Energy Consumption Analysis for LLM Training in Edge Environments

Abstract

With the growing demand for real-time generative AI applications on edge devices, on-device fine-tuning techniques have gained increasing attention in power- and resource-constrained environments. This study performs LLM training experiments using GPT-2 Small and DistilGPT2 models on NVIDIA Jetson Xavier NX and Orin Nano devices, analyzing both training time and energy consumption. The experimental results are expected to contribute to the design of energy-efficient LLM deployment frameworks for edge environments.

Publication
Korean Institute of Communications and Information Sciences (KICS) 2025
Daeun Choi
Daeun Choi
Backend Developer

A backend developer majoring in IT Information Engineering at Jeonbuk National University.