Google Developing 'Frozen v2' Chip for Gemini Efficiency Gains

Alphabet's Google is developing a new server chip designed to make its Gemini AI models more efficient, according to Tom's Hardware. The processor, known internally as 'Frozen v2', could deliver between six and ten times more tokens per unit of power than Google's current tensor processing units.

The chip would embed parts of Gemini's architecture directly into the silicon to reduce computational steps and data movement. Engineers aim to deploy it by 2028, focusing on improving inference performance while reducing power usage. The project originates from research led by DeepMind chief scientist Jeff Dean.

Unlike earlier versions that proposed baking model weights into silicon, Frozen v2 will retain flexibility by fixing only parts of the model architecture, allowing updates across future Gemini releases. The design remains experimental and will complement Google's existing TPU hardware instead of replacing it.

This chip development follows broader efforts by major AI companies to create specialized hardware that reduces dependence on suppliers such as Nvidia and addresses capacity constraints in AI infrastructure.