Google has introduced its seventh-generation custom TPU architecture, codenamed Ironwood, designed to accelerate agentic AI capabilities. The chip is built on top of simulated reasoning and “thinking” capabilities, allowing it to proactively gather data and generate outputs.
The Ironwood processor boasts higher throughput compared to previous Google Tensor Processing Units (TPUs) and can be used in clusters of up to 9,216 liquid-cooled chips. This design enables seamless communication between chips through a newly enhanced Inter-Chip Interconnect (ICI).
Google plans to use Ironwood in its Gemini models, as well as make it available to developers for cloud-based AI projects. The processor will be available in two configurations: a 256-chip server and the full-size 9,216-chip cluster, expanding context windows and accelerating inference.
Ironwood represents a major shift towards agentic AI capabilities, allowing AI systems to act autonomously on behalf of users. This development is expected to unlock more powerful AI capabilities for Google’s Gemini models and developers alike.
Source: https://arstechnica.com/gadgets/2025/04/google-unveils-ironwood-its-most-powerful-ai-processor-yet