Sustainability is a hot topic worldwide, and generative AI is also a popular conversation starter. What does this have in common, exactly? There is a significant overlap between the two, and a recent study titled “Making AI Less Thirsty” provides extensive analysis and data on how much water is used while training powerful AI models such as ChatGPT and Bard.
Everything you need to know about ChatGPT’s Consumption!
The study is being conducted by academics from the University of Texas at Arlington and the University of Colorado Riverside. It gauges and compares how much water and power are used during AI training, which is extremely energy intensive. In order to keep the data centers operating, water is needed to cool them. Microsoft, which partners with OpenAI, used 185,000 gallons of water to train GPT-3, researchers claim. It is equivalent to cooling a nuclear reactor with water.
The research article claims that when training GPT-3, Microsoft consumed as much water to cool its US-based data centres as it would have taken to make 370 BMW cars or 320 Tesla electric vehicles. These figures would have tripled if they had done this training in their even larger data centers in Asia. In addition, the paper mentions “For a simple conversation of 20-50 questions and answers, ChatGPT needs to ‘drink’ a 500 ml bottle of water. While a 500ml bottle of water may not seem like much, considering ChatGPT’s billions of users, the total combined water footprint for inference is still massive.”
The paper raises many questions, such as where the water goes. Does it just vanish into thin air? The water comes from rivers, lakes, and other sources of freshwater. There’s also a distinction to be made between withdrawal and consumption. The research focuses on consumption, which accounts for the majority of water use. On the plus side, the water isn’t gone for good. Instead, it is released into the atmosphere via cooling towers, where it eventually returns as rain.
But keep in mind that data centres cannot simply use any water. To avoid rust and bacteria, it must be extremely clean and fresh. Also, the centres are in charge of all the water utilized to generate the enormous amounts of power they require. Who would have thought that requesting information fromChatGPT and receiving assistance could be such a thirsty endeavour?
Data centers that house training data for artificial intelligence are cooled in a variety of methods using water. The location and layout of the data center, the choice of a cooling system, and the temperature of the area all affect how much water is used. The two main cooling methods used in data centers, air-cooled and water-cooled systems, both require water.
Fans are typically used in air-cooled systems to move air around the servers and other equipment. As opposed to air-cooled systems, water-cooled systems use water to absorb heat from the machinery and transfer it to a chiller or external cooling tower.
Although water-cooled systems use more water than air-cooled systems do, they are more energy-efficient and capable of performing better cooling. Depending on the system’s design and the region, the amount of water needed in water-cooled systems might vary significantly. To reduce water usage, several data centers employ recycled or reclaimed water.
Given that this technology is not the first or the only one to require water to keep data towers cool, it would be naive to assume that major tech companies are unaware of the excessive water usage necessary to operate these systems. Yet, with the rising reliance and importance of AI projected, it is imperative to take immediate action and lessen the water footprint of these huge systems before it poses greater harm to the environment.
