At the Workshops & Tutorials Day of the Helmholtz AI Conference 2026 (HAICON26), Philipp Huber from the Junior Research Group Robust and Efficient AI (RAI) delivered an invited talk in the session “Current Status of the Benchmarking Field: Lessons Learned from the First Half of the UNLOCK Initiative.” The presentation focused on the growing importance of energy-efficient neural network training as AI models continue to increase in scale and computational demands.
Drawing on research recently presented at the SIAM Conference on Parallel Processing (SIAM PP), Philipp discussed how parallelization parameters in data-parallel training can significantly influence energy consumption and overall system efficiency. The talk also highlighted power fluctuations that arise from alternating phases within the training process, providing new insights into the complex relationship between training performance and powerdraw.
The findings contribute to ongoing efforts within the benchmarking community to develop more comprehensive evaluation methodologies that account not only for model accuracy and runtime but also for energy efficiency and sustainability.
Contact at SCC: Dr. Charlotte Debus