arXiv 2608.09998: Training phase accounts for majority of emissions across six DL models in CPU multi-label classification tests

The systematic review covers Green AI literature and compares existing carbon estimation tools before running an empirical test. Six models performed multi-label classification on CPU hardware with total emissions tracked across data preparation, training, and inference. Training consistently exceeded other lifecycle stages in CO2-equivalent output.

Results showed no linear relationship between architectural size and F1 score improvements, confirming earlier observations that parameter count alone does not predict performance returns. The study used direct CPU power logging rather than GPU-centric estimators common in prior work, producing lower absolute numbers but preserving the same phase ranking.

These measurements align with Strubell et al. 2019 findings on NLP energy costs and extend them to CPU environments where many production inference workloads now run. The absence of GPU data limits direct comparison to large-scale pretraining reports but highlights measurable emissions even at modest hardware scales.

Operational takeaway is that model selection criteria must incorporate measured training emissions rather than relying on parameter counts or reported accuracy alone. Future pipelines will require standardized reporting thresholds before scaling experiments.