Key facts
- Google is developing a new server chip, codenamed 'Frozen v2,' to improve AI model efficiency.
- The chip aims to integrate Gemini model elements directly into hardware.
- The new chip is expected to be significantly more efficient than current AI hardware.
- Deployment is tentatively planned for 2028, though design is ongoing.
- This initiative is separate from Google's existing Tensor Processing Units (TPUs).
Google is reportedly developing a new server chip, internally referred to as 'Frozen v2,' designed to more efficiently serve its artificial intelligence models, including its Gemini family of large language models. The chip aims to incorporate elements of these AI models directly into the hardware, a move expected to enhance computing capacity and alleviate an ongoing AI computing crunch that has reportedly strained resources and led Google Cloud to decline deals with external clients.
The new chip is anticipated to be significantly more efficient, potentially serving six to 10 times more AI tokens per unit of power compared to Google's latest custom AI chips. Engineers are still finalizing the chip's design and the extent to which model information will be hardwired, with a potential deployment timeline set for 2028. This 'Frozen' project is intended to create a new line of homegrown chips that will complement, rather than replace, Google's existing Tensor Processing Units (TPUs).
Shares of Alphabet, Google's parent company, rose 3% in early trading following the report. This development comes amid broader industry efforts to create specialized hardware for AI workloads, driven by the increasing demand for computational power required to train and run advanced AI models.