Energy Profiling of Deep Neural Networks on Android Devices: Benchmarking and Analysis

Sowmiya Sree C., S. R. Dwaraknaath, Dhanush Kiran Jai

2025

Abstract

The deployment of deep neural networks (DNNs) on mobile devices, such as those running the Android operating system, has become increasingly prevalent due to the growing demand for on-device AI capabilities in applications like image recognition, natural language processing, and augmented reality. However, the interplay between DNN architectural design parameters and the underlying hardware of mobile devices leads to non-trivial interactions that significantly impact performance metrics such as energy consumption and runtime efficiency. These interactions are further complicated by the diversity of Android devices, which vary widely in terms of processing power, memory capacity, and hardware accelerators like GPUs and NPUs. As a result, understanding the energy usage characteristics of DNNs on Android devices is critical for designing energy-efficient architectures and optimizing neural networks for real-world applications. Beyond the technical development of the app, this project seeks to explore the broader implications of energy usage in DNNs on mobile devices. Through extensive benchmarking, we will analyze how factors such as model complexity, hardware-software interactions, and optimization techniques like quantization and pruning influence energy efficiency. For instance, we will investigate how the number of layers, filters, and operations in a DNN affect energy consumption, and how different processors and accelerators (e.g., Snapdragon vs. Exynos, GPU vs. NPU) impact performance. We will also compare the energy efficiency of popular inference frameworks like TensorFlow Lite and PyTorch Mobile, shedding light on the trade-offs between accuracy, latency, and energy consumption. This project bridges the gap between deep learning research and practical mobile application development by providing a tool for benchmarking DNN energy usage and offering actionable insights for designing energy-efficient neural networks. By addressing the critical need for sustainable AI solutions, this work contributes to the ongoing effort to make on-device AI more accessible, efficient, and environmentally friendly. The Android app and accompanying analysis will serve as valuable resources for researchers, developers, and practitioners seeking to optimize DNN performance on mobile devices, ultimately enabling the next generation of intelligent, energy-efficient applications.

Paper Citation

in Harvard Style

C. S., Dwaraknaath S. and Jai D. (2025). Energy Profiling of Deep Neural Networks on Android Devices: Benchmarking and Analysis. In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies - ICRDICCT`25; ISBN 978-989-758-777-1, SciTePress, pages 558-564. DOI: 10.5220/0013886400004919

in Bibtex Style

@conference{icrdicct`2525,
author={Sowmiya C. and S. Dwaraknaath and Dhanush Jai},
title={Energy Profiling of Deep Neural Networks on Android Devices: Benchmarking and Analysis},
booktitle={Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies - ICRDICCT`25},
year={2025},
pages={558-564},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013886400004919},
isbn={978-989-758-777-1},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies - ICRDICCT`25
TI - Energy Profiling of Deep Neural Networks on Android Devices: Benchmarking and Analysis
SN - 978-989-758-777-1
AU - C. S.
AU - Dwaraknaath S.
AU - Jai D.
PY - 2025
SP - 558
EP - 564
DO - 10.5220/0013886400004919
PB - SciTePress