What is Green AI?

Half a century ago, Gordon Moore observed that the density of transistors on a given area of silicon would double about every two years – an established and proven pattern in the IT industry. However, one other intriguing pattern has recently emerged, one that is not commonly known – The compute power used to train large accurate Artificial Intelligence (AI) models has doubled every 3.4 months since 2012 (see).

AI has ushered in many big and promising possibilities; however, the demand for considerable computational resources necessitates substantial energy consumption.  Accurate AI models have clear benefits. However, they come with environmental implications. Researchers at the University of Massachusetts assessed the energy consumption spent when training several common large AI models. Their findings (see) include a shocking statistic. Training large models can produce an equivalent of about 626,000 pounds of carbon dioxide, which is about five times the lifetime emission of an average American car.

Advances in AI this past decade have been characterized by the development and deployment of large and complex AI models pushing the boundaries of innovation. Similar to the oil industry, once extracted and refined, data, like oil, can be a highly profitable commodity. However, like fossil fuels, the process of extracting value from data into an AI model has a costly environmental impact. The substantial amount of energy that powers the computing resource used to train big AI models in its current form is not sustainable concerning both the global electricity supply and the associated carbon footprint.

Green AI is a term that refers to research that aims to reduce the environmental impact of training large AI models.  It was coined by Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni (see). Green AI is a paradigm that gauges the success of AI not simply by the metric of the accuracy of AI models but also by their carbon footprint. Find out more about Green AI here.