Go down

The use of artificial intelligence (AI) is increasing year by year, spanning various domains (commercial activities, artistic creation tools, weather prediction, etc.) within our society. Concerning its monetary cost, it varies depending on the economic model chosen by businesses.

However, there is a hidden cost that few people are aware of. This is the environmental cost (carbon dioxide [CO2] emissions) of training an AI model. 

Take, for example, the training (not forgetting its usage) of the GPT-3 model; depending on the continent, the level of carbon emissions can vary from one to two-fold (Taddeo M. and al. "Artificial Intelligence and the Climate Emergency: Opportunities, Challenges, and Recommendations," June 8, 2021). 

The research team found that the carbon footprint of different AI models is not necessarily taken into account during their design. This team has developed a framework and a series of measures to apply in order to, if not halt the development of AI models, at least reduce carbon emissions (Henderson P. and Al., "Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning," The Journal of Machine Learning Research, Volume 21, Issue 1, Article No: 248, pp 10039–10081). 

The Surface Calculation of Carbon Footprint

To calculate the carbon footprint, there are applications available to optimize the performance of AI models (e.g., CodeCarbon). There are also calculators that can predict the carbon footprint of an AI project. For example, there's the Green Algorithms project (http://calculator.green-algorithms.org) developed by Loïc Lannelongue, Jason Grealey, and Michael Inouye from the University of Cambridge (United Kingdom) and the Baker Heart and Diabetes Research Institute (Melbourne, Australia). 

However, calculating the carbon footprint is primarily an overall estimate rather than precise results ("Measuring the environmental footprint of the digital world, a real headache," Next, Mathilde Saliou, 2023). 

Indeed, there are several factors to consider when calculating the carbon footprint of an AI model. These factors include the algorithm design (is the AI model code efficient?), electricity consumption (a factor that varies depending on the location and its production mode), test duration (on average, tests last 24 hours), and server hosting type (Cloud carbon footprint: Do Amazon, Microsoft, and Google have their head in the clouds? Carbon4, 2022). 

Water Footprint

In addition to its carbon footprint, the use of AI leaves another type of imprint on our environment that is underestimated or even ignored by developers. This is freshwater consumption. To train OpenAI's AI model, GPT-3, Microsoft's data centers in the United States consumed 700,000 liters of freshwater. This is equivalent, in the automotive sector, to creating 370 BMW cars or 320 Tesla cars. 

As mentioned by Professor Shaolei Ren ("How much water does AI consume? The public deserves to know," 2023, OECD.AI), training AI models requires freshwater for the air conditioning systems of server rooms. Additionally, freshwater is used in the cooling system of power plants.  

For information, in Microsoft's data centers, for every kWh consumed, freshwater consumption ranges from 1.8 liters to 12 liters. To calculate water consumption, Professor Shaolei Ren and his associates applied this formula: 

Water Footprint =
Server Energy * WUE On-site + Server Energy * PUE * WUE Off-site

Energy Waste

It's a fact that data centers are significant carbon emitters, generating heat from air conditioners and servers. Additionally, as mentioned in the previous paragraph, they are heavy consumers of freshwater. 

However, it's important not to forget that heating (gas, wood, oil) also pollutes just as much, if not more. Not to mention the absurdity of using potable water for flushing toilets. 

The goal isn't to stop heating altogether but rather to harness "digital" heat sources as a new form of fuel.

Conclusion

This brief investigation demonstrates that firstly, it is difficult to calculate the carbon footprint of AI usage. It is more of a rough estimate than a precise calculation ("Measuring the environmental footprint of the digital world, a real headache," Next, Mathilde Saliou, 2023). This depends primarily on the diverse nature of energy production methods (hydroelectric, thermal, nuclear, coal, etc.). Additionally, it depends on the nature of the infrastructure hosting the data (shared, private, etc.). 

In this study, we have observed that depending on the nature (traditional or generative) of the trained model (Crawford, "Generative AI’s environmental costs are soaring—and mostly secret," 2024), there is a difference in the amounts of CO2 emitted. 

We have also seen that in addition to emitting CO2, infrastructure requires freshwater for the operation of cooling systems to train AI models. 

Apart from the consumption of energy and water resources and the production of CO2, the use of hardware for system infrastructures (servers, storage systems) and network infrastructures poses another ecological problem. This involves the extraction and use of rare earth minerals, as well as the management of electronic waste (Marcella Evite P. Bitcoin’s Climate Impact: Carbon Emissions and Beyond). 

While the industry sector has its share of responsibility, the research sector also bears responsibility. The availability of scientific publications on the CO2 cost (and more generally, the environmental cost) of AI models is rather limited, as noted by researchers Roy Schwartz, Jesse Dodge, Noah A. Smith, Oren Etzioni (Schwartz R. and al. Green AI, 2019). The research team observed that between 2012 and 2019, there were more scientific publications on the accuracy of AI model calculations than on the efficiency of AI models. 

Rare also the research studies on improving AI models or research studies on increasing the energy efficiency of production centers or infrastructures used for model training. It is worth noting that few companies are attempting to implement systems to recover the heat energy released by data centers. For example, French companies like EQUINIX or ILLIAD are developing systems to utilize the heat emitted by data centers to redistribute warm water in the urban heating system. 

Instead of wasting time listening to incompetent people, it would be wiser to improve infrastructure and AI models. To support innovation and research in energy sectors or to leverage collateral effects (heat emissions, CO2 emissions, etc.) to our advantage rather than shooting down new technologies. 

Among the solutions proposed, in addition to the certainty of reducing carbon emissions, but without returning to the Stone Age. There are geological solutions using the mineral olivine (Olivine weathering, Campbell Nilsen, Work in Progress, via Contrepoints). Or CO2 capture (Captage du carbone : ce qu'il faut savoir, Maxime Bilodeau, Science Presse). While these solutions may make you smile, they at least have the merit of existing and trying to find solutions.

And let's not forget the harmful effects of deforestation, since trees play an important role in capturing CO2 and producing oxygen (photosynthesis). 

It should not be neglected that other sectors of activity are also sources of carbon emissions. As an indication, in 2022 worldwide, industry accounted for 24%, agriculture and electricity production each accounted for 22%, and transportation accounted for 15%.

While mankind emits around 8 billion tonnes of CO2 per year, Mother Nature emits around 380 million tonnes of CO2 every year through volcanic activity (Weathering of rocks impacts climate change, CNRS, Phys.org via Contrepoints).

Another "natural" source of carbon emissions is forest fires. In 2023, a sad record year for forest fires, countries such as Canada, the United States (the state of Hawaii, especially the island of Maui), Chile, Argentina and Greece experienced major forest fires. These forest fires accounted for 23% of carbon emissions in this sad record year (Nearly a quarter of carbon emissions burned in 2023 come from fires in Canada, Sarah Boumedda, Le Devoir).

In addition to the graph titled "Global Greenhouse Gas Emissions," carbon dioxide (CO2) comprises approximately 80% of greenhouse gases. The remaining greenhouse gases include methane (CH4), nitrous oxide (N2O), hydrofluorocarbons (HFCs), and water vapor (H2O).

The best for last, is the implementation of a carbon footprint monitoring system based on... Deep Learning and Machine Learning (Carbon Footprint Monitoring System Using Machine Learning and Deep Learning Techniques, IEEE Xplore, 2023). I'll leave you to ponder this idea. 

 

Bibliography/Sitography

  • Meetup FinOps #13 | Les politiques environnementales des hyperscalers par Carbone 4, 14 juin 2022, YouTube. Lien : https://youtu.be/tmeiJQbL9Xo  
  • Meetup FinOps #18 | Empreinte environnementale de l'IA + Synergies FinOps/GreenOps,6 décembre 2023, YouTube. Lien : https://youtu.be/lqgSeQ-YiNg  

 

Other articles (not treated) related to the publication

 

  • Le captage, l’utilisation et le stockage du carbone sont nécessaires de toute urgence (ONU), 3 mars 2021, ONU Info, Organisation des Nations Unis/United Nations. Lien : https://news.un.org/fr/story/2021/03/1090762  

 

Tutti sono invitati a partecipare al progetto STULTIFERA NAVIS!

 

Alexandre Martin

Alexandre Martin / Business Analyst

gnosis.grenoble@gmail.com https://bio.site/AlexandreMartinGnosis
El tribalismo digital, entre la furia y la farsa: pinchemos la burbuja de la polarización artificial en internet
Evolving Dynamics of Customer Experience: An Empirical Analysis of Interaction-Centric Approaches