Support our independent tech coverage. Chrome Unboxed is written by real people, for real people—not search algorithms. Join Chrome Unboxed Plus for just $2 a month to get an ad-free experience, access to our private Discord, and more. Learn more about membership here.
START FREE TRIAL (MONTHLY)START FREE TRIAL (ANNUAL)
While we usually focus on the gadgets in our hands, the real magic of the AI era happens in the massive data centers powering them. This week, Google officially launched Ironwood, its seventh-generation Tensor Processing Unit (TPU), and the specs are absolutely staggering.
Google calls Ironwood its “most powerful, capable, and energy-efficient TPU yet,” and it represents a massive leap forward for the hardware that underpins everything from Gemini to Google Search.
Built for the “Age of Inference”
The biggest shift here is the focus. While previous chips were heavily focused on training AI models, Ironwood is purpose-built for inference—the actual moment when you ask an AI a question and it generates an answer.
As you move from building models to having people actually using them in daily life, this speed is critical. Ironwood delivers more than 4X better performance per chip for both training and inference compared to the previous generation. This means faster, smoother, and more responsive AI interactions for everyone.
A massive supercomputer network
These chips don’t work alone; instead, they are designed to be clustered into massive “superpods.” With Ironwood, Google can link up to 9,216 chips in a single domain.
These chips are connected by a lightning-fast network running at 9.6 Tb/s, allowing them to access a mind-boggling 1.77 Petabytes of shared High Bandwidth Memory. This massive connectivity removes the data bottlenecks that usually slow down complex AI tasks, making the training and running of cutting-edge models significantly more efficient.
Chips designed by chips
Perhaps the most meta part of this announcement is how Ironwood was created. Google researchers actually used AI to design the chip itself. Using a method called AlphaChip, Google utilized reinforcement learning to generate superior layouts for the chip, continuing a cycle where AI helps build better hardware, which in turn powers better AI.
Ironwood is available now for Google Cloud customers, marking another huge step in Google’s full-stack approach to dominating the AI landscape.
SUBSCRIBE TO UPSTREAM
Get Chrome Unboxed delivered straight to your inbox
Upstream is our flagship, curated newsletter with the top stories, most click-worthy deals, giveaways, and trending articles from Chrome Unboxed sent directly to your inbox a few times a week. Join 31,000+ subscribers.
