Researchers at the Allen Institute for AI have developed a method of calculating the carbon emissions of artificial intelligence (AI) systems that boasts greater accuracy.
The new approach records the energy usage of server chips as a series of measurements and aligns that data with a series of data points detailing local emissions per kilowatt-hour (kWh) of the energy consumed.
The researchers found graphics processing units training the biggest model to only 13% completion emitted nearly as much carbon as powering a U.S. home for a year does. In contrast, training the smallest models generated as much carbon as charging a phone.
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2022 SmithBucklin, Washington, DC, USA