Green AI: A Framework for Energy-Efficient and Sustainable Machine Learning Systems

Abstract

Artificial Intelligence is improving rapidly and changing how all industries operate. However, a significant environmental threat comes along with the advancements. It has been calculated that the training of a large AI model creates as much CO2 as the lifetime emissions of 5 cars. Furthermore, data centers use as much electricity as entire countries. This research focuses on methods of creating energy-efficient and sustainable artificial intelligence systems. Using the data from academic articles, we conducted systematic relevance screening and determined the most important studies. Along with valuable industry case studies, we created a practical framework which we call the “Green AI Pyramid.” Within this framework, we focus on 4 key steps. These include: (1) Evaluating and Assessing Energy Use, (2) Selection of Energy Efficient Algorithms, (3) Use of Smart Hardware, and (4) Emission Reduction through Offsetting. By following our recommendations, organizations can make energy consumption reductions of 40-90% with no loss of accuracy to the AI systems. An example of this is Google, which was able to use AI to reduce cooling energy by 40%, and Huawei, which was able to reduce model size by 75%.