Google Develops 'Frozen v2' Chip to Embed Gemini AI Directly in Hardware, Report Says

Google is reportedly developing a new server chip that could embed parts of its Gemini artificial intelligence model directly into hardware, aiming to boost efficiency and address computing capacity constraints, according to a report by The Information.

The chip, internally codenamed 'Frozen v2', is designed to run Gemini AI models more efficiently by hardwiring certain components of the model into the chip's architecture. This approach could significantly reduce the computational power needed for AI tasks, making them faster and more energy-efficient.

The report says the company plans to deploy the chip as soon as 2028, but engineers are still finalising the design and determining exactly how much of the AI model will be hardwired. The chip is estimated to be six to ten times more efficient than Google's latest custom AI chips, measured in terms of AI tokens served per unit of power.

Google's move comes as the company faces limitations in AI computing capacity, which has reportedly prevented Google Cloud from entering into deals with outside customers for AI services. By embedding AI directly into hardware, Google hopes to alleviate these constraints and offer more powerful and efficient AI capabilities.

The development of 'Frozen v2' underscores the growing trend of custom silicon for AI workloads, as tech giants seek to gain a competitive edge by optimising hardware for their specific algorithms. Google has previously designed custom chips like the Tensor Processing Unit (TPU) for machine learning, but 'Frozen v2' represents a deeper integration of AI model and hardware.

The Information's report is based on sources familiar with the project, though Google has not officially confirmed the chip's development or specifications.