Google has introduced its seventh-generation Tensor Processing Unit (TPU), known as Ironwood, which signifies a major advancement in its proprietary AI hardware. Ironwood delivers the necessary computational power for advanced Gemini models through its targeted design, which enhances simulated reasoning abilities referred to as “thinking” by Google and enables more robust “agentic AI” applications that mark the emerging “age of inference.”

The company maintains that Gemini models’ performance depends directly upon its infrastructure through custom AI hardware, which boosts inference speed and enlarges context windows. The Ironwood TPU brings Google’s peak scalability and raw power to date and will enable AI systems to independently gather information and produce results for users in alignment with Google’s agentic AI objectives.

Ironwood provides a major throughput improvement over earlier models. Google intends to operate these chips in extensive liquid-cooled clusters that can hold up to 9,216 units. A newly advanced Inter-Chip Interconnect (ICI) allows these chips to exchange data efficiently at high speeds throughout the extensive system.

The powerful design serves purposes beyond Google’s internal operations. Developers looking to run demanding AI projects in the cloud will also be able to leverage Ironwood through two distinct configurations: Developers aiming to host complex AI workloads in the cloud have two options through Ironwood: either a 256-chip server or a full-scale 9,216-chip cluster.

Google’s Ironwood pods reach an incredible 42.5 Exaflops of inference computing power when fully configured. Google states that each Ironwood chip delivers a peak performance of 4,614 TFLOPs which demonstrates a substantial advancement from earlier generations. The latest TPUs from Google showcase a substantial improvement in memory capacity as each chip now holds 192GB which represents a sixfold increase from the previous Trillium TPU generation. The memory bandwidth rose significantly until it reached 7.2 Tbps which reflects a 4.5 times enhancement.

Google benchmarks Ironwood with FP8 precision, but direct comparisons to other AI hardware remain difficult due to different measurement methods. The company’s assertion about Ironwood “pods” running 24 times faster than the comparable parts of the world’s most powerful supercomputer needs careful evaluation because certain systems lack native FP8 support. Google’s direct comparison does not feature its TPU v6 (Trillium) hardware.

Google claims that Ironwood delivers double the performance efficiency compared to v6 based on power usage. The company representative explained that Ironwood serves as the direct successor to the TPU v5p model, but Trillium was developed as an upgrade to the weaker TPU v5e. At FP8 precision, Trillium achieved computational performance in the region of 918 TFLOPS.

Ironwood stands as a major progression for Google’s AI system in spite of benchmarking difficulties. Ironwood delivers exceptional speed and efficiency improvements beyond earlier TPUs while leveraging Google’s established strong infrastructure that supports fast advancements in large language models and simulated reasoning. The current market-leading Gemini 2.5 model from Google operates on the earlier generation of TPUs. Ironwood’s boosted inference speed and improved efficiency foreshadow new AI breakthroughs next year and mark the start of both the “age of inference” and more advanced agentic AI.