Reports have emerged that Google's parent company is working on a new server chip designed to significantly boost the efficiency of its Gemini AI model, leading to a rise in its share price.
The chip, known by the internal codename Frozen v2, aims to permanently etch a portion of the Gemini model's architecture directly onto the silicon.
According to internal engineering estimates, the new chip can reportedly process six to ten times more computational tokens per unit of power compared to Google's latest generation of Tensor Processing Units (TPUs).
Following an exclusive report on the development this past Monday, shares of Alphabet Inc climbed by 3%. The report detailed that Google is developing the server chip, internally called Frozen v2, with the primary goal of optimizing the operational efficiency of its Gemini series of large language models.
The media outlet explained that by permanently embedding parts of Gemini's foundational architecture into the chip's silicon, the design can reduce the amount of computation and data transfer required when processing AI requests.
Internal calculations by Google engineers reportedly show that, under the same power consumption, the Frozen v2 chip's ability to process tokens could be six to ten times greater than that of the company's current, most advanced TPUs. The Frozen series represents a specialized product line within Google's custom chip portfolio and is not intended to replace its general-purpose TPUs.
Google's current plan is to formally deploy this chip in 2028. The project was initiated to address a severe internal computing power shortage. This scarcity has not only intensified conflicts between different internal departments but has also reportedly forced Google Cloud to turn away external customer orders for computing capacity.
Just last month, Google was said to be paying nearly $1 billion per month to SpaceX to purchase computing power, a move aimed at bridging its capacity gap and fulfilling computing service commitments to enterprise clients.
This approach involves a trade-off in flexibility: if future versions of Gemini undergo changes to their underlying architecture, this chip would not be compatible with the updated models. The report states that Google currently views Frozen v2 as a technological pilot project, with production volumes not expected to match the scale of its TPU manufacturing.
A request for comment sent to the company has not yet received a response.
Comments