What is Green AI, and how do you develop it?
Recent studies estimate that data centers represent nearly 10% of global electricity consumption. In other words, if digital were a country, it would be the third largest energy consumer. Among the technologies massively deployed in data centers, Artificial Intelligence (AI) is one of the most energy-consuming. While AI has a critical role to play in overcoming major environmental and societal challenges like biodiversity, waste management, or energy performance, its impact on the environment is not negligible. AI for Green must be accompanied by Green AI, which means an AI respectful of the environment.
With the rise of AI over this last decade, assessing its environmental impact has become urgent. Market studies show that 70% of companies will deploy at least one AI solution at scale by 2030. Indeed, the subject represents a real issue since the single learning phase of recent Deep Learning algorithms can emit tens or, even worse, hundreds of tons of CO2, thus equaling the environmental impact of several cars over their entire lifetime. Parallel to that, the power used for training a single AI algorithm has multiplied by 300,000 over the last six years, a trend that is only going upward.
However, making AI sustainable is complex and requires a combination of capabilities to be successful: —First, quantify the energy consumption of an AI and, more generally, its environmental impact by offering tools to measure and/or quantify precisely.
—Second, develop and propose concrete methods and levers to reduce the environmental impact of AI.
—Third, raise awareness in the AI community to promote eco-responsible AI and disseminate good practices.
—And fourth, establish norms and certifications to standardize the implementation and development of eco-responsible AI.
Today, environmental impact assessment focuses mainly on carbon emissions related to the use phase of a product or service. While global warming is a real challenge for humanity in the decades to come, the scarcity of raw materials such as metals and oxides is already a high-stakes issue for the industrial and digital sectors. So, the need to think more about sustainably beyond global warming is critical to meet future business needs. We must consider other environmental criteria, such as the scarcity of raw materials, water use, societal impact, and the entire product or service life cycle.
We expect that tools to quantify an AI's global and multi-criteria environmental impact, from extracting raw materials for electronic components to the end of its life, will emerge in the coming years. They will depend on standardized methodologies for systemic quantification of the environmental impacts of products, such as Life Cycle Assessment.
We have been building our expertise in green AI since 2019, particularly working to make AI sustainable by design. By anticipating the environmental impact of AI from the design phase, we can reduce the carbon emissions of AI-based solutions by up to 90%. We deeply believe that green AI is one of the next decade’s biggest challenges, where AI sustainability meets business growth and financial gain.