Sign In

Communications of the ACM

ACM TechNews

AI's Carbon Footprint Problem


View as: Print Mobile App Share: Send by email Share on reddit Share on StumbleUpon Share on Hacker News Share on Tweeter Share on Facebook
Datacenters use a lot of electricity.

Artificial intelligence has a terrible carbon footprint.

Credit: Dean Mouhtaropoulos/Getty

Researchers at Stanford University, Facebook AI Research, and Canada's McGill University have developed a tool to measure the hidden cost of machine learning.

The "experiment impact tracker" quantifies how much electricity a machine learning project will consume, and its cost in carbon emissions.

The team first measured the energy cost of a specific artificial intelligence (AI) model—a challenge because a single machine often trains several models concurrently, while each session also draws power for shared overhead functions like data storage and cooling.

The researchers then translated power consumption into carbon emissions, whose blend of renewable and fossil fuels varies by location and time of day, by tapping into public sources about this energy mix.

The tracker determined that some machine learning algorithms consume more energy than others, and the researchers have incorporated an easy-to-use tool into the tracker that produces a website for comparing the energy efficiency of different AI models.

From Stanford Institute for Human-Centered Artificial Intelligence
View Full Article

 

Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA


 

No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account