted from the analysis. According to Google, Ironwood delivers double the efficiency in performance per watt compared to Trillium, which shows notable improvements in energy efficiency. According to a Google spokesperson, Ironwood succeeded the TPU v5p whereas Trillium came after the TPU v5e. The Trillium TPU reached its peak FP8 performance at approximately 918 TFLOPS.