Hardware Dimensioning for Environmental Sustainability: benchmark of AI algorithms and environmental impact

Artificial Intelligence (AI) has seen unprecedented performance improvements in recent years, surpassing human capabilities in tasks such as image classification, natural language inference, and generative modeling. These advancements have been largely driven by the exponential increase in computational resources dedicated to training state-of-the-art AI models, with compute requirements growing by a factor of 4-5x per year since 2010. However, this rapid expansion has also resulted in a significant rise in energy consumption and $\mathsf{CO_2}$ emissions, contributing to AI’s growing environmental impact. For instance, the training of Meta AI’s LLaMA models required $2,638$ $\mathsf{MWh}$ of energy, leading to approximately $1,015$ $\mathsf{tCO_2eq}$ emissions. This work addresses the sustainability challenges of AI by extending HADA (HArdware Dimensioning for AI Algorithms), a framework that leverages Machine Learning (ML) and optimization to determine the optimal algorithm-hardware configuration under performance and budget constraints. Our contribution expands HADA to incorporate energy consumption and carbon emissions as key optimization criteria, allowing for carbon-aware hardware selection. Through benchmarking experiments on local machines and high-performance computing (HPC) clusters, we demonstrate how AI workloads can be optimized not only for efficiency and cost but also for sustainability. This research highlights the potential of hardware-aware AI resource allocation in mitigating AI’s environmental footprint, contributing to the broader goal of green and sustainable AI computing.