Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Computation to Consumption: Exploring the Compute-Energy Link for Training and Testing Neural Networks for SED Systems

Constance Douwes, Romain Serizel

TL;DR

The paper investigates how compute cost relates to energy consumption in training and testing neural networks for sound event detection, focusing on MLP, CNN, RNN, and CRNN architectures on DESED-based audio tagging. Using forward FLOPs measured by a profiler and backward FLOPs approximated by a $2:1$ ratio, the study links energy to computation while tracking GPU energy with CodeCarbon and GPU/memory usage with Nvidia SMI. Results reveal architecture-dependent relationships: energy generally scales with $FLOPs$, but the strength and shape of this relation vary across MLP/RNN versus CNN/CRNN, and training often incurs higher energy than testing due to memory traffic. A key finding is the strong correlation between energy and GPU utilization across phases, suggesting GPU load as a practical predictor for energy and supporting the development of architecture-aware, hardware-normalized energy indicators for greener AI in audio processing.

Abstract

The massive use of machine learning models, particularly neural networks, has raised serious concerns about their environmental impact. Indeed, over the last few years we have seen an explosion in the computing costs associated with training and deploying these systems. It is, therefore, crucial to understand their energy requirements in order to better integrate them into the evaluation of models, which has so far focused mainly on performance. In this paper, we study several neural network architectures that are key components of sound event detection systems, using an audio tagging task as an example. We measure the energy consumption for training and testing small to large architectures and establish complex relationships between the energy consumption, the number of floating-point operations, the number of parameters, and the GPU/memory utilization.

From Computation to Consumption: Exploring the Compute-Energy Link for Training and Testing Neural Networks for SED Systems

TL;DR

The paper investigates how compute cost relates to energy consumption in training and testing neural networks for sound event detection, focusing on MLP, CNN, RNN, and CRNN architectures on DESED-based audio tagging. Using forward FLOPs measured by a profiler and backward FLOPs approximated by a ratio, the study links energy to computation while tracking GPU energy with CodeCarbon and GPU/memory usage with Nvidia SMI. Results reveal architecture-dependent relationships: energy generally scales with , but the strength and shape of this relation vary across MLP/RNN versus CNN/CRNN, and training often incurs higher energy than testing due to memory traffic. A key finding is the strong correlation between energy and GPU utilization across phases, suggesting GPU load as a practical predictor for energy and supporting the development of architecture-aware, hardware-normalized energy indicators for greener AI in audio processing.

Abstract

The massive use of machine learning models, particularly neural networks, has raised serious concerns about their environmental impact. Indeed, over the last few years we have seen an explosion in the computing costs associated with training and deploying these systems. It is, therefore, crucial to understand their energy requirements in order to better integrate them into the evaluation of models, which has so far focused mainly on performance. In this paper, we study several neural network architectures that are key components of sound event detection systems, using an audio tagging task as an example. We measure the energy consumption for training and testing small to large architectures and establish complex relationships between the energy consumption, the number of floating-point operations, the number of parameters, and the GPU/memory utilization.
Paper Structure (15 sections, 4 figures, 1 table)

This paper contains 15 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Computational cost
  4. Energy consumption
  5. Experiments
  6. Task description
  7. Models
  8. Training and test
  9. Results
  10. Relationship between energy and computational cost at test
  11. Relationship between energy and computational cost at training
  12. Training and test comparisons
  13. GPU and memory utilization
  14. Discussion and future works
  15. Conclusions

Figures (4)

  • Figure 1: Energy consumption at test for various neural network architectures and configurations, as a function of FLOPs (top) and of parameters (bottom). The three columns show: (1) all architectures together, (2) only MLP/RNN (in blue and green), and (3) only CNN/CRNN (in red and purple).
  • Figure 2: Energy consumption for training various neural network architectures and configurations, as a function of FLOPs (top) and of parameters (bottom). The three columns show: (1) all architectures together, (2) only MLP/RNN (in blue and green), and (3) only CNN/CRNN (in red and purple).
  • Figure 3: Average power during training (circles) and test (triangles) as a function of FLOPs/S.
  • Figure 4: Relationship between the energy consumption and the GPU utilization (left) and memory utilization (right) for training and test.