● Considering the power consumption and the huge number of users, the global digital environment is now generating a lot of greenhouse gases.
● The complex task of assessing carbon footprint is a particular focus of academic research.
● Statistical learning methods in artificial intelligence require enormous amounts of power, making this issue even more urgent.
Internet users find it hard to imagine that interacting with ChatGPT or watching YouTube-recommended videos actually emits greenhouse gases. Meanwhile, for researchers, the carbon footprint produced by computers and other digital devices is a very hot topic in the age of global warming. And that fossil fuels (coal, oil, gas) are being burned to generate electricity in the grid that we charge our batteries with and connect to machines, and most notably all the networks and They are keen to remind us that we must not forget the infrastructure of the Internet. Data centers that store data and applications consume a lot of electricity around the world.
In France, the issue is the subject of the National Digital and Environmental Program launched in 2022 by the French Institute of Computer Science and Automation. And the concerns raised by this issue are shared around the world, especially because of its staggering scale. The total amount of digital emissions we are expected to generate in the future. Earlier this year, Soumya Sudhakar, Vivian Shih, and Sertak Karaman of the Massachusetts Institute of Technology (USA) published a model that simulated the potential emissions from on-board data processing for electrically powered self-driving cars. announced the expected results. Sensors and artificial intelligence (AI). They noteworthy that the carbon footprint of the computing needed for a global fleet of 1 billion I conclude that they are at least the same.
A problem made even more urgent by the rise of AI
A study published in mid-February 2023 models emissions generated by machine learning from 2012 (a breakthrough year in the field) to 2021. His two authors are professional researchers working at Hugging Face and postdoctoral fellows at the University of Quebec for Artificial Intelligence. The institute has 77 sciences drawn from five fields: image classification, object detection, machine translation, conversational agents or chatbots, and named entity recognition (an aspect of natural language processing that classifies words into strings). We selected 95 ML algorithms mentioned in the paper. Categories: People, Places, Companies, Dates, Quantities, Addresses, etc.).
Gathering all the information needed to perform a detailed carbon footprint estimate is very difficult.
The aim was not to assess the exact amount of carbon dioxide associated with each of these, but to outline the main trends. “It is very difficult to collect all the information necessary to make a detailed carbon footprint estimate.” Hug Face’s Sasha Luccioni points out: “Articles on AI tend not to reveal how much computing power was used or where the training took place.”
Low performance levels do not necessarily mean low emissions
This project focused on the training phase of a learning model, which requires a large amount of computational power. The first finding was that 73 of the 95 models were trained using power generated primarily from coal, natural gas, and oil. As an example, a model powered by energy from coal produced an average CO2 equivalent of 512g per kilowatt-hour. This compares to 100.6g for a model powered primarily by hydroelectric power (some greenhouse gases were produced but converted to CO2 equivalents). provide a single number). Second, in this context, it is important to note that higher electricity consumption does not necessarily mean higher carbon dioxide emissions, given the low emissions of models operating on hydropower. is. Another finding is that performance does not necessarily correlate with reduced carbon footprint when comparing two fossil fuel-powered models.
The carbon footprint of machine translation algorithms has been declining since 2019
However, the researchers did not observe “A systematic increase in carbon footprint for individual tasks.” The footprint generated by image classification models and chatbots continued to grow, while the footprint generated by machine translation algorithms has decreased since 2019.
Still, the fact that there has been an overall increase cannot be denied. The learning model produced an average of 487 tonnes of CO2 equivalent in 2015-2016. By 2020-2022, this figure, which is for training only, will reach 2020 tons. The introduction also made a big impact. Sure, a single ChatGPT request is minimally costly to perform from an energy perspective, but the millions of requests sent daily to ever-growing chatbots are even more problematic. “That’s what I’m working on now.” says Sasha Luccioni. “However, it is still a complex task given how the model is deployed, the hardware used, scaling, etc. all have a significant impact on the energy required and the carbon emitted.”
