Google announced Ironwood as its seventh-generation Tensor Processing Unit (TPU), which functions as a custom-built engine to enhance Gemini models and introduce a new phase of AI functionality known as the “age of inference.” This new chip demonstrates a foundational change in Google’s AI system design because it was built to handle the sophisticated computing requirements of tasks involving simulated reasoning, which Google terms “thinking.” The company reveals Ironwood’s integral role in its infrastructure to work alongside advanced AI models for improved inference speeds and greater capacity to handle large amounts of contextual data.

Google expects Ironwood to enable more robust “agentic AI” functionalities. The concept defines AI systems that can independently perform tasks for users by collecting data and analyzing information to deliver appropriate results without direct commands. The company imagines future AI assistants that function more proactively and supportively, which Ironwood enables through its substantial computational power.

Ironwood showcases remarkable advancements in both performance capabilities and architectural design beyond any former Google TPUs. Next-generation AI demands enormous computational capacity, which this system meets by functioning through massive liquid-cooled clusters that consist of up to 9,216 distinct chips. The chips within this system have advanced connections that allow for rapid communication through the improved Inter-Chip Interconnect (ICI). The advanced interconnect technology provides essential support for efficient data transfer while reducing communication bottlenecks in large-scale distributed computing systems.

This scalable architecture will support both Google’s internal research and development work as well as external developers using Google Cloud Platform. Google plans to offer Ironwood in two distinct configurations: Ironwood will be offered in two hardware formats: one with 256 chips for smaller research operations and another with 9,216 chips to handle intensive AI tasks and operate large production environments.

When fully configured, Ironwood pods operate at maximum capacity, delivering a tremendous 42.5 Exaflops of inference computational power. These processing capabilities surpass any existing standards for complex AI applications. The peak throughput of each Ironwood chip reaches 4,614 TFLOPs, which represents a notable advancement beyond earlier TPU generations as specified by Google.

The memory capacity has been significantly upgraded to handle these enhanced processing requirements. The Ironwood chip contains 192GB of high-bandwidth memory, which represents six times the memory capacity available in the earlier Trillium TPU. The TPU’s enhanced on-chip memory capacity enables efficient storage and retrieval of larger datasets and model parameters while minimizing memory transfers to boost performance. The memory bandwidth now reaches 7.2 Tbps, which indicates a performance boost by 4.5 times, resulting in quicker data retrieval and processing operations.

Benchmarking and Contextualizing Ironwood’s Capabilities

Google offers context for Ironwood’s capabilities, which helps understand its performance due to the complexity of comparing AI hardware performance because different benchmarking methodologies exist. The company evaluates Ironwood’s performance primarily through its FP8 precision benchmarking standard. While Google asserts that Ironwood “pods” operate 24 times faster than similar segments of the world’s leading supercomputers, this claim requires careful evaluation because many top supercomputers lack native FP8 precision support in their architecture. The variation in hardware capabilities between devices creates discrepancies in both accuracy and relevance when making direct comparisons.

The direct performance comparisons provided omitted Google’s TPU v6 (Trillium) system. According to Google Ironwood outperforms TPU v6 (Trillium) by achieving double the power-efficiency which marks a substantial improvement in performance per watt. A Google spokesperson explained that Ironwood serves as the direct replacement for TPU v5p whereas Trillium followed the TPU v5e. Trillium reached its highest FP8 performance level at roughly 918 TFLOPS.

Next generation of Agentic AI systems.

The combination of Ironwood’s faster processing speed and larger memory capacity along with its superior power efficiency will play a critical role in transforming Google’s AI environment while promoting the creation of advanced AI applications. Ironwood builds on existing strong infrastructure behind advanced AI systems such as Gemini 2.5 to develop superior features for agentic AI.

These systems will have the ability to collect data from multiple sources and take appropriate actions without needing comprehensive instructions from human users. Google sees Ironwood as the powerhouse that will drive the next stage of more sophisticated and self-sufficient AI interactions which will establish the groundwork for major advances in natural language processing and machine learning to create more effective AI agents.