Google is developing a custom chip called Frozen v2 that embeds elements of its Gemini AI model directly into silicon. The company says the chip will deliver 6 to 10 times the efficiency of its current TPU hardware. Deployment is expected as early as 2028.
What Frozen v2 does differently
Most AI chips run models as software on specialized processors. Frozen v2 takes a different approach by building parts of the Gemini model into the chip's physical design. That means certain calculations happen at the hardware level, cutting down on the energy and time needed to process AI tasks. By integrating model components into the chip, Google can reduce data movement, which often consumes more energy than the computation itself.
This is a shift from running Gemini on TPUs, which are optimized for a broader range of models. Frozen v2 is designed specifically for Gemini, making it a purpose-built accelerator for Google's latest large language model.
Efficiency gains over TPUs
Google's Tensor Processing Units are already among the most efficient AI chips on the market. Frozen v2 aims to improve on them by a factor of 6 to 10. That means a task that takes 10 units of energy on a TPU could take just 1 or 2 units on Frozen v2. For a company running millions of AI queries per day, the savings add up quickly.
The gains come from hard-coding parts of Gemini's architecture into the chip. This eliminates the overhead of loading model layers as software and reduces latency. Google has not disclosed specific benchmarks, but the efficiency target suggests a significant leap in performance per watt.
Deployment timeline and next steps
The chip is still in development. Google expects to deploy it as early as 2028, which gives the company several years to finalize the design, test prototypes, and prepare for mass production. The exact products and services that will use Frozen v2 have not been announced.
For now, Google is focused on getting the silicon right. The 2028 target leaves room for iteration and testing. The industry will be watching to see if Frozen v2 can deliver on its efficiency promises.
