Google is building a chip with Gemini baked into the silicon (thenextweb.com)

Google is reportedly developing a groundbreaking AI chip, informally called “Frozen v2,” which aims to integrate the architecture of its Gemini AI model directly into the silicon. This innovative approach, which could make the chip 6 to 10 times more efficient than Google’s current custom AI chips, responds to significant AI capacity constraints within the company. By hardwiring Gemini’s neural network architecture into the hardware, Google seeks to eliminate the overhead associated with general-purpose chips that require constant model loading, potentially revolutionizing AI infrastructure. The deployment of this technology is anticipated by 2028. The significance of Frozen v2 lies in its potential to drastically improve performance and reduce operational costs for AI applications, particularly in real-time scenarios like voice assistants. While the efficiency gains are compelling, there are risks; a fixed architecture may become outdated as AI evolves. However, updates to model weights could mitigate some of this rigidity. This move signals a shift in the AI landscape towards custom silicon solutions tailored to specific models, suggesting that competitors will need to innovate quickly to keep pace if Google achieves success with this project.