Google has revealed its seventh-generation Tensor Processing Unit (TPU) called Ironwood because of its ongoing commitment to AI progress. Google’s groundbreaking custom chip marks a major transformation in hardware strategy because it surpasses small improvements to fulfill the intricate requirements of its advanced Gemini models. Ironwood delivers exceptional performance in simulated reasoning tasks, which Google defines as “thinking” and stands ready to usher in a new age of AI development.

The creation of Ironwood demonstrates Google’s understanding of how advanced artificial intelligence models need purpose-built infrastructure to function effectively. Google positions the Ironwood chip as a vital component of its strategic plan to boost inference speeds while expanding AI model context windows to fully realize the capabilities of “agentic AI.” Google describes this new paradigm as the “age of inference” because it foresees AI systems that will take proactive actions to assist users.

Ironwood showcases a major advancement in performance combined with innovative architectural design. Ironwood achieves substantially enhanced throughput compared to earlier TPUs while being engineered for use inside massive clusters that use liquid cooling. The new Inter-Chip Interconnect (ICI) connects clusters of up to 9,216 individual chips to allow efficient high-speed data communication. Google’s scalable architecture serves both its internal research teams and external developers using Google Cloud across configurations, which extend from 256-chip servers to comprehensive 9,216-chip clusters.

Ironwood’s Technical Specifications

The power of Ironwood’s computational abilities becomes evident when you look at its fundamental specifications. Ironwood pods at full configuration deliver an impressive 42.5 Exaflops of inference computing performance. Ironwood chips achieve peak performance levels of 4,614 TFLOPs, which represents major progress beyond earlier TPU generations. Ironwood maintains its enhanced processing capabilities through a substantially improved memory architecture. The 192GB high-bandwidth memory installed in each chip represents a sixfold enhancement compared to the Trillium TPU’s memory capacity. The memory bandwidth now stands at 7.2 Tbps, which represents an improvement of 4.5 times.

Google released benchmarks to evaluate Ironwood’s performance based on FP8 precision as the primary metric. Although Ironwood “pods” reportedly deliver a 24-fold performance boost over the world’s leading supercomputers, according to the company, this statement requires careful analysis. Google recognizes that some supercomputing systems lack native FP8 precision support, which impacts benchmark comparisons. Direct performance comparisons with Google’s TPU v6 (Trillium) were not part of the evaluation. Google confirms that Ironwood delivers double the power efficiency performance compared to Trillium, which reflects advancements in energy efficiency. According to a representative from Google, Ironwood succeeds the TPU v5p and Trillium succeeds the TPU v5e. Trillium reached a peak FP8 performance level of 918 TFLOPS.

Ironwood’s influence extends past simple performance metrics. Google expects Ironwood’s improved processing speed alongside its superior memory capacity and power efficiency to transform its AI ecosystem significantly. Ironwood will enable the creation of advanced AI models, which are projected to produce new discoveries in natural language processing and machine learning domains while advancing agentic AI development. The future generation of AI systems will function proactively with the ability to independently collect information and make decisions while representing users with minimal direct guidance. Google’s AI innovations move forward with Ironwood acting as a crucial component of their transformative progress.