Researchers slash LLM energy use by 90% for code tasks

A team of researchers from Delft University of Technology and University of Porto has developed Morph, a novel energy-aware knowledge distillation framework that dramatically reduces the computational footprint of large language models (LLMs) for software engineering tasks.

The study challenges the industry-standard metric of FLOPs (floating point operations) for estimating energy consumption, demonstrating that it often fails to accurately reflect real-world power usage. Instead, the researchers propose using direct energy-surrogate models that measure actual CPU and GPU consumption during model optimization.

In controlled experiments across clone detection, vulnerability prediction, and code summarization tasks using models like CodeT5+, the team achieved remarkable efficiency gains: distilled student models reduced inference energy consumption by up to 90% and memory usage by 86%, while maintaining competitive accuracy. This breakthrough enables sustainable LLM deployment on consumer hardware and resource-constrained platforms that were previously inaccessible due to power limitations.