"Joule", a sister journal of "Cell", published a paper this week called "The Growing Energy Footprint of artificial intelligence."
The paper was published by Alex De Vries, founder of the scientific research institution Digiconomist, which claimed that AI inference performance may consume a large amount of electricity in the future, and by 2027, AI's electricity Consumption may be equivalent to that of the Netherlands for a year. of electricity usage.
Alex De Vries said that the outside world has always believed that training an AI model is "the most power-consuming stage of AI." However, Vries cited reports from SemiAnalysis and Google and pointed out that the "inference link" consumes more power. From 2019 to 2021, An estimated 60% of AI-related energy consumption comes from model inference.
Alex De Vries also calculated based on the 18.3 TWh power consumption disclosed by Google in 2021, claiming that AI accounted for 10%-15% of Google’s overall power consumption at that time. After Google fully implemented “AI search”, the company The power consumption for AI may reach up to 27.4 TWh, which is close to the electricity used by Ireland in an entire year (29.3 TWh).
Looking at the entire market, as the new CoWoS factory invested by NVIDIA's partner TSMC is expected to be mass-produced in 2027, Alex De Vries also estimates that the total power consumption of all AI servers delivered by NVIDIA will reach 85-134 TWh, equivalent to one year's electricity use in the Netherlands.
Alex De Vries believes that the AI industry should improve hardware efficiency and model calculation efficiency to improve the power consumption of the industry. He also suggests that developers should not only focus on optimizing AI but also consider the necessity of using AI to reduce the power consumption of the AI industry. energy consumption costs .
refer to
The growing energy footprint of artificial intelligence
