New research in Discover Artificial Intelligence compared three deep learning models trained to classify dermoscopic skin lesion images. ResNet-50, EfficientNet-B3, and Vision Transformer B/16 all landed within one percentage point of each other on overall accuracy, but training energy ranged from 0.0143 kWh to 0.0630 kWh per run. The vision transformer used about 4.4 times the energy of ResNet-50 without delivering a measurable accuracy gain. Under Turkish grid conditions, that translated to 5.5 to 24.4 grams of CO2 per training run, and inference energy gaps were similar. The study also found that parameter count did not predict energy use, and that a popular software estimator called CodeCarbon reported zero GPU energy on the test hardware. Direct board-level measurements were needed. The authors argue that clinical AI documentation should report measured training energy, carbon emissions under a specified grid, inference energy, and sensitivity on the most important diagnostic class. For anyone tracking the carbon cost of AI, the lesson is simple: accuracy alone hides a wide range of energy outcomes, and energy-aware model selection is a cheap way to cut emissions without hurting performance.
