Reproduce “Energy-Efficient GPU DVFS for Fine-Tuning of SLMs on Resource-constrained Embedded Devices” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jurn-Gyu Park, Sanzhar Zholdybayev, Aidar Amangeldi, Ademi Zhanuzakova
Researchers from Nazarbayev University characterized the energy-performance trade-offs for small language model (SLM) fine-tuning on embedded GPU platforms and proposed an interpretable machine learning model to dynamically select energy-optimal GPU frequencies. This approach achieves an average energy saving of 13.11% on unseen GLUE benchmark tasks, enhancing the feasibility of on-device AI adaptation.
Key insights
- SLM fine-tuning on embedded GPUs exhibits a U-shaped energy consumption pattern across GPU frequencies, with optimal energy points typically found at intermediate frequencies.
- The stability and clarity of this U-shaped energy-frequency pattern are correlated with the number of model parameters, with larger models demonstrating more consistent optimal points.
- Simple, interpretable machine learning models can effectively identify and adapt to energy-optimal GPU frequencies for diverse SLM fine-tuning workloads on resource-constrained embedded platforms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.