On April 25, 2025, INRIA, a prominent partner in the SustainML project, presented significant research findings at the CHI Conference on Human Factors in Computing Systems (CHI 2025). Their paper, titled "Should I Choose a Smaller Model?: Understanding ML Model Selection and Its Impact on Sustainability", delves deeply into the sustainability considerations within Machine Learning (ML) model selection processes, underscoring the pressing need for environmentally responsible computing practices.
Understanding Sustainability in Machine Learning
Machine Learning models, particularly large models like Large Language Models (LLMs), have become integral to numerous applications across sectors. However, their complexity and widespread deployment have raised substantial sustainability concerns, primarily due to their intensive computational requirements and associated carbon footprints. The INRIA paper emphasizes that while most sustainability efforts traditionally focus on optimizing training processes, the inference phase—the stage when ML models are actively deployed—can significantly outweigh the training phase in environmental impact. This shift in resource consumption highlights the necessity for a more comprehensive, lifecycle-oriented sustainability assessment of ML applications.
Insights from Interviews with ML Developers
To gain deeper insights, INRIA researchers conducted detailed interviews with ML developers, uncovering several critical aspects of current practices:
- Sustainability Often Overlooked: Developers predominantly focus on performance metrics such as accuracy and interpretability, leaving sustainability concerns frequently misunderstood or completely neglected.
- Misunderstanding of Energy Consumption: Many developers underestimate the environmental costs associated with the inference phase, failing to recognize that inference often surpasses training in terms of energy use and environmental impact.
- Limited Use of Sustainability Tools: Despite available tools like CodeCarbon and Tracarbon that track energy consumption and carbon emissions, developers rarely integrate these into their regular workflows due to lack of awareness and standardized methods.
INRIA’s Recommendations for Enhancing Sustainability
The INRIA study provides several actionable recommendations aimed at fostering sustainability in ML practices:
- Education and Awareness: Sustainability considerations should be systematically integrated into ML education and professional development programs to raise awareness among developers.
- Transparency and Standardized Reporting: Establishing clear, standardized methods for measuring and publicly reporting the sustainability impacts of ML models can significantly improve transparency and accountability.
- Balanced Decision-Making: Encouraging developers to critically evaluate whether smaller, more specialized models can adequately fulfill their tasks, potentially reducing unnecessary energy consumption and environmental impact.
Strategic Alignment with SustainML
These findings resonate strongly with the goals and philosophy of the SustainML project, which aims to embed sustainability deeply within the ML lifecycle. INRIA's insights underline the critical need for systemic changes in model selection practices, suggesting that:
- Sustainability should be integrated throughout the ML lifecycle management.
- Developers and organizations should prioritize task-specific, energy-efficient models over unnecessarily complex general-purpose models.
- Enhanced documentation and transparency regarding the environmental impacts of ML practices are crucial for informed decision-making.
Moving Towards Sustainable ML Practices
The research presented by INRIA at CHI 2025 represents a substantial advancement toward embedding sustainability in ML development processes. By highlighting the urgent need for systemic change, INRIA encourages developers, researchers, and organizations to adopt more responsible and sustainable ML practices. SustainML is positioned to leverage these insights to continue developing innovative frameworks and tools that promote sustainable computing, aligning technological advancement harmoniously with environmental responsibility.
All the SustainML participants: DFKI, Inria, IBM, University of Copenhagen, UpMem, Technische Universität Kaiserslautern, and eProsima.
MORE INFORMATION ABOUT SUSTAINML:
If you have any questions, please get in touch with
This project has received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement No 101070408.