A new scientific paper introducing FLOPpy, an open-source Python library designed to establish a new paradigm for evaluating the computational cost of Machine Learning and Deep Learning algorithms, has been published. The library is now publicly available on PyPI and GitHub, enabling researchers and developers to use it and contribute to its ongoing development.

Evaluating the computational efficiency of Artificial Intelligence models has become a key challenge, particularly in the context of Green AI. Commonly adopted metrics, such as execution time and energy consumption, are highly dependent on the underlying hardware, system workload, and other external factors, making objective comparisons and the reproducibility of experimental results difficult.

FLOPpy addresses these limitations by introducing a hardware-independent approach that directly measures the algorithmic computational effort during model execution. The library dynamically tracks the computational workload and provides deterministic metrics, enabling fair and reproducible comparisons across models executed on different hardware platforms.

Key Features of FLOPpy

• Fully deterministic metrics, free from fluctuations caused by hardware or execution environments;
• Hardware-independent evaluation, allowing computational cost to be compared consistently across laptops, GPU clusters, and edge devices;
• Bit-Operation (BOP) support, enabling accurate measurement of the computational impact of quantization techniques, including INT8 and other low-precision representations;
• Seamless integration with PyTorch and Scikit-learn, requiring no modifications to model code while automatically tracking the entire training lifecycle, including the forward pass, backward pass, loss computation, and optimizer updates.

The goal of the project is to provide the scientific community with a rigorous, reproducible, and easy-to-use framework for evaluating the computational efficiency of Artificial Intelligence models, ultimately supporting the development of more sustainable and objectively comparable AI algorithms.

The work was carried out by Francesco Scala, Liliana Martirano, and Luigi Pontieri.

Further Information

Official website and documentation: https://floppy.icar.cnr.it/

Scientific article: https://doi.org/10.1016/j.softx.2026.102865