3 min Read
You don’t need to read an entire library to answer a specific question. You just need to know where to look. Following this principle, AI could be made up to 90 percent more efficient through compression techniques and smaller models. This is the conclusion reached by researchers at Saarland University.
Data centers consume vast amounts of energy, which has further increased with the adoption of AI and generative AI. According to Bitkom, the power demand of German data centers has more than doubled over the course of a decade – and this trend is set to continue sharply upward.
AI, in particular, is a significant energy consumer. Therefore, it will not only be crucial to generate more energy to satisfy this demand. It will be equally important to develop more energy-efficient hardware and algorithms.
The training of AI models like ChatGPT is particularly compute- and energy-intensive, consuming terawatt hours of electricity. The same applies to AI-generated text and image creation. With increasing demand, it will be necessary to build more and larger data centers, which require extensive cooling and contribute to higher CO2 emissions, potentially undermining the EU’s goal of achieving climate neutrality by 2050.
Smaller Models Are More Efficient
Professor Wolfgang Maaß from Saarland University and the German Research Center for Artificial Intelligence (DFKI), along with two colleagues, are pursuing a new approach to make AI significantly more energy-efficient. The team has been selected to present their research at the Federal Ministry for Economic Affairs and Climate Action’s stand at the Hannover Messe at the end of March and beginning of April 2025.
To curb AI’s energy appetite and conserve resources, the research team is developing leaner, more demand-oriented AI models. “By making the models smaller and more efficient, we contribute to greater sustainability on the one hand. On the other hand, this also opens up access to powerful AI models for medium-sized and smaller companies,” quotes the SME Blog team member Sabine Janzen.
“The smaller AI models no longer require extensive infrastructure. This makes them accessible to everyone, not just the big players,” says the doctoral scientist. She further explains her team’s approach: “We are working, among other things, with a technique called knowledge distillation. This is a kind of compression technology that allows us to make the models smaller. As a result, AI models consume significantly less energy for comparable performance.”
Knowing where, rather than knowing everything
The approach of researchers in Saarland and Kaiserslautern: You don’t need to read an entire library to find answers to specific questions. It’s enough to read only the books with the relevant answers – following the motto “knowing where.”
To achieve this, they extract small, focused, and efficient “student models” from large “teacher models.” This allows them to distill only the necessary knowledge for a specific task area and reduce it to the essentials. Thus, it would be possible to slim down language and data models by up to 90 percent.
“These student models deliver comparable performance but can potentially operate with up to 90 percent less energy,” says Janzen. Particularly with visual AI models, which process images, good results have already been achieved in saving energy using the method of “neural architecture search.”
“The approach of researchers in Saarland and Kaiserslautern: You don’t need to read an entire library to find answers to specific questions.”
Practical deployment shows initial successes
The research team has focused on the particularly energy-intensive machine learning with neural networks. The otherwise necessary labor-intensive manual work is automated here to help the AI find the best architecture itself. “We try out different network structures and optimize them to create a model with high performance, efficiency, and reduced costs,” says Janzen.
For the practical test, the team led by Maaß and Janzen is collaborating with Stahl Holding Saar. The task is to teach the artificial neural networks to sort steel scrap so that only the recyclable portion remains for the production of high-quality steels.
Together with partners, the Saarbrücken team is also working on a concept with “recommendations for action for sustainable data centers and energy-efficient AI” as well as a “tool that enables reliable forecasts of the exact energy consumption and associated costs of AI models,” as doctoral student Hannah Stein explains. Decision-makers should thus be better able to estimate which models consume how much energy.
Source header image: Pexels / Pixabay
