Studies on Environmental Impact and Energy Efficiency of Artificial Intelligence and Machine Learning

The consideration of environmental impact and energy efficiency when developing artificial intelligence (AI) and machine learning (ML) models has received some traction recently as ML models get bigger.It has been shown to increase the energy efficiency of models without significantly affecting accuracy.We reviewed five recent studies that focus on investigating and refining the energy efficiency of ML systems through numerous models and methods with either software or hardware variations.The first article proposed using the total number of Floating-Point of Operations (FPO) to better measure efficiency without depending on hardware, location, and AI approaches used in the model.The second article evaluated and compared two deep learning frameworks, Pytorch and Tensorflow, and how each has its strengths and weaknesses regarding its efficiency in the training phase versus the inference phase.The third article analyzed energy-efficient architectures for ML, which includes utilizing more approximation to increase efficiency while having a minimal effect on accuracy significantly.The fourth article focused on holistic architecture solutions for creating efficient run times and energy consumption of neural networks.The last article developed a Java tool to help with increasing energy efficiency on edges with minimal decrease in accuracy.Lastly, we touched on foundation models and their energy efficiency considerations.