The climate crisis unprecedented obliges all countries, and especially those that are part of the United Nations, given that they are signatories to the 2030 Agenda for Sustainable Development, to promote strategies and initiatives to combat it. It is not, however, just a question of ethics or responsibility. Several studies show that Sustainability is key to the success of private initiatives[1],[2], as legislation is increasingly stricter and investors and end consumers are more aware and critical.
Green artificial intelligence, in English Green AI, It is an emerging paradigm that aims to respond to and integrate the criteria of environmental responsibility and sustainability into artificial intelligence systems. This paradigm recognizes global efforts towards sustainability and integrates both the classic metrics of algorithm quality, precision and speed, and metrics that refer to the ecological footprint, as can be seen in the recent article describing the Llama2 model.[3], and where an entire section is dedicated to discussing their broadcasts.
It is essential to address how the industry Artificial Intelligence can contribute to the advancement of the 2030 Agenda. Although the answer may seem obvious, we can identify two key areas: a) the reduction of emissions originating from the same generation of AI and b) collaboration in facing environmental challenges, in this article, we propose to explore, contextualize and exemplify each of these in detail.
It is clear that AI can play a decisive role in the success of Sustainable Development Goals (SDGs) offering answers to specific challenges. In fact, the Generalitat of Catalonia, in its report of July 2019[4], already includes examples of how AI can contribute to the achievement of each of the 17 SDGs included in the 2030 Agenda.
There are countless examples to illustrate the potential of AI in this area. We could talk, for example, about the agreement between Microsoft and the Generalitat de Catalunya[5], which addresses the watershed management to guarantee compliance with Catalan legislation, the wild animal count through the use of drones or intelligent detection of pollution in wastewater treatment plants. Another example is a project carried out by NTT Data and ProDrone to identify and locate an invasive plant that is highly toxic to both people and animals.[6].
However, while it is true that AI can be a boost on the path towards sustainability, we cannot ignore that it is also part of the problem.[7]. The University of Massachusetts was one of the first institutions to raise awareness about this issue, estimating that only the training process of a deep learning model of natural language processing (NLP) emitted five times more CO2 than the average consumption of a car throughout its entire useful life[8]. Moreover, according to recent studies, at the current rate of AI adoption, the global share of energy consumption in data centers will increase from 2% in 2019 to 10% in 2025.[9]. Other studies state that global energy consumption will increase from 190TWh in 2018 to 3000TWh in 2030.[10].
The Green Software Foundation[11], of which NTT has been a part since 2021, is a non-profit organization founded by the Linux Foundation that Its main objective is to reduce by 45% the emissions produced by Information and Communication Technologies (ICT). The foundation has identified three areas of action for reducing emissions, which are applicable in the field of AI:
- The first is Cloud optimization or data centers. One of the main challenges in this area is to reduce the use of cooling systems, which consume up to 40% of energy. For example, at NTT DATA we have tested the immersion system for refrigeration, which has managed to save up to 97%[12] the energy consumption dedicated to refrigeration. Another example is that of Google, which has used machine learning techniques with the same objective, obtaining improvements of up to 40%.[13].
- The second area is the optimization of the hardware itself, such as Cloud servers or the different components of personal computers. It is necessary to develop designs specialized in executing AI workflows. A clear example is the evolution of generic Central Processing Units (CPUs) towards specialized Graphics Processing Units (GPUs) and, about five years ago, the emergence of Tensor Processing Units (TPUs), designed by Google specifically to accelerate ML processes.
- Finally, the third area consists of optimize the software through i) model compression techniques to reduce training and execution time, as well as, in turn, reduce memory usage, ii) the use of pre-compiled programming languages, such as C++ or Java, which have higher processing speed than the popular Python, and can, therefore, reduce the energy cost of their execution and iii) using formats for data that optimize its processing and storage, such as Parquet, a column-oriented format that is much more efficient than CSV.
The truth is that the uses of AI to respond to challenges within the framework of sustainability are infinite and, certainly at this point, unimaginable. However, it is crucial that its production and exploitation are also sustainable. As is usual in the face of global challenges, with only one conjunction of research, public and private initiative, where the entire ecosystem of hardware and software production intervenes and with a clear commitment to the development of AI from an environmental perspective, AI can become truly green and a true tool in achieving the SDGs.[1] https://online.hbs.edu/blog/post/business-sustainability-strategies [2] https://www.forbes.com/sites/forbesbusinesscouncil/2021/02/10/why-corporate-strategies-should-be-focused-on-sustainability/?sh=6daed0a27e9f [3] Touvron, Hugo, et al. “Call 2: Open Foundation and Fine-Tuned Chat Models.” https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/ (2023).
[4]http://www.accio.gencat.cat/web/.content/bancconeixement/documents/informes_sectorials/informe-tecnologic-inteligencia-artificial.pdf [5] https://news.microsoft.com/es-es/2022/08/04/microsoft-y-la-generalitat-de-catalunya-collaborate-en-la-puesta-en-marcha-de-tres-proyectos-de-inteligencia-artificial-para-la-mejora-de-la-sostenibilidad-medioambiental/ [6] https://nttdata-solutions.com/us/services/innovation-consulting/giant-hogweed-case/ [7] CIDAI Masterclass: NTT DATA Generative AI & LLMs: https://www.youtube.com/watch?v=4zyuoIIA_rw&t=2312s [8] https://arxiv.org/abs/1906.02243 [9]https://observer.com/2019/08/artificial-intelligence-bitcoin-cloud-computing-climate-change/ [10] https://dl.ndl.go.jp/view/prepareDownload?itemId=info%3Andljp%2Fpid%2F11546567&contentNo=1 [11] https://greensoftware.foundation/ [12] https://www.nttdata.com/jp/ja/news/release/2022/060601/ [13] https://www.deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-by-40