Cross-Machine Comparable Benchmark for Machine Learning Energy Consumption

Luc Dop, Nawmi Nujhat, Sabina Gradinariu, Vincent van Vliet.

Group 24.

Paper.Source code.

This paper introduces a machine-agnostic benchmark for measuring and comparing the energy consumption of machine learning models during inference. Built on top of MLPerf and extended with energy tracking tools like EnergiBridge and CarbonTracker, the benchmark enables fair cross-machine evaluations. Using EfficientNet and Vision Transformer on the MNIST dataset, the study highlights differences in model energy efficiency. The work aims to support sustainable software engineering by promoting standardized and transparent energy measurement practices.