When machine learning models retire, decay, or become obsolete: A review on algorithms, software, and hardware
M.Z. Naser
Abstract
The exponential growth of artificial intelligence (AI) has raised serious concerns about its environmental impact, with large-scale AI models now consuming substantial amounts of energy. From this lens, this review systematically examines the current landscape of energy consumption in AI systems across their lifecycle, from training to deployment and inference. This review also introduces a holistic framework for assessing and improving AI systems through six interconnected dimensions: grid-optimized scheduling, resource-efficient architectures, energy-aware algorithms, environmental impact metrics, algorithmic optimization, and adaptive computation. Furthermore, we evaluate the trade-offs between model performance and environmental impact and propose methodological standards for measuring and reporting energy consumption in AI research and development. The paper concludes with actionable best practices to foster a more sustainable AI ecosystem and identifies promising directions and persistent challenges in reducing AI's carbon footprint.