<ul><li>Neural Architecture Search (NAS) aims to enhance deep learning by refining model architectures, but it leads to high energy consumption during the search process.</li><li>Energy-aware benchmarking through surrogate-based methods helps in estimating model quality with reduced training costs.</li><li>Three design principles for energy-aware benchmarks are proposed: reliable power measurements, wide GPU usage range, and holistic cost reporting.</li><li>Study highlights the impact of GPU measurement API on benchmark quality, emphasizing the importance of calibration experiments for improved energy reporting.</li></ul>

Guidelines for the Quality Assessment of Energy-Aware NAS Benchmarks

Discover more