Summary

  • Researchers at MIT Technology Review have attempted to put some numbers to the ever-increasing energy usage associated with training and running AI models.
  • The team looked at three types of AI model in particular.
  • Open source language models; image generation models; video generation models
  • After training, the inference phase is the most energy hungry part of an AI models life, and the part where all the carbon emissions associated with AI can be calculated.
  • There are no standards for reporting these figures and companies are reluctant to share information, so the best proxy is open-source models.
  • Variables to consider in the calculations are the number of steps taken to generate an image and the quality of the results.
  • Once figures for energy usage have been calculated, they can be doubled to also provide an estimate of the carbon footprint created by the model.
  • This, in turn, can be localised to a particular area’s carbon emissions and energy usage, to get a truly accurate, though constantly shifting, figure.

By James O’Donnell, Casey Crownhart

Original Article