@article{MD3E11493,
title = "Optimizing On-Device VSR Performance via Adaptive Thermal Management",
journal = "The Journal of Korean Institute of Communications and Information Sciences",
year = "2025",
issn = "1226-4717",
doi = "10.7840/kics.2025.50.1.1",
author = "Hye-jin Park, Rhan Ha",
keywords = "super-resolution, thermal management, mobile device, intermittent computing, machine learning",
abstract = "VSR (Video Super Resolution) tasks on mobile devices cause a rapid rise in device temperature due to heavy computational load in GPU. Mobile systems commonly employ DVFS (Dynamic Voltage-and-Frequency Scaling) for heat dissipation, but it automatically reduces GPU frequency without considering application behavior, causing a sudden drop in task performance. The occurrence of thermal throttling causes sudden work delays, hindering task efficiency and user satisfaction. Therefore, we propose he first adaptive thermal management technique ATM (Adaptive Thermal Management system for VSR) to address thermal issues. ATM adaptively controls the inference stage of VSR tasks based on device temperature changes to mitigate the rate of temperature increase and delay the onset of thermal throttling. Experiments confirmed that ATM effectively prevents thermal throttling during a given task period, while also preventing a 1.56x decrease in model inference speed and achieving an overall task throughput improvement of 1.8x."
}