Sustainable Computing Optimization in Large-Scale Machine Learning Training

The rapid growth of large-scale machine learning (ML) models has led to unprecedented computational demands, raising critical concerns about energy consumption, and environmental sustainability. To address these challenges, it is imperative to develop innovative approaches that reduce resource usage without sacrificing model performance. In this paper, we propose sustainable adaptive green engine (SAGE), a comprehensive framework designed to optimize computational efficiency during large-scale ML training. SAGE integrates multiple key techniques, including dynamic model sparsification that adaptively prunes redundant parameters to reduce computational overhead; resource-aware training scheduling that flexibly adjusts training processes based on real-time energy consumption; and communication-efficient distributed learning strategies to minimize overhead in multi-node environments. By combining these modules, SAGE dynamically adapts model complexity and computation pathways according to workload characteristics and energy constraints. This adaptive approach enables the system to maintain high accuracy even under stringent resource limitations, making it particularly suitable for deployment on energy-sensitive platforms such as mobile and edge devices. Furthermore, the framework significantly reduces overall energy consumption during training, promoting sustainable AI development while preserving scalability and generalization capabilities. Extensive evaluations on diverse benchmark datasets demonstrate that SAGE achieves a balanced and effective solution across performance, computational cost, and sustainability metrics. Our work represents a practical step toward environmentally friendly AI, providing a valuable blueprint for future research on green and efficient ML systems.