Representational image Business Insider India

A new study has found that trning AI models like and consume lakhs of tonnes of water.
The amount of water it took to train ChatGPT 3 is the same amount it takes to make 370 BMW and 320 Tesla electric cars.
The paper also notes that ‘ChatGPT needs to drink a 500 ml bottle of water for a simple conversation of roughly 20-50 questions and answers.

The world is talking about sustainability, and the other buzzing topic in the mainstream is generative AI. Now, what is the connection here? Well, there's a huge intersection of both, and a new study called "Making AI Less Thirsty" reveals comprehensive research and numbers about how much water is consumed when training large AI models like OpenAI's ChatGPT and Google's Bard.

Researchers from the University of Colorado Riverside and the University of Texas Arlington conduct the study. It measures and compares the environmental impact of AI training that requires huge amounts of constant electricity and tons of water. The water is used to cool the data centres, which is essential to keep them running.

According to the estimate of water consumption presented by researchers,

, which is partnered with OpenAI, consumed a whopping 185,000 gallons of water in training GPT-3 alone. To put things in perspective, it is equivalent to the amount of water needed to cool a nuclear reactor.

The research paper mentions that Microsoft used enough water to cool its US-based data centers while training GPT-3 that they could have produced 370 BMW cars or 320 Tesla electric vehicles. And that's just in the good old USA conditions. If they had done this training in their even larger data centers in Asia, those numbers would have been tripled. That's enough cars to start a whole new Microsoft-themed taxi service.

The paper also notes that "ChatGPT needs to 'drink' a 500 ml bottle of water for a simple conversation of roughly 20-50 questions and answers. While a 500ml bottle of water might not seem too much, the total combined water footprint for inference is still huge, considering ChatGPT’s billions of users."

Where from, where to?

The paper raises a lot of questions like - where does the water go? Does it disappear into thin air? Well, the water comes from rivers, lakes, and other freshwater sources. And there's a difference between "withdrawal" (physically taking water from a source) and "consumption" (when the water is lost through evaporation in the data center).

The study is around consumption, which is where most of the water use comes from. But on the brighter side, the water isn't gone forever. Instead, it gets released into the air through cooling towers, although it takes some time to return as rain.

But note that data centers can't just use any water. It has to be super clean and fresh to prevent corrosion and bacteria. And the centers are responsible for all the water used to produce the massive amounts of power they need. So who knew asking questions and taking help from ChatGPT could be such a thirsty business?

Cooling the air

Water is used in many ways to cool data centers that store artificial intelligence training data. The amount of water used depends on several factors, including the location and design of the data center, the type of cooling system used, and the region's climate. Water is used in two main types of cooling systems used in data centers —- air-cooled and water-cooled systems.

Air-cooled systems generally use fans to circulate air around the servers and equipment. In contrast, water-cooled systems use water to absorb heat from the equipment and transport it to an external cooling tower or chiller.

Water-cooled systems generally consume more water than air-cooled systems, but they are more efficient in terms of energy consumption and can provide better cooling performance. The amount of water used in water-cooled systems can vary widely, depending on the system's design and the climate. Some data centers use recycled or reclaimed water to minimize their water consumption.

It would be foolish to think that these tech giants aren’t aware of the excess water consumption required to run these systems, especially since it's not the first and only technology that demands water to keep data towers cool. But with the increasing reliance and prominence predicted around AI, it is necessary to act early and reduce the water footprint of these heavy systems before it raises a larger threat to the environment.

Search across your past Websets

Articles: water usage, AI, reputable publications

Search Criteria

Enrichments

Find more results