Is AI environmentally sustainable?
Workshop Details
Practical Information
None
Prerequisites: NONE
About this Workshop
The rapid proliferation of AI applications is placing significant pressure on infrastructure. The extensive computational requirements of AI algorithms and systems raise serious concerns about the long- and potentially medium-term sustainability of producing and operating the necessary servers and associated hardware.
This workshop addresses these concerns. The way environmental impacts of AI can be measured and modelled will be first discussed. Then efforts to design less unsustainable AI solutions will be presented. The workshop is organized around technical presentations, followed by active discussions between the speakers, the organizers, and the public.
Key take-aways for the public:
Environmental impacts of IA: what are we talking about? Impact measurements: how to do this? Who can do this? IA systems eco-design? Which trade-offs to explore Schedule and speakers: to be announced by mid-February
Workshop Program
13:30–13:40
Pamela Delgado, Sébastien Rumley, HES-SO
Context and workshop introduction
13:40–14:30
Peva Blanchard, Kleis Computing
Modeling AI environmental impacts: what does it look like?
14:30–14:40
Coffee break
14:40–15:10
Pamela Delgado, HES-SO, Sébastien Pittet, Exoscale
OptiCloud and Exoscale: toward sustainable and responsible cloud consumption for AI
Training, fine-tuning, and running AI models require specialized infrastructure with a high environmental impact. This presentation shows a concrete example of how this can be improved from both the cloud provider side and the cloud customer side.
15:10–15:30
Adam Grossenbacher, HEIA-FR
More Power, More Accuracy? An Energy-Aware Evaluation of Document AI
This study challenges the "scale is all you need" assumption by demonstrating that compact expert VLMs achieve state-of-the-art document parsing accuracy while consuming less energy than generalist models.
15:30–15:35
Coffee break
15:35–15:50
Loïc Guibert, HEIA-FR
Impact of a carbon tax on the total lifecycle cost of servers
We measured the electricity consumption of different machines performing AI workloads. We also estimated the carbon footprint of manufacturing those machines. We discuss how these elements can be combined and compared with real server costs.
15:50–16:20
Jean-Damien Beaud, e-Durable, Viet Hang Nguyen, DeepAlps.ch
Sustainable AI for SMEs: creating practical value with reused GPUs
Starting from concrete SME needs, we show how AI can already deliver value with reused GPUs, while clearly addressing key limits (energy, obsolescence, resources) and the need for better measurement and standardization to compare options and make informed decisions.
Speakers & Organizers
Peva Blanchard
Associate, Kleis Technology
Kleis Technology
Peva Blanchard, PhD, was a researcher in distributed algorithms (Univ Paris-Sud, EPFL) before joining Kleis Technology as an associate founder. Nowadays, his work focuses on automating environmental impact calculations.
Pamela Delgado
Associate Professor, HEIG-VD / HES-SO, Yverdon-les-Bains
IICT at HEIG-VD/HES-SO
Pamela Delgado is a professor at the IICT at HEIG-VD. Before joining HES-SO, she worked at the SDSC and obtained her PhD from EPFL. Her research focuses on efficiently managing large-scale/limited GPU resources at the intersection between systems and machine learning with a focus on sustainability.
Sébastien Rumley
Associate Professor, HEIA-FR, HES-SO, Fribourg
HEIA-FR, HES-SO
Sébastien Rumley is professor of software engineering in the engineering school of Fribourg. Expert in large, distributed computing systems architecture, his research interest lay at the intersection of energy systems, IT systems, and sustainability.