<![CDATA[The exponential escalation of computational requirements for training and deploying Deep Learning models has precipitated an energy crisis, necessitating a critical reevaluation of the trade-off between algorithmic performance and environmental sustainability. This study aims to reconcile these conflicting demands by developing and validating a novel Dynamic Energy-Aware Pruning (DEAP) framework designed to maximize inference efficiency without compromising predictive accuracy. Employing a rigorous quantitative experimental design, we benchmarked state-of-the-art neural architectures, including ResNet-50 and Large Language Models (LLMs), across diverse hardware environments. The research utilized real-time telemetry to measure total energy consumption (Joules), thermal output, and carbon intensity () against standard accuracy metrics. Empirical results demonstrate that the proposed framework achieved a 42% reduction in energy consumption and stabilized hardware thermals, while maintaining predictive performance within a strict 1.5% non-inferiority margin compared to dense baselines. We definitively conclude that algorithmic sparsity effectively decouples high-level intelligence from excessive power usage, establishing a viable engineering paradigm for “Green AI” that aligns the trajectory of artificial intelligence with global decarbonization targets.]]>
Copyrights © 2025
