Timing Trick Cuts Energy Used in LLM Training by Up to 14 Percent
Researchers at the University of Twente have developed a method to reduce the energy consumption of training large language models by up to 14%. The technique, known as dynamic voltage-frequency scaling (DVFS), involves intelligently adjusting the clock frequencies of a GPU's computational core and memory. By fine-tuning these frequencies on a per-kernel basis, the researchers achieved significant energy savings without compromising training speed. AI
IMPACT Reduces the significant energy footprint of LLM training, potentially lowering costs and environmental impact.