• Google Unveils New AI Processor: The seventh generation of Google's custom TPU architecture, Ironwood, is designed for Google's powerful Gemini models and represents a major shift towards "agentic AI" and the "age of inference".
• Infrastructure and Hardware: The model's capabilities depend on Google's infrastructure and custom AI hardware. Ironwood is the most scalable and powerful TPU yet, with higher throughput and more memory (192GB per chip, 6 times more than last-gen Trillium TPU) and increased memory bandwidth (7.2 Tbps, 4.5x improvement).
• Cluster and Configuration: Ironwood is designed to operate in clusters of up to 9,216 liquid-cooled chips and can be used in two configurations: a 256-chip server or the full-size 9,216-chip cluster. In its larger incarnation, it can generate 42.5 Exaflops of inference computing.
• Benchmark and Comparison: Measuring AI throughput is difficult, and Google is using FP8 precision as a benchmark. While it claims Ironwood "pods" are 24 times faster than comparable segments of the world's most powerful supercomputer, there are some limitations and its TPU v6 hardware is absent from the comparison chart. However, Ironwood is twice as powerful per watt compared to TPU v6.
• Impact on AI Ecosystem: Ironwood is a significant improvement for Google's AI ecosystem, enabling faster and more efficient AI. Google's existing infrastructure has led to rapid improvements in LLMs and simulated reasoning, and Ironwood sets the stage for more breakthroughs in the coming year with the market-leading Gemini 2.5 model running on last-gen TPUs currently.