Operationalizing Green AI Analytics: A Reproducible Framework for Monitoring Model Complexity Across Classical and Deep Learning
Main article
Abstract
Green AI has moved the discussion on artificial intelligence from model accuracy alone to the broader question of how computationally sustainable, comparable, and reproducible model development can become. Yet in practice, complexity reporting remains fragmented. Classical machine learning studies often report fit time or prediction speed, deep learning studies usually foreground parameter counts or forward-pass FLOPs, and work on quantized or compressed models frequently emphasizes memory savings without offering a unified account of training-, inference-, and precision-sensitive workload. This article develops a reproducible Green AI analytics framework for operational monitoring of model complexity across both classical and deep learning pipelines. The framework integrates four analytic layers—pipeline metadata, workload metrics, reproducibility controls, and decision analytics—and evaluates 104 standardized benchmark configurations spanning classical learners, compact vision models, encoder models, and large language models under multiple precision regimes. The analysis shows three consistent patterns. First, BOP-aware monitoring changes model rankings in ways that FLOPs-only reporting often misses, particularly under INT8 and INT4 settings. Second, lightweight classical models remain highly competitive on the efficiency frontier for structured-data tasks, whereas compact neural architectures dominate only after accuracy thresholds exceed roughly 0.85. Third, adding reproducibility metadata and phase-specific reporting materially improves cross-study interpretability. Regression results further indicate that precision-aware workload measures explain a larger share of modeled sustainability burden than FLOPs alone. Rather than proposing another isolated profiler, the article demonstrates how Green AI analytics can be operationalized as a reporting and benchmarking practice that supports transparent model selection, deployment governance, and efficiency-aware research design.
