Towards an Energy Consumption Index for Deep Learning Models: A Comparative Analysis of Architectures, GPUs, and...

Area of Science:

• Computer Science
• Environmental Science
• Electrical Engineering

Background:

• Growing global demand for computational resources in Artificial Intelligence (AI) applications raises concerns about energy consumption and environmental impact.
• Existing methods for evaluating Deep Learning (DL) model energy efficiency lack standardization and adaptability across diverse architectures.
• The need for a quantifiable metric to assess and compare the energy footprint of DL models is critical for sustainable AI development.

Purpose of the Study:

• To introduce and validate a novel energy consumption index for evaluating the energy efficiency of Deep Learning (DL) models.
• To provide a standardized and adaptable framework for comparing the energy consumption of various DL architectures.
• To analyze the energy consumption of representative DL models during training and inference phases.

Main Methods:

• Developed a new energy consumption index to quantify DL model energy efficiency.
• Selected representative DL architectures (AlexNet, ResNet18, VGG16, EfficientNet-B3, ConvNeXt-T, Swin Transformer) for case studies.
• Utilized sensor-based (OpenZmeter v2) and software-based (CarbonTracker, CodeCarbon) tools for energy measurements on TITAN XP and GTX 1080 GPUs using the Imagenette dataset.

Main Results:

• Demonstrated significant variations in energy efficiency among different DL architectures and GPU configurations.
• Quantified energy consumption differences during both training and inference stages for the evaluated models.
• Highlighted the trade-offs between model performance and energy utilization across tested architectures.

Conclusions:

• The proposed energy consumption index offers a flexible and standardized approach for evaluating DL model energy efficiency.
• The findings provide crucial insights for optimizing AI systems towards greater sustainability and reduced environmental impact.
• This framework supports accurate energy evaluations applicable to diverse computational settings, advancing sustainable AI practices.