New TPU Ironwood for AI
- SAN FRANCISCO (April 9, 2025) – Google announced its seventh-generation Tensor Processing Unit (TPU), named Ironwood, at the Cloud Next '25 event.
- Ironwood's capabilities extend to powering AI agents capable of performing tasks on behalf of users,marking a significant advancement in AI-driven automation.
- While Google has been developing TPUs for over a decade, previous generations primarily focused on training AI models.
Google unveils Ironwood TPU for AI Inference at Cloud Next ’25
SAN FRANCISCO (April 9, 2025) – Google announced its seventh-generation Tensor Processing Unit (TPU), named Ironwood, at the Cloud Next ’25 event. This new chip is designed specifically for the inference of artificial intelligence (AI) models, offering Google Cloud customers enhanced capabilities for AI applications.
Ironwood’s capabilities extend to powering AI agents capable of performing tasks on behalf of users,marking a significant advancement in AI-driven automation.
Ironwood: Key Features
While Google has been developing TPUs for over a decade, previous generations primarily focused on training AI models. Ironwood represents a shift, being the first TPU engineered specifically for inference – the process of applying trained AI models to new data.
The new TPU is designed to handle the computational demands of “thinking models,” which provide reasoning capabilities similar to those seen in models from the Gemini series and OpenAI. These models require significant parallel processing power and efficient memory access.
Google engineers designed Ironwood to minimize data transfer bottlenecks and reduce latency. Google Cloud clients can choose between configurations featuring 256 chips or a larger setup with 9,216 chips. The larger configuration delivers performance up to 42.5 Exaflops per pod.
To put that in outlook, Google claims this performance is 24 times greater than that of El Capitan, currently considered the world’s most powerful supercomputer, which clocks in at 1.7 Exaflops.
A single Ironwood chip boasts 4,164 TFLOPS of processing power. Each chip is equipped with 192 GB of HBM memory, a sixfold increase compared to the previous-generation TPU, Trillium. Memory bandwidth reaches 7.4 TBPS, 4.5 times greater than Trillium.
To manage the heat generated by this performance, Ironwood employs a liquid cooling system. This design also improves efficiency, doubling the performance per watt compared to Trillium.
Ironwood is slated to power current and future Gemini models. Google anticipates making Ironwood available to developers and Google Cloud customers by the end of the year.
