Normalizing Energy Consumption for Hardware-Independent Evaluation

TL;DR

This work tackles the problem of hardware-dependent training energy consumption in ML for signal processing by proposing a hardware-agnostic normalization framework based on reference points, regression type, and optional computational metrics. The approach is validated across four GPUs and 43 audio-tagging model configurations, with energy measured via CodeCarbon and forward-work quantified using FLOPs and parameter counts. Key findings show that dual reference points yield more robust normalization, and that incorporating FLOPs and/or parameters generally improves predictive accuracy, as assessed by metrics such as $R^2$ and $MSE$. These results enable more accurate, energy-aware ML development and fair cross-hardware comparisons in energy efficiency.

Abstract

The increasing use of machine learning (ML) models in signal processing has raised concerns about their environmental impact, particularly during resource-intensive training phases. In this study, we present a novel methodology for normalizing energy consumption across different hardware platforms to facilitate fair and consistent comparisons. We evaluate different normalization strategies by measuring the energy used to train different ML architectures on different GPUs, focusing on audio tagging tasks. Our approach shows that the number of reference points, the type of regression and the inclusion of computational metrics significantly influences the normalization process. We find that the appropriate selection of two reference points provides robust normalization, while incorporating the number of floating-point operations and parameters improves the accuracy of energy consumption predictions. By supporting more accurate energy consumption evaluation, our methodology promotes the development of environmentally sustainable ML practices.

Figures (7)

• Figure 1: Normalization of energy consumption using different reference points for the T4-A40 GPU pair.
• Figure 2: Normalization of energy consumption using different reference points for the T4-RTX GPU pair.
• Figure 3: Impact of reference point selection on the linear regression between GPU pairs using different data sampling strategies.
• Figure 4: Illustration of different regression types to model the energy consumption of the RTX-GTX hardware pair.
• Figure 5: Comparison of regression types among linear, polynomial and support vector regression for normalizing energy consumption across different hardware.