Energy Sustainability Analysis of Deep Neural Network

Artificial Intelligence (AI) applications are becoming more widespread, raising significant environmental concerns due to the high energy use required to train large Deep Neural Networks (DNNs). To address this issue, we conduct a detailed quantitative analysis of the energy consumption of various models across different training steps and datasets. Our study measures energy use at each training stage with a comprehensive review of selected models. By tracking energy consumption during each training step, we find that the backpropagation phase is the most energy-intensive. Additionally, we evaluate the power limits and performance features of Graphics Processing Units (GPUs), collecting empirical data on their behavior under different GPU power settings. Based on these insights, we explore the integration of locally installed renewable energy sources, such as solar power and battery systems, with the electrical grid to enhance the energy sustainability of GPU operations. We introduce and test an innovative approach for managing energy and computing resources, aiming to optimize energy use and reduce operational costs. Our results demonstrate that this method can reduce energy consumption by more than 40% and operational costs by almost 25%, paving the way for greener AI solutions.