Google builds custom Frozen v2 chips for Gemini AI
Custom Frozen v2 chips under development by Google will directly incorporate Gemini AI model design
Google is developing a new custom server chip internally codenamed Frozen v2 to run its Gemini AI models and boost processing capabilities. The initiative aims to integrate elements of the Gemini model architecture directly into the physical circuitry, rather than relying on general-purpose AI chips that continuously load models into memory and shuffle data back and forth.
The Information reported that the project aims to hardcode portions of the neural-network blueprint straight into hardware. This approach could make the chip six to ten times more efficient than Google's latest custom Tensor Processing Units (TPUs), measured by AI tokens served per unit of power. The company launched the effort partly due to internal AI computing capacity constraints that created friction and forced Google Cloud to turn away certain outside customer deals.
Engineers are currently finalising the design and determining the exact amount of model information to permanently hardwire into the silicon. The project aims to complement Google's existing TPU lineup rather than replace it, with targeted deployment planned as early as 2028. A Google Cloud spokesperson confirmed that company teams constantly research and experiment with innovations, noting that co-designing hardware and software from the ground up ensures systems remain integrated and highly optimised.
Bloomberg News separately reported last week that Google delayed launching its latest Gemini AI model after it fell short of internal performance goals. The technology company continues working to improve the model's capabilities, focusing particularly on coding performance.
The new Frozen v2 hardware represents a distinct set of homegrown chips operating alongside existing Google infrastructure. By hardcoding software elements into the hardware, the organisation seeks to relieve server pressure and optimise processing speeds across its enterprise cloud services. If successful, the deployment will mark a significant shift in how custom chips process complex artificial intelligence workloads.
