The carbon footprint caused by digital systems represented, according to 2020 estimates, 4% of global greenhouse gas emissions, a figure comparable to emissions from other areas of activity such as road transport or aviation.
Energy consumption attributable to digital systems is estimated to grow at a rate of 6 to 9% per year and could represent 20% of global energy consumption by 2030, with an ecological footprint (which includes the carbon footprint and other factors) that would represent 2.7% of the global total.[1].
Without a doubt, the widespread, massive and growing use of applications based on artificial intelligence contributes to this increase in the environmental impact of digitalization, as can be deduced if, apart from CO emissions2, figures related to other environmental impact elements such as energy consumption and the use of water to cool the large data processing centers that make up the global “cloud” where a good part of AI-related computing takes place are evaluated. Let’s look at some figures.
Training large language models the size of GPT-3 (with 175 billion parameters) can be time-consuming. 1300 MWh[2] (equivalent to a CO emission2 of 502 tons) and require around 1.8 liters of cooling water per kWh (where this is the average value of the water use efficiency), which would lead us, in the case of GPT-3, to a consumption of 2,340 m3 of water for cooling. If we scale these values to a global level, we can get an idea of the high magnitude of the environmental impact of the use of AI-based applications.
Given this reality, the sustainability of the operations that AI requires is a growing concern and consequently research and innovation work is being developed to reduce this environmental impact that goes in several directions and addresses potential solutions in the field of software on the one hand, and in the physical field on the other.
In the software level measures we can highlight the following. First of all, a series of good practices should be considered for code optimization in order to make it more efficient in the use of computing resources. Another measure to consider is to improve the efficiency of AI models through pruning algorithms that remove unnecessary neurons and weights from a neural network, or the distillation of knowledge, a technique used primarily in deep learning to compress models, particularly massive deep neural networks. The use of domain or application-tailored AI models (as opposed to general models) are usually smaller and less expensive to train and therefore have a lower environmental impact. Another possibility to gain efficiency in the code is the use of new software paradigms based on quantum algorithms, which can perform complex operations and process large databases faster than classical algorithms.
In relation to hardware, the main actions aimed at reducing the environmental impact of AI are the following. In the field of microelectronics, work is being done on new semiconductor materials with better performance in terms of energy consumption and thermal dissipation than silicon. While they may be useful in certain applications, their difficult scalability and high cost do not yet make them viable to replace the silicon semiconductors that are the basis of current devices on which AI operates, although research is progressing and results such as those recently announced by MIT are being achieved.[3], in which they present a transistor design that overcomes the inherent limitations of silicon devices and could be a viable technology and significantly reduce data center energy consumption, improving processing capabilities for AI and machine learning applications. The design and implementation of new and more efficient refrigeration technologies, such as microfluidics integrated into the semiconductor chip, is another way at the device level to increase energy efficiency and reduce the use of water as a cooling material. With regard to large data processing centers, it is necessary to seek solutions at scale. These range from greater use of renewable energies to meet the demand for electricity for use of recycled water for refrigeration.
Aside from these options, other paths emerge that explore completely different approaches. One of these was recently presented by Rika Nakazawa, Executive Commercial Director at NTT, at the latest edition of the AI&BD Congress. In this case, the idea consists of installing data processing centers on satellites in Earth orbit and is based on using solar energy captured by the satellite’s panels and thermal dissipation through heat sinks that would evacuate the heat into space. This proposal, which may seem very futuristic, is actually part of the strategy of NTT, the large Japanese telecommunications operator, to improve global sustainability.[4].
But, with regard to digitalization in general and sustainability in particular, there is an element that must be taken into consideration and that often goes unnoticed. It is the concept of “dark data” or dark data, referring to the set of data that has been generated in the digital sphere that has ceased to be used but remains stored in the cloud and, therefore, continues to consume energy and other resources. Within dark data there is data generated by companies, administrations and other organizations, but also the set of images, messages, posts, etc. generated by individuals and which remain forgotten. It is necessary for citizens, as a whole, through education and training, to become aware of this phenomenon and its implications for the sustainability of the planet and therefore make appropriate and ethical use of the digital channels that are available to them.
[1] Zulfiqar, M., Tahir, SH, Ullah, MR et al. Digitized world and carbon footprints: does digitalization really matter for sustainable environment?. Environ Sci Pollut Res 30, 88789–88802 (2023).
[2] Luccioni, AS, Viguier, S., & Ligozat, AL (2023). Estimating the carbon footprint of bloom, a 176b parameter language model. Journal of Machine Learning Research, 24(253), 1-15 (2023)
[3] https://www.unite.ai/mit-research-team-engineers-quantum-solution-to-computings-energy-problem/
[4] NTT and SKY Perfect JSAT Agree to Establish Space Compass Corporation -Novel Space Integrated Computing Network Enterprise to Aid Realization of a Sustainable Society- | Press Release | NTT
