There is a simple way we can significantly reduce AI energy use

Being wiser with the AI ​​models we use to perform tasks could save 31.9 terawatt-hours of energy this year alone — the equivalent of the output of five nuclear reactors.

Thiago da Silva Barros At the University of Côte d’Azur in France, he and his colleagues looked at 14 different tasks for which people use generative AI tools, from text generation to speech recognition and image classification.

They then examined public leaderboards, including those hosted by machine learning center Hugging Face, to see how different models performed. The models’ energy efficiency during inference — when the AI ​​model produces an answer — was measured by a tool called CarbonTracker, and the model’s total energy use was calculated by tracking user downloads.

“Based on the size of the model, we estimated the energy consumption, and based on that, we could try to make our estimates,” says da Silva Barros.

The researchers found that across all 14 tasks, switching from the best-performing models to the most energy-efficient models for each task reduced energy use by 65.8%, while making the output only 3.9% less useful – a trade-off they suggest may be acceptable to the general public.

Because some people are already using the most economical models, if people in the real world switched from high-performance models to more energy-efficient models, they could achieve a 27.8 percent reduction in overall energy consumption. “We were surprised by how much money could be saved,” says the team member. Frederic Girouard At the French National Center for Scientific Research.

However, this requires change from both users and AI companies, da Silva Barros says. “We have to think in the direction of running smaller models, even if we lose some performance,” he says. “And companies, when they develop models, it is important to share some information about the model that allows users to understand and evaluate whether the model is consuming a lot of energy or not.”

Some AI companies are working to reduce the power consumption of their products through a process called model distillation, where large models are used to train smaller models. He says this is already having a big impact Chris Priest At the University of Bristol in the United Kingdom. For example, Google recently claimed Improved energy efficiency by 33x In Gemini over the past year.

However, pushing users to choose the most efficient models is “unlikely to limit the power surge from data centers as the authors suggest, at least in the current AI bubble.” says Priest. “Reducing the energy per claim will simply allow more customers to be served more quickly with more advanced consideration options,” he says.

“Using smaller models can certainly lead to lower energy consumption in the short term, but there are many other factors that need to be taken into account when making any kind of meaningful projections for the future,” he says. Sasha Lucioni In a face hug. It warns that rebound effects such as increased use “need to be taken into account, as well as broader impacts on society and the economy”.

Lucioni points out that any research in this area relies on external estimates and analysis due to a lack of transparency on the part of individual companies. “What we need, to do this kind of more complex analysis, is more transparency on the part of AI companies, data center operators, and even governments,” she says. “This will allow researchers and policy makers to make informed predictions and decisions.”