Processing data takes a lot of power, and researchers from the University of Massachusetts, Amhurst have found that training a single machine learning algorithm can emit as much carbon dioxide as five average American cars do during their entire lifelines (including their manufacture).
In a newly published paper, the team report that the process can produce over 626,000lb of CO2 equivalent.
- What is AI? Everything you need to know
- What is machine learning? Your complete guide
- AI and ethics: looking inside the black box
The processors themselves use a lot of juice, but so does all the cooling necessary to keep them from overheating.
As the team note, training machine learning algorithms often means running supercomputers for weeks or even months at a time. Some of the power used might come from renewables (solar, wind and tidal), but the vast majority won't.
It's also not possible to suddenly increase generation of renewable energy to meet a surge in demand. If a supercomputer is guzzling power, suppliers need to start burning more fossil fuels to satiate it.
Brain drain
The team specifically looked at machine leaning algorithms for natural language processing, which is increasingly important as interfaces start to move away from keyboards and screens to speech recognition and voice synthesis in devices like smart speakers.
And this is only the initial training; further development of the algorithm will generate more CO2.
"We recommend a concerted effort by industry and academia to promote research of more computationally efficient algorithms, as well as hardware that requires less energy," the team concludes.
