- Google is developing a new server chip integrating Gemini model blueprints directly
- The chip, called Frozen V2, aims to boost AI model efficiency by reducing data movement
- Frozen V2 could improve efficiency by six to eight times over current custom AI chips
Google is developing a new server chip that will directly integrate the blueprints of its flagship Gemini model, according to a report from The Information on Monday.
With the blueprints directly built into the chip, the amount of data movement is expected to be reduced, along with the number of decisions needed to be taken during inference, boosting the AI models' efficiency.
Informally named 'Frozen V2' it is expected to boost the efficiency of the company's in-house AI models. The company is looking to deploy the chip by 2028 at the earliest.
Shares of Google's parent company Alphabet were up 1% in pre market trade.
The chip could provide over six to eight times more efficiency compared to the company's latest custom AI chips that are out now, with regards to the number of AI tokens served for each unit of power that is consumed.
This could also help in making the company's power and computing infrastructure more cost-effective.
The firm is facing a dearth of availability of computing capacity for its expanding infrastrucuture, according to reports. This has also led to Google Cloud declining deals with select external companies.
ALSO READ: Google Poised To Lose Two More Senior AI Staffers To Anthropic
These new chips will be developed to be used alongside the company's Tensor Processing Units or TPUs, which are chips custom-built for AI and machine learning workloads.
The chips are likely being built specifically for Google's Gemini models, while the TPUs are expected to be used for more general purposes.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.