EVENT DOSSIER
The researchers proposed an empirically calibrated state-aware dynamic voltage and frequency scaling (DVFS) scheduler aimed at addressing thermal throttling issues in passive cooling hardware. This scheduler combines time-domain guards, absolute temperature limits, and derivative-triggered mechanisms as protective strategies. In tests with the YOLOv8n model on a Raspberry Pi 5 with passive cooling, this method successfully eliminated all thermal throttling events during a 30-minute workload. Compared to a reactive baseline based solely on temperature, the frame rate increased by 6.8%, and per-frame energy consumption decreased by 1.9%; the optimized passive scheduling was more efficient in terms of energy efficiency than the active cooling reference system, although the active cooling system remained superior in original throughput. Ablation experiments confirmed that the presence of idle guards is crucial for running repeatability, and when the ambient temperature is 27°C or higher, non-linear leakage currents can disrupt the control mechanism, causing the passive operating range to be shut down.
研究人员提出了一种经验校准的状态感知动态电压和频率缩放（DVFS）调度器，用于消除被动冷却硬件上的热节流问题。该调度器利用时域守卫、绝对温度界限及导数触发机制作为防护，在被动冷却的 Raspberry Pi 5 上运行 YOLOv8n 模型时，成功消除了持续 30 分钟工作负载中的所有热节流事件。与仅基于温度的反应式基线相比，该方法帧率高出 6.8%，每帧能耗降低 1.9%；其优化后的被动调度在能效上优于主动冷却参考系统，尽管主动冷却在原始吞吐量方面仍更优。隔离消融实验表明，驻留守卫对运行可重复性至关重要，而边界探测显示当环境温度高于或等于 27°C 时，非线性漏电流会破坏基于 DVFS 的控制，导致被动工作范围关闭。