量子位·2025-08-22 05:51Core Viewpoint - Google has effectively countered public concerns regarding the energy consumption of AI models, particularly its Gemini model, by presenting data that shows significantly lower energy usage and carbon emissions than expected [2][4][11]. Group 1: Energy Consumption and Emissions - A single query using Gemini consumes only 0.24 watt-hours (wh), emits 0.03 grams (g) of CO₂ equivalent, and uses approximately 5 drops of water [3][19]. - Over the past year, Gemini's energy consumption has been reduced to 1/33 of its previous level, and carbon emissions have decreased to 1/44, while still providing higher quality responses [6]. Group 2: Measurement and Methodology - Google emphasizes that many calculations regarding AI energy consumption reflect theoretical efficiency rather than actual efficiency during large-scale operations [8]. - The company has developed a comprehensive method to measure energy consumption, which includes factors such as idle machines, CPU and memory usage, data center overhead, and water usage [13][14][17][18]. Group 3: Efficiency Improvements - The efficiency of Gemini is attributed to a full-stack approach that includes custom hardware, efficient model architectures, and robust service systems [22][27]. - The latest TPU Ironwood has a performance efficiency that is 30 times better than the first publicly available TPU, significantly outperforming general-purpose CPUs in inference tasks [28]. Group 4: Data Center Operations - Google's data centers have an average Power Usage Effectiveness (PUE) of 1.09, making them among the most efficient in the industry [33]. - The company is committed to increasing the use of clean energy to achieve carbon-free operations and has optimized cooling systems to balance energy, water resources, and emissions [33].