Water Footprint of Large AI Models

Artificial Intelligence (AI) has revolutionized the world by providing unprecedented benefits to industries ranging from healthcare to finance. However, a recent study by the University of Colorado Riverside and University of Texas at Arlington, titled “Making AI Less Thirsty,” sheds light on an aspect of AI that has remained under the radar – its water footprint.

Measuring the “water footprint” of AI models

The study aims to measure the “water footprint” of large AI models like OpenAI’s ChatGPT. It found that the water consumed to run ChatGPT, which is used by billions of users worldwide, is “extremely large.” The study highlights the importance of addressing water footprint of AI models in order to address global water crisis.

Withdrawal and consumption

The study distinguishes between “withdrawal” and “consumption.” Withdrawal refers to the physical extraction of water from rivers, lakes, and other sources, while consumption relates to the water loss due to evaporation when used in data centers. According to the recent study, the conversation comprising 20-50 questions with the AI chatbot in a single system may consume a “500 ml bottle of water.”

Water usage concentration

According to the researchers, when addressing consumption in this context, “water cannot be recycled,” which is where most of the AI’s water usage is concentrated. The water consumed also refers to clean freshwater used for generating electricity as well as for cooling servers to run AI models. The study claims that while training GPT-3 alone, Microsoft may have consumed a staggering 700,000 liters of water, which is the same amount used to make 370 BMW cars.

Water use in less energy-efficient data centers

The study highlights that if the data had been generated in the tech giants’ less energy-efficient data centers in Asia, water use may have increased threefold. The number is predicted to surge with the newly-launched GPT-4 AI system since it has a larger model size.




Leave a Reply

Your email address will not be published. Required fields are marked *