Latin America. There is no artificial intelligence without energy: a data center dedicated exclusively to AI products and services today consumes as much electricity as 100,000 homes, according to the International Energy Agency (IEA).
In recent years, artificial intelligence has ceased to belong exclusively to the world of research to conquer more and more spaces in our lives and this has been accompanied by a significant increase in its energy needs.
According to the IEA, data centers currently consume 1.5% of all the electricity produced in the world and, if nothing changes, their energy demand will double by the end of the decade.
On the road to making something change and reducing the energy footprint of AI, two studies by the Open University of Catalonia (UOC), with the participation of researchers Fernando Sevilla Martínez and Laia Subirats Maté, from the NeuroADaS Lab (Cognitive Neuroscience and Applied Data Science Lab) group, propose alternatives towards a more sustainable and efficient AI and, also, more affordable. The articles have been published in open access in IEEE Networking Letters and in the International Journal of Intelligent Systems.
"Energy efficiency must become a central parameter in the design of AI. It's not just about making models faster or with better performance, but about making them sustainable, ethical and accessible," says Subirats Maté, also an associate professor in the UOC's Faculty of Computer Science, Multimedia and Telecommunications. "Designing energy-efficient AI not only benefits the planet, but also enables AI to be deployed in small devices such as robots and sensors, reduce operating costs for businesses and data centers, and improve resilience in environments with limited connectivity or power."
Greener neural networks
The first of the published works, led by PhD student Fernando Sevilla Martínez from the UOC and with the participation of the Universitat Autònoma de Barcelona, the Centre de Visió per Computador (CVC/UAB), the Telecommunications Technology Centre of Catalonia (CTTC) and the Volkswagen group, has shown that it is possible to develop low-power and high-performance impulse neural networks (a type of AI that mimics the functioning of the human brain) using inexpensive and affordable components, such as Raspberry Pi 5 and the BrainChip Akida accelerator. This study paves the way for energy-efficient distributed artificial intelligence networks, applicable in fields such as transport, environmental monitoring or the industrial Internet of Things (IoT).
"The methodology we propose allows us to train, convert and execute these impulse neural network models without the need for a graphics processing unit or connection to a data center or the cloud, with a consumption of less than ten watts of energy," the authors detail. "In addition, thanks to other technologies such as Message Queuing Telemetry Transport, Secure Shell and Vehicle-to-Everything communication, multiple devices can collaborate with each other in real time, and share results in less than a millisecond and with an energy expenditure of just ten to thirty microjoules per operation."
According to the researchers, this not only has implications from the point of view of energy consumption, but also from a social and ethical point of view, as it allows AI to be available to anyone and reinforces data privacy. This makes it possible for schools or hospitals, rural areas with limited infrastructure or groups of citizens with few resources to use efficient, sustainable, accessible and distributed artificial intelligence.
Towards efficient autonomous driving
The second of the studies, also led by Fernando Sevilla Martínez from the UOC and with the same participants as the previous one, analyzes in detail how impulse neural networks can reduce the energy consumption of autonomous driving systems, compared to convolutional networks, which are widely used in artificial vision systems such as those used in some autonomous vehicles. To do this, they compare both technologies in tasks such as predicting steering wheel turning angles or detecting obstacles. The researchers' proposal also involves introducing a new way of measuring the real efficiency of systems, in order to achieve a better balance between accuracy and energy consumption.
"The tests we have carried out with different architectures show that impulse neural networks with a certain coding achieve an optimal balance between performance and low consumption, and use between ten and twenty times less energy than convolutional networks," explain researchers from the UOC's NeuroADaS Lab group, attached to the eHealth Centre. "This demonstrates that neural networks can drive more sustainable AI even without the need for specialized hardware, marking a key milestone towards efficient computing in intelligent and autonomous transportation," they add.
According to the authors, both studies provide valuable data in research to achieve AI systems that consume less energy and, therefore, are also more affordable and accessible. "The first work provides a practical workflow with less electrical demand, less heat generation and the possibility of deploying AI directly without data centers, the so-called edge computing," they conclude. "And the second introduces a metric that combines performance and energy consumption, allowing us to drive the design of more sustainable AI."
