Almost everyone uses artificial intelligence in one way or another these days. What not everyone realizes is that even the smallest response from a chatbot consumes energy and resources. Training and operating AI models with massive amounts of data consumes hundreds of terawatt-hours worldwide. On a global scale, all of this leaves a massive ecological footprint. And with the rapid development of the technology, demand will rise sharply. New data centers will seal off vast areas of land, as will new power plants—including those running on fossil fuels—that must supply the additional energy required. CO2 emissions will rise just as dramatically as the water demand for cooling.

A consortium led by Professor Wolfgang Maaß, who conducts research at Saarland University and the German Research Center for Artificial Intelligence (DFKI), has made it its mission to counteract this trend and make artificial intelligence more energy-efficient. For three years, the consortium has been researching various methods and developing and testing new technologies. The researchers will present their findings at the project’s conclusion.

Compressed AI requires nearly 90 percent less energy

The team is focusing, on the one hand, on smaller, more needs-based AI models to curb AI’s energy consumption and conserve resources. Today, AI uses massive data models. A chatbot, for example, uses the entire data model—with trillions of parameters—to generate its response: In a figurative sense, it searches an entire library instead of just the books with relevant content. The researchers have therefore developed AI models that skip processing irrelevant parameters altogether, making them more energy-efficient.

To do this, they filter the knowledge actually needed for each specific task from large teacher models and create customized student models that are up to 90 percent smaller. “We’re achieving good results by compressing the AI models—that is, making them smaller and more efficient. In our test runs, we were able to show that the student models deliver comparable performance but use up to 89 percent less energy,” says Sabine Janzen, a postdoctoral researcher on Wolfgang Maaß’s team. The leaner AI models, tailored for specific use cases, operate without requiring extensive infrastructure. “This makes powerful AI models accessible even to small and medium-sized enterprises, which was previously impossible simply because of the models’ size,” says Janzen.

Automatically Finding the Best AI Model While Using 40 Percent Less Energy

For AI models that process and generate digital image data, the researchers are using a different method known as “neural architecture search.” This process automatically identifies the best architecture for artificial neural networks. The team was able to demonstrate that they can reduce the size of the models by nearly 90 percent and cut energy consumption by 40 percent. And the models don’t sacrifice any performance in the process. “We were even able to improve the model’s accuracy with this approach,” says Sabine Janzen.

In machine learning with artificial neural networks, the learning processes unfold similarly to those in the human brain. While the human brain is a master of energy efficiency—it has continuously optimized itself through evolution and processes information very efficiently—its artificial counterpart, with its efficient algorithms, requires an enormous amount of computing power and electricity: Artificial neural networks are still painstakingly assembled by humans today and fine-tuned until they deliver good results. “We automate this process using neural architecture search. In this process, we test various

Scrap Sorting Test Case Highly Transferable

To test such more energy-efficient AI methods on an industrial scale, researchers have been collaborating with SHS – Stahl-Holding-Saar since 2022 as part of the Escade project. SHS developed a highly capable AI model that automatically classifies steel scrap and uses camera images to identify which type of steel scrap is being delivered to the steel mill site. Since 2024, SHS has been using its in-house developed, optimized AI model to classify steel scrap: Using camera images, the AI identifies the different types of scrap, allowing the steel scrap to be sorted by type before being used in production.

The goal of the Escade project was to make an AI model with roughly equivalent performance as energy-efficient as possible. To this end, the research team developed a powerful visual AI model for steel scrap classification. This new model was compressed so that it is compact, energy-efficient, and performs similarly to an AI model created using traditional methods. This makes the visual computing process more energy-efficient. To achieve this, the partners first trained the model using the complete dataset and all available information, and then compressed it using knowledge distillation and automatically assembled neural networks. The techniques developed can be applied to other visual AI models as well, thereby contributing to energy savings on a broader scale.

Further savings potential for sustainable data centers

In addition, the consortium identifies potential savings for data centers. Together with its partners, the Saarbrücken-based research team developed a concept and recommendations for energy-efficient AI that enable data centers and AI users to plan more effectively and identify inefficient processes. To this end, the team developed a tool that enables reliable forecasts of the exact energy consumption and costs of AI models. “This tool makes it possible, for example, to schedule processes that require high computing power when electricity prices are low. Until now, decision-makers have found it difficult to estimate how much energy specific models will consume, making it hard to plan economically,” explains doctoral candidate Hannah Stein, who is researching energy-efficient AI methods.

Neuromorphic chip technologies are proving to be promising

In addition, the research consortium is working on neuromorphic chip technologies in the field of hardware—microprocessors that also mimic the functioning of the human brain. “Our findings in this hardware area suggest that this technology is also significantly more energy-efficient than conventional chips: In our tests, they are already up to six times more efficient. But further research is needed here, and we need more time for that,” explains Sabine Janzen.

Background

The ESCADE (Energy-Efficient Large-Scale Artificial Intelligence for Sustainable Data Centers) project received approximately five million euros in funding from the Federal Ministry of Research, Technology, and Space (BMFTR) as part of the “Green Tech Innovation Competition” for a three-year period.

In addition to the research team led by Professor Wolfgang Maaß as coordinator (German Research Center for Artificial Intelligence DFKI and Saarland University), the project consortium includes the Technical University of Dresden, Bielefeld University, the Central German Data Center NT Neue Technologie AG (NT.AG), SHS - Stahl-Holding-Saar GmbH & Co. KGaA, SEITEC GmbH, the Austrian research organization Salzburg Research, and the subcontractors eco2050 Institute for Sustainability, SpiNNcloud Systems GmbH, and elevait GmbH & Co. KG.

Video