Energy sustainability analysis of machine learning algorithms

AI applications are pervasive today and most innovative applications embed AI solutions. With the release of the ChatGPT-4 model by OpenAI, large language models (LLM) have received increasing attention and research. However, AI methods require a significant amount of computational resources, which leads to considerable energy consumption. Taking OpenAI's GPT-3 model as an example, training it once approximately requires 1287 MWh of energy[1]. The recent recommendations from the European Commission emphasize the need to govern and reduce CO2 emissions from software assets. Therefore, it is necessary to identify tools and methods able to provide a reliable and realistic estimation of the energy consumption of AI solutions and their environmental impact. The aim of our research is the quantification and investigation through ad hoc tools, for assessing the environmental impact of AI training. These ad hoc tools are envisioned as the preliminary step for implementing green AI approaches. The thesis first analyzes the energy consumption of different parts of the forward step in deep learning models. It then analyzes the energy consumption of different steps during the training process of the model. Based on this, a model is constructed to estimate the energy consumption for a given dataset and model. Then the thesis examines the relationship between different energy consumption limits and training time of a GPU during the training process. On this basis, a model is constructed that integrates solar power, batteries, and the electrical grid as power sources to power the GPU for training the deep learning model, evaluating the minimum costs under different start times and durations of operation.