AI, energy and resource conservation: opportunities and challenges

In times of climate change and limited resources, it is more urgent than ever to find sustainable, peaceful and efficient solutions. Artificial intelligence (AI) has the potential to play a key role in overcoming these global challenges, for example through better climate models or the expansion of renewable energies. At the same time, however, there is growing concern about the enormous energy consumption of AI systems themselves. Large AI models such as ChatGPT require considerable computing resources, resulting in a significant carbon footprint.

A new study shows that by 2027, AI could consume as much electricity as the Netherlands does in an entire year, or around half a percent of global electricity consumption. Data scientist Alex de Vries estimates that AI chips will consume between 85 and 134 terawatt hours (TWh) of electricity per year by 2027. Research at the L3S is not only concerned with the enormous challenges of AI, but also with the potential applications for greater climate and resource protection: the scientists are developing solutions to make AI more energy-efficient and are researching possible applications for the expansion of renewable energies and environmental protection.

Energy efficiency of AI systems

Training large neural networks requires enormous computing resources and therefore a considerable amount of energy. Even training relatively small large language models consumes many times the amount of CO2 emitted by several cars over their entire lifetime. Large language models also require around 30 times as much energy for inference as a Google search query. However, there are promising approaches to improving the energy efficiency of AI.

One example at L3S is the development of automated Green ML methods for driver assistance systems, which are being installed in more and more cars. These systems optimize energy consumption by performing only the necessary calculations while maintaining a high level of accuracy. Using automated machine learning (AutoML), these approaches can later be transferred to other applications. Another exciting field is Hypersparse Neural Networks, which aim to learn only the most relevant weights of neural networks, thereby reducing the number of operations required and thus the energy consumption. This requirement can be reduced even further with so-called binary neural networks, which are extremely easy to implement in hardware. In some cases, savings of over 95 percent in energy consumption are possible without compromising quality.

The development of energy-efficient hardware accelerators for neural networks is the focus of the L3S project ZUSE-KI-mobil. These accelerators are specially designed for use in embedded applications and help to reduce energy consumption in mobile and networked devices.

AI in use for renewable energies

AI can also contribute directly to the promotion of renewable energies. For example, the L3S project WindGISKI is developing an AI-based GIS platform that identifies potential areas for wind turbines. This platform helps to identify the best locations for wind turbines and thus maximize the efficiency and expansion of wind energy.

Another notable L3S project is OffshorePlan, which deals with construction site logistics for wind turbines. AI optimizes the planning and execution of logistics processes, which lowers costs and reduces the environmental footprint of construction projects.

Protecting the environment and resources with AI

In addition to promoting renewable energies, AI can also make a significant contribution to environmental protection and resource conservation. One outstanding example is the L3S project SWIFTT (Satellites for Wilderness Inspection and Forest Threat Tracking). Here, AI-supported satellite data is used to identify threats to forests at an early stage and take appropriate protective measures.

The L3S project GLACIATION shows how the development of a distributed knowledge graph can improve the efficiency of big data analysis and thus reduce carbon emissions. This knowledge graph makes it possible to process large amounts of data more efficiently and thus helps to reduce energy consumption.

Research at L3S

At the L3S research center, we work intensively at the interface between AI, energy efficiency and resource conservation. Our goal is to develop new methods and technologies that both increase the performance of AI systems and minimize their energy consumption. We cover the entire spectrum: from basic research into new architectures for neural networks and specialized hardware to the efficient development of new AI applications with AutoML and AI applications for sustainability.

The potential of AI in the area of energy efficiency and resource conservation is enormous. Through the continuous further development and targeted use of AI technologies, we can not only increase the efficiency of our systems, but also make a significant contribution to protecting our environment. The projects presented in this issue of Binaire provide an impressive insight into the many ways in which AI can contribute to sustainability.