Phoeni6: a Systematic Approach for Evaluating the Energy Consumption of Neural Networks

TL;DR

Phoeni6 introduces a systematic, containerized framework to evaluate neural network energy consumption under fair comparison and reproducibility constraints. It integrates a NoSQL database model, modular containerized tools, and an eleven-step methodology to automate energy evaluations and data management. Two case studies (AlexNet vs MobileNet; image format effects) show that network design and data formats significantly influence energy use, with MobileNet typically more energy-efficient and BMP formats more favorable than PNG. The work emphasizes portability, transparency, and reproducibility, providing open-source tooling to support scalable, energy-aware AI development.

Abstract

This paper presents Phoeni6, a systematic approach for assessing the energy consumption of neural networks while upholding the principles of fair comparison and reproducibility. Phoeni6 offers a comprehensive solution for managing energy-related data and configurations, ensuring portability, transparency, and coordination during evaluations. The methodology automates energy evaluations through containerized tools, robust database management, and versatile data models. In the first case study, the energy consumption of AlexNet and MobileNet was compared using raw and resized images. Results showed that MobileNet is up to 6.25% more energy-efficient for raw images and 2.32% for resized datasets, while maintaining competitive accuracy levels. In the second study, the impact of image file formats on energy consumption was evaluated. BMP images reduced energy usage by up to 30% compared to PNG, highlighting the influence of file formats on energy efficiency. These findings emphasize the importance of Phoeni6 in optimizing energy consumption for diverse neural network applications and establishing sustainable artificial intelligence practices.

Figures (24)

• Figure 1: Overview of the Phoeni6 system architecture, illustrating the key components and their interactions for energy consumption evaluation. The modular design of Phoeni6 is highlighted, showcasing its scalability potential.
• Figure 2: Steps of the proposed methodology for energy evaluation, covering all phases from setup to energy calculation.
• Figure 4: Data model for storing investigation results, detailing attributes for energy consumption evaluation.
• Figure 5: This diagram illustrates the coordination of containerized applications during the energy evaluation process. It highlights the steps managed by the Manager-app, including initialization, data collection, warming up devices, neural network execution, and logging results. This workflow ensures adherence to fair comparison (FC) and reproducibility (RR) principles by standardizing the evaluation stages.
• Figure 6: Initial setup steps for the first case study: This diagram illustrates the workflow used in the first case study to evaluate the energy consumption of neural networks when varying image sizes. The process begins with dataset preparation, where the original dataset is filtered and resized to generate additional images for evaluation. Phoeni6 coordinates the classification process, leveraging its modular architecture, including job-apps for resizing and filtering datasets. These steps ensure reproducibility, portability, and scalability of the energy consumption analysis while maintaining consistency across neural network models.