GreenAuto : An Automated Platform for Sustainable AI Model Design on Edge Devices Conference: The 26th International Workshop on Mobile Computing Systems and Applications; February 26–27, 2025; La Quinta, CA, USA The 26th International Workshop on Mobile Computing Systems and Applications (HOTMOBILE

Xiaolong Tu, Dawei Chen, Kyungtae Han, Onur Altintas, Haoxin Wang · 2025

Read GreenAuto : An Automated Platform for Sustainable AI Model Design on Edge Devices Conference: The 26th International Workshop on Mobile Computing Systems and Applications; February 26–27, 2025; La Quinta, CA, USA The 26th International Workshop on Mobile Computing Systems and Applications (HOTMOBILE by Xiaolong Tu, Dawei Chen, Kyungtae Han, Onur Altintas, Haoxin Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-14995.txt.

Introduction

Xiaolong Tu and four co-authors build GreenAuto, a platform that designs smart models for small edge tools with an eye on low energy use. Their paper is from two thousand twenty-five.

The platform runs the full path: search, build, ship, and grade. Its search rides a front of best trade-offs, in a wide design space, steered by slope steps. Fast gauges of each small core's energy give a world view of the cost of each pick. That view steers the hunt toward lean models. The platform takes its own measures and refines the hunt as it goes. It finds lean, strong models with no hand work at all.

Catalog record

CREATORXiaolong Tu, Dawei Chen, Kyungtae Han, Onur Altintas, Haoxin Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-14995

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GreenAuto : An Automated Platform for Sustainable AI Model Design on Edge Devices Conference: The 26th International Workshop on Mobile Computing Systems and Applications; February 26–27, 2025; La Quinta, CA, USA The 26th International Workshop on Mobile Computing Systems and Applications (HOTMOBILE ’25), February 26–27, 2025, La Quinta, CA, USA DOI: 10.1145/3708468.3711890 ISBN: 979-8-4007-1403-0/25/02 CCS: Computing methodologies Neural networks CCS: Human-centered computing Ubiquitous and mobile computing systems and tools Xiaolong Tu, Dawei Chen, Kyungtae Han, Onur Altintas, Haoxin Wang Abstract . We present GreenAuto , an end-to-end automated platform designed for sustainable AI model exploration, generation, deployment, and evaluation. GreenAuto employs a Pareto front-based search method within an expanded neural architecture search (NAS) space, guided by gradient descent to optimize model exploration. Pre-trained kernel-level energy predictors estimate energy consumption across all models, providing a global view that directs the search toward more sustainable solutions. By automating performance measurements and iteratively refining the search process, GreenAuto demonstrates the efficient identification of sustainable AI models without the need for human intervention. 1. Introduction Sustainable Artificial Intelligence (AI) refers to the development and application of AI technologies aimed at addressing environmental challenges and promoting sustainability ( wu2022sustainable ; tu2023deepen2023 ) . This field spans two key areas: the sustainability of AI itself and the use of AI for advancing sustainability ( van2021sustainable ; vinuesa2020role ) . While there is growing interest in utilizing AI to support environmental sustainability, research focusing on AI’s own sustainability remains relatively scarce. On-device learning, especially on mobile and edge devices, is gaining traction for model personalization and improved data privacy. However, the sustainability of on-device learning often receives little attention. One reason is that it is not intuitively measurable and cannot be assessed using conventional metrics. Another challenge lies in the typically short runtime of models on edge devices, which makes tracking sustainability difficult. Since most edge devices are electrically powered, and electricity generation is a significant source of carbon dioxide emissions ( electricitygeneration ) , energy consumption serves as a practical and quantifiable proxy for evaluating sustainability. The importance of energy efficiency becomes even more apparent when considering the scale of edge devices. By 2030, the number of edge devices is projected to reach 40 billion ( estimate ) , their collective carbon footprint could surpass 292.42 million metric tons annually, 1.145 times France’s carbon emissions in 2023. A major contributor to this carbon footprint is the runtime of AI models on these edge devices, such as video analytics and voice assistants ( wang2020energy ; wang2022leaf+ ) . As AI-driven applications become more prevalent, their cumulative energy consumption across billions of devices will have a significant impact on global carbon emissions. Addressing this pressing issue requires a deeper understanding of how AI models consume energy during runtime and how to optimize their energy efficiency for specific edge hardware. This optimization is crucial for reducing the environmental impact of AI at scale. As the prevalence of on-device learning grows, integrating sustainability into its design and evaluation processes is essential to align technological advancements with environmental goals. To address this, we design GreenAuto , an end-to-end automated platform to optimize AI models’ sustainability from model itself and the model search process. We use energy consumption and converted carbon emissions as representatives, offering a straightforward approach for measurement and evaluation of the sustainability. GreenAuto is designed based on energy predictors, integrates with NAS search space, Pareto front-based model search, and automated energy measurements in a completed pipeline to search the sustainability models automatically. It demonstrates strong potential for sustainable AI model design by streamlining the search process and generating energy-efficient models. GreenAuto offers several unique features: Sustainable model search: GreenAuto extends existing NAS spaces and uses pre-trained energy predictors to estimate energy consumption for all model candidates. Pareto front optimization and gradient descent techniques are employed to efficiently balance energy efficiency and accuracy, significantly reducing training time and human intervention. Automated energy measurement: GreenAuto automates energy measurement on edge devices without human intervention. It iteratively and interactively collects energy and latency data to inform the model search. Scalability and flexibility: GreenAuto is modular and reconfigurable, allowing researchers to replace components, adjust parameters, and prioritize different objectives (e.g., accuracy or energy efficiency). This flexibility enables easy customization for diverse research needs and use cases. 2. System Design Figure 1 . System overview of GreenAuto , operating autonomously without human intervention. system pipeline In this section, we introduce GreenAuto , our end-to-end platform designed to automate the environmentally sustainable design of DNN models for edge devices. GreenAuto contains three essential components: (1) energy efficiency driven model search space (§ 2.1 ) that includes over 900,000 DNN models tailored for energy efficiency, (2) Pareto front-based model search algorithm (§ 2.2 ) that, combined with gradient descent-based model sampling, efficiently identifies promising model candidates from the search space, and (3) automated energy measurement function (§ 2.3 ) that leverages an external power monitor to capture accurate power consumption data during model execution. Fig. 1 illustrates the end-to-end pipeline of GreenAuto and its implementation. GreenAuto operates autonomously without requiring human intervention. The process begins by generating an energy efficiency-driven model search space. We create new DNN model variants by expanding the structural parameters, such as channel size, kernel size, and stride, of foundational model architectures derived from existing neural architecture search (NAS) spaces, like NAS-Bench-201 ( dong2020bench ) . This expanded search space focuses on models with greater potential to optimize the trade-off between accuracy and energy consumption. Next, GreenAuto uses our pre-trained kernel-level energy predictors ( 10419248 ) and an accuracy proxy to generate initial energy and accuracy estimates for each model. These estimates effectively accelerate the search process and reduce system overhead in identifying energy-efficient models. We design and implement a Pareto front-based optimization algorithm, integrated with gradient descent-based model sampling in GreenAuto , to iteratively explore the search space with minimal overhead. By leveraging energy and accuracy estimates as a guide, the algorithm prioritizes DNN models that better balance inference accuracy and energy consumption. After each iteration of exploration, an automated energy efficiency assessment is conducted. GreenAuto precisely profiles the actual energy consumption of selected model candidates when they are deployed and running on the target edge device using an external power monitor. The evaluation results from each iteration are continuously used to update the Pareto front, progressively refining the set of optimal models. Additionally, GreenAuto is highly reconfigurable. By adjusting search spaces, hardware devices, or objective weights, it can easily be adapted by other researchers to meet specific optimization needs, making it a versatile tool for a wide range of applications (discussed in § 2.4 ). Table 1 . Convolutional layer configurations Configurations Original space Expanded space Output channel 16, 32, 64 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 16, 32, 64, 128, 256 Kernel size 1, 3 1, 3, 5, 7 Stride 1 1, 2 2.1. Energy Efficiency-Driven Search Space GreenAuto initially expands the foundational search space to prioritize energy efficiency. For example, it extends NAS-Bench-201 ( dong2020bench ) , a widely used cell-based search space benchmark for comparing NAS algorithms. NAS-Bench-201 offers a standardized framework for evaluating cell-based architectures. Inspired by our previous work ( 10419248

Read the complete text