Google is working on a new server chip designed to be closely tailored to the Gemini models. The project, informally known as ‘Frozen v2’, could be deployed in data centres by 2028. However, the work is still at an early stage, and the final design of the processor has not yet been finalised.
The most significant change is expected to be the storage of some of the model architecture information directly within the integrated circuit. This would mean the chip would not have to perform all operations in the same way as more general-purpose accelerators. According to unofficial data, Frozen v2 could process between six and ten times more tokens per unit of energy consumed than Google’s current AI chips.
For the company, this would mean the ability to deliver Gemini-based services more quickly and cost-effectively. This is significant because the development of generative artificial intelligence requires an ever-increasing number of servers, energy and specialised processors. According to reports, a shortage of available computing power has already forced Google Cloud to turn down some contracts with external clients.
Frozen v2 is not intended to replace the TPU processors that have been in development for years. It is intended to be an additional, more specialised solution designed primarily to run Google’s models. The company is currently rolling out Ironwood chips, designed for both training models and their subsequent deployment. Google states that Ironwood offers more than four times the performance per chip compared to the previous-generation Trillium.
This new project is part of a wider trend among the largest technology companies to develop their own chips. This helps to reduce costs and dependence on external suppliers, primarily Nvidia. However, the price to pay for very high performance may be reduced flexibility, as hardware tailored to a specific model version is more difficult to utilise following a significant change to its architecture.
Google has not officially confirmed plans to roll out Frozen v2. The company has merely stated that it is constantly testing new solutions, and not every project makes it into production. Following the publication of this information, Alphabet’s share price rose by around 3 per cent, indicating that investors viewed the potential to curb rising AI infrastructure costs positively.
