Exploration and systematic assessment of the resource efficiency of Machine Learning

Abstract

Estimations of today’s energy consumption of information and communication technologies (ICT) range from 2 to 9 % of the total produced energy and forecasts for the year 2030 predict an increase up to 21 %. Even though these numbers are controversial, it cannot be denied that the consumption growth of large impact factors, like data centers, networks, consumer devices, and the production of ICT needs to be reduced. In addition to Green IT, which is primarily focused on hardware, software is increasingly seen as an energy consumer with considerable savings potential. In this paper, we take a look at software for artificial intelligence (AI) and especially machine learning (ML). We describe a method for in-depth measurement and analyses of the energy consumption and hardware usage of ML algorithms and a series of experiments where we use the method on convolutional neural networks (CNN). We also compare existing estimation methods with our own. As outlook, we propose a holistic approach along the AI life cycle and additional experiments and assessments that could show potential efficiency improvements and consumption savings in AI.