Publication
Measuring and Comparison of the Energy Consumption of Different Machine Learning Methods
Christoph Tholen; Carolin Leluschko; Lars Nolle; Frederic Theodor Stahl
In: Max Bramer; Frederic Theodor Stahl (Hrsg.). Artificial Intelligence XLII. SGAI International Conference on Artificial Intelligence (AI-2025), Cham, Pages 172-184, ISBN 978-3-032-11402-0, Springer Nature Switzerland, 2026.
Abstract
The increasing use of artificial intelligence (AI) is driving up energy demand, particularly during the training of complex models. While existing approaches to measuring energy consumption are mostly based on software estimations, this paper presents a hardware-based method for directly measuring the energy usage of AI applications. Using a Shelly Pro 1 PM device, energy consumption is recorded independently of hardware, operating system, or programming language. The method is evaluated across three use cases: (1) regression of photosynthetically available radiation, (2) plastic litter detection in remote sensing imagery, and (3) classification of Higgs boson data. In the first use case, linear regression consumed the least energy (mean: 92.48 Ws), while an artificial neural network (ANN) consumed 320.46 Ws on average. In the second use case, MobileNetV2 required 343,499 Ws per training run, whereas EfficientNet v2 s consumed 1,174,249 Ws. In the third use case, the proposed method measured an average of 1,227.1 Ws per run, while CodeCarbon overestimated the same process at 5,468.6 Ws. These results highlight significant discrepancies between models and between measurement methods, offering a more accurate foundation for sustainable AI deployment.