Towards Green NeRF: an Exploration of Energy Influence Factors in NeRF Models
Zeke Raphael Lim Sy, Shihao Luo, Jean Atsumi Flaherty, Truong Cong Thang · 2024
Recently, Neural Radiance Field (NeRF) has become a promising AI solution for reconstructing complex scenes, which are important for various applications, such as robots, education and Virtual Reality. Training a model has been proved to be a lengthy task, leading to new NeRF architectures which achieved massive reductions in training time without sacrificing the quality. However, until now, energy efficiency has never been considered in the design of these models. With growing concerns about the energy consumption of AI-related systems in recent years, there is a clear gap in the NeRF research which must be addressed. To this end, we conduct an empirical experiment to investigate the impacts of different NeRF architectures and datasets on energy consumption. Our findings reveal the significant impact of key factors on energy consumption and identify the most energy-efficient model. Additionally, we discuss potential parameters and highlight areas warranting further investigation.