AI's Hidden Environmental Footprint

Thank you for reading my article. I regularly write about Digital transformation and how it is reshaping the future of the enterprise.

To read my future content, follow me on LinkedIn.

——

Climate change always brings images of thick, dark smoke billowing from factory chimneys, obscuring the once-blue sky. It's natural for our minds to associate climate change with something ‘tangible’ that our senses can perceive, such as exhaust fumes of packed traffic in Delhi or smog over Shanghai.

However, amidst the apparent culprits, there are some hidden or 'intangible' agents causing environmental damage without much recognition or awareness. Today we are experiencing a revolution with AI, Machine learning, Blockchain, and other ground-breaking technologies taking the world by storm. At first glance, these technologies might seem eco-friendly due to their lack of obvious mechanical or physical processes leaving any carbon footprint. But in reality, AI and other technologies have a profound impact on the planet's ecology and climate.

Take, for instance, the mining of Bitcoin — the most popular cryptocurrency — which requires an incredibly fast computing processor called GPU (Graphic Processing Unit). The use of these GPUs consumes a substantial amount of energy and as more people mine Bitcoin, the algorithm becomes more complex and hence requires even more computational resources. To put things into perspective, the energy consumed annually to mine Bitcoin surpasses the entire energy consumption of a country like Argentina with a population of 45 million1.

Another booming industry, Generative AI or Large Language Models (LLMs), shares similar computing needs. Training these models requires multiple instances of GPUs to handle billions of calculations within seconds, resulting in massive energy consumption. Additionally, Data Centers where these models are trained, require a huge amount of water to cool the equipment — similar to your cars needing a coolant to keep the engine cool.

AI's ecological toll

Research from the University of Massachusetts Amherst2 indicates that training only one Transformer model (with a neural architecture search) generates approx. 350 tonnes of CO?. That is the equivalent of 205 return flights between New York and London! Another similar 2022 study3 on the carbon footprint of training a 176 billion parameter LLM called BLOOM estimated that 25 tonnes of CO? was generated in training that model.

The following chart, reproduced from Stanford University's Human-Centred Artificial Intelligence (HAI) Institute's 2023 AI Index report ?, illustrates the CO? emissions from various studies. Notably, the GPT-3 model stands out as the biggest carbon emitter at 500 tonnes. The average power consumption to train GPT-3 model was estimated to be 1,287MWh — enough to power an average American home for 120 years.

?

Beyond carbon emissions, another pressing issue lies in the water footprint of these AI models. A 2023 study reveals that training the GPT-3 model in Microsoft's U.S. data centres consumes a substantial 700,000 litres of clean, fresh water. This consumption increases to 3.5 million litres when the offsite water footprint (e.g. to generate electricity used for training) is taken into consideration?. This high water consumption is a cause for concern, especially as interactions with LLMs increase, leading to greater demands on water for data centre cooling. In other words, ChatGPT needs to ‘drink’ a 500ml water bottle for an average session of 20-50 Q&A?.

Future of Research

While AI holds great promise in optimizing energy usage and aiding in carbon emission reduction, it also poses a significant threat to the environment and its limited resources. Raising awareness about the potential environmental impact of these technologies is crucial, and organizations training LLMs should provide more transparency regarding their energy usage and emissions.

AI's presence is undeniable, and it will undoubtedly reshape how we work, interact, and innovate. Nevertheless, world leaders, tech experts, organisations and regulators must act urgently by supporting research institutes to conduct independent studies on AI's environmental impact. Currently, the cost of training models for such experiments prohibits the effective study of the impacts.

Establishing a comprehensive framework to achieve a Net-Zero AI footprint is also essential, considering the rapid progress in Generative AI development and usage. Remember, when it comes to the environment…we don't have a Plan B.

…………………….

References:

1. Hinsdale, J. (2022).?Cryptocurrency’s Dirty Secret: Energy Consumption. [online] State of the Planet. Available at: https://news.climate.columbia.edu/2022/05/04/cryptocurrency-energy/ .
2. Strubell, E., Ganesh, A. and McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP.?arXiv:1906.02243 [cs]. [online] Available at: https://arxiv.org/abs/1906.02243v1 .
3. Luccioni, A.S., Viguier, S. and Ligozat, A.-L. (2022). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. doi:https://doi.org/10.48550/arxiv.2211.02001 .
4. Artificial Intelligence Index Report 2023: Introduction to the AI Index Report. (2023). Available at: https://aiindex.stanford.edu/wp-content/uploads/2023/04/HAI_AI-Index-Report_2023.pdf .
5. Li, P., Yang, J., Islam, M.A. and Ren, S. (2023). Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models.?arXiv (Cornell University). doi:https://doi.org/10.48550/arxiv.2304.03271 .