Google Develops New AI Chip Frozen v2 for Gemini Efficiency Boost

Google is working on a specialized server chip named Frozen v2, aimed at enhancing the performance of its Gemini AI models. This chip is set to integrate part of Gemini's architecture directly into its hardware, which engineers project will provide a significant efficiency boost—between six to ten times the output of existing Tensor Processing Units (TPUs). Google plans to deploy this new hardware by 2028.

The move comes as Google faces increasing demand for AI processing power, which it has struggled to meet. Earlier in the year, Google informed Meta that it could not fulfill its request for additional compute power for Gemini, forcing Meta to limit its AI resource usage. This situation underscores the necessity for Google to enhance its infrastructure, leading to the decision to develop a chip dedicated solely to its AI models.

Unlike Google's traditional Tensor Processing Units, which are versatile and can run various AI models, Frozen v2 will hardwire specific elements of Gemini’s architecture into its layout, thereby eliminating redundant data processing and greatly improving energy efficiency. By integrating this architecture, the chip is expected to allow Google to handle more AI queries—potentially enabling the processing of ten queries for the same power cost as one.

While the specifics of this new chip remain limited, it represents a strategic effort by Google to bolster its technological capabilities in the competitive AI landscape, amidst pressures to scale its offerings effectively.