Skip to content
hw.dev
hw.dev/signal/dvfs-passive-cooling-edge-inference-rpi5-2026
SignalarXiv

Calibrated DVFS Eliminates Thermal Throttling on Passively Cooled Edge Inference Hardware

An empirically calibrated DVFS scheduler with time-domain guards eliminates all thermal throttling events during sustained 30-minute YOLOv8n workloads on a passively cooled Raspberry Pi 5, hitting 6.8% higher frame rate and 1.9% less energy per frame than a reactive baseline.

#embedded#ai-hardware#tools
Read Original

Passively cooled edge hardware eliminates fan failure modes and standby power, but the tradeoff has always been thermal throttling under sustained inference load. This paper maps the control problem precisely: a state-aware DVFS scheduler with absolute temperature bounds and derivative triggers that catch sharp thermal spikes before the governor steps in. On a passively cooled Raspberry Pi 5 running YOLOv8n continuously for 30 minutes, the calibrated scheduler eliminates all throttling events and beats a temperature-only reactive baseline by 6.8% frame rate (Cohen's d = 8.73) while consuming 1.9% less energy per frame. The open question of whether passive could beat active cooling on energy efficiency is now answered for this platform: yes, at the cost of raw throughput.

The dwell guard is the structural contribution. Without it, run-to-run reproducibility collapses because short-cycle frequency changes destabilize thermal state. The scheduler also maps the operating envelope boundary: at ambient temperatures above 27C, nonlinear leakage defeats DVFS-based control entirely. That boundary is a product spec for every passively cooled edge device running inference in summer field conditions or enclosures with poor convection, and until now it was typically discovered during board-bring-up rather than at design time.

The code and datasets are open source on GitHub. The meaningful takeaway for embedded teams is that fan-less AI edge products are viable across a wider operating envelope than thermal derate curves suggest, but only with a calibrated governor -- not a generic SoC thermal policy. Teams building inference appliances for outdoor, industrial, or medical enclosures where mechanical cooling is impractical now have a validated scheduler pattern and a clear constraint: size the thermal mass for the 27C ambient ceiling, or accept that the operating envelope closes.