Sustainable Edge Vision via Empirically Calibrated DVFS: Eliminating Thermal Throttling on Passively Cooled Hardware
Aayush Marasini, Zhaoxian Zhou
cs.AR · cs.CV · cs.LG
Quick verdict
Calibrated DVFS scheduling reportedly eliminates throttling on a passively-cooled Pi 5 at lower J/frame than active cooling; careful single-board study, but baseline and boundary evidence are weak.
Summary
The paper presents an empirically calibrated, state-aware DVFS scheduler for passively cooled edge SoCs, evaluated on a Raspberry Pi 5 running YOLOv8n (RDD2022 subset, OpenVINO FP32) for sustained 30-minute inference. Scheduler thresholds derive from measured sensor noise (σT=0.58 °C), derivative noise, an RC step-response time constant (10 s), and calibration-phase thermal plateaus, rather than manual tuning; absolute temperature bounds dominate, with a dwell guard (20 s) and a derivative trigger retained as a spike safeguard. The scheduler records zero throttling events over 30 minutes, versus 1823 for unthrottled full-speed operation, and exceeds a temperature-only reactive baseline by 6.8% FPS while using 1.9% less energy per frame; passive scheduling also beats an actively cooled reference on J/frame (0.531 vs 0.563) while giving up 18.6% throughput. A dwell-guard ablation shows removing it raises run-to-run FPS standard deviation sixfold with negligible mean change. Single-run boundary probes suggest the passive envelope closes near 27 °C ambient, where nonlinear leakage defeats DVFS control. Key limitations: one platform and workload, three replications per condition, exploratory single-run boundary findings, and an INT8 result specific to the OpenVINO stack.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An empirically calibrated, state-aware DVFS scheduler eliminates all observed thermal throttling during sustained 30-minute YOLOv8n inference on a passively cooled Raspberry Pi 5 (vs 1823 events uncontrolled).
- The guarded scheduler beats a temperature-only reactive baseline by 6.8% FPS (d=8.73) with 1.9% lower energy per frame, because it retains access to the efficient S1 state that the reactive policy cannot reach.
- Passive calibrated scheduling achieves lower energy per frame than an actively cooled reference (0.531 vs 0.563 J/frame) while delivering 81.4% of its throughput.
- The dwell guard is necessary for run-to-run reproducibility: removing it raises FPS standard deviation about sixfold at a negligible (<1%) throughput cost.
- The passive operating envelope closes at approximately 27 °C ambient, where nonlinear leakage current defeats DVFS-based control (established only by single-run exploratory probes).
Strengths
Calibration-first controller design with derived, frozen thresholds
Section III.E/V-F derive hysteresis from sensor noise (σT=0.5835 °C → 3σT≈1.75 °C), the proactive trigger from derivative noise (6.6×σṪ=0.5 °C/s), and dwell from a step-response RC time constant (2×τ=20 s); Table II set-points were fixed before evaluation and never retuned, addressing the classic tuned-on-test-set criticism.
Mechanistic explanation of the baseline gap, not just an aggregate win
Section V-B/VI-E show Proactive holds S1 for 54.2% of the run while Reactive spends 0% time in S1; Section VI-E explains the lock-in structurally: Trec(S2→S1)=71 °C sits below the S2 plateau (73.4–73.9 °C), so the recovery condition can never fire.
Well-isolated ablation attributing gains to the dwell guard
Strong measurement and reproducibility discipline
Honesty about negative and null results
Weaknesses
Entire study rests on one board, one workload, one model, one ambient band
Small n and a headline boundary established by a single exploratory run
Reactive baseline comparison is partly predetermined by threshold placement
Key fine-grained comparisons are asserted without statistical testing
The INT8 result likely reflects a quantization configuration pathology
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This systems paper proposes an empirically calibrated, state-aware DVFS scheduler for passively cooled edge SoCs and evaluates it on a single Raspberry Pi 5 running YOLOv8n (RDD2022 subset, OpenVINO FP32 IR) for 30-minute sustained inference. All guard parameters derive from on-device measurements — sensor noise (σT=0.5835 °C → ≥3σT hysteresis floor), post-EMA derivative noise (6.6×σṪ = 0.5 °C/s), a step-response RC time constant (τ=10 s → 20 s dwell) — plus calibration-phase thermal plateaus (Table II), fixed before evaluation. Central claims: (1) zero throttle events in 30 minutes vs 1823 for uncontrolled S0; (2) +6.8% FPS (d=8.73) and −1.9% J/frame versus a temperature-only reactive baseline; (3) lower J/frame than an actively cooled reference (0.531 vs 0.563) at 81.4% of its throughput; (4) the dwell guard is necessary for run-to-run reproducibility; (5) the passive envelope closes at ≈27 °C ambient.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does DVFS scheduling remove the need for a fan on a Raspberry Pi 5 running continuous YOLO inference?
According to this paper, yes — but only below roughly 27 °C ambient temperature. A calibrated three-state DVFS scheduler (2400/1800/1500 MHz) sustained 30 minutes of YOLOv8n inference on a fanless Pi 5 with zero thermal-throttling events, while an uncontrolled full-speed run throttled 1823 times as it hit the 85 °C hardware limit (plateau ≈84.8 °C). Single exploratory probes at 27 °C and 31 °C ambient showed throttling returns, because non-linear leakage current defeats the lowest-frequency state.
How much more efficient is passive cooling than a fan for edge AI inference?
The paper reports the passive scheduler consumed 0.531 J per frame versus 0.563 J/frame for the actively cooled reference (about 5.7% less energy per frame), because the fan draws extra power and the passive scheduler spends time in low-frequency states. The trade-off: the passive system delivered 81.4% of the active system's throughput (11.82 vs 14.53 FPS). Note this efficiency win was reported without a statistical test, at n=3 runs per condition.
Why does a simple temperature-threshold controller perform worse than this scheduler?
In the paper's setup, a reactive threshold-only baseline cascades to the lowest frequency state (S2, 1500 MHz) early in the run and its recovery threshold (71 °C) sits below the S2 thermal plateau (≈73.4–73.9 °C), so it can never climb back to the efficient mid state (S1, 1800 MHz) — it spends 0% of the run in S1. The guarded scheduler uses dwell times and confirmation samples to hold transitions and spends 54.2% of the run in S1, yielding 6.8% higher FPS and 1.9% lower energy per frame. A reviewer-style caveat: that recovery threshold choice itself partly pre-determines the baseline's failure.
Is the paper's code and data available for reproduction?
Yes, the paper points to a public GitHub repository (github.com/Aayush-Marasini/sustained-edge-vision) with SHA256-frozen model weights, dataset partitions, and the benchmark video, plus per-run metadata and fixed seeds. However, exact replication requires the same physical setup: a Raspberry Pi 5 with the official active cooler (fan disconnected), the official 27 W USB-C PSU, an inline POWER-Z KM003C power meter, and a DHT11 ambient sensor.
What are the main limitations of this DVFS study?
Everything was measured on a single Raspberry Pi 5 board with a single workload (YOLOv8n) at 23±2 °C ambient, with n=3 runs per condition. The 27 °C thermal boundary comes from single, explicitly exploratory probe runs (Table VIII). The INT8 finding (−43% FPS) is specific to the OpenVINO 2026.0.0 runtime on Cortex-A76 and is acknowledged as stack-dependent. The scheduler's derivative (dT/dt) trigger never actually fired during the nominal-ambient experiments, so its individual contribution is untested.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The review is based on the extracted PDF text only; figures (Figs. 1-5) are unavailable, so visual claims (e.g. the S0→S1→S2→S1 trajectory in Fig. 4, Pareto layout in Fig. 5) could not be independently checked.
- The cited GitHub repository was not inspected; repository contents, completeness, and integrity of the frozen artifacts are taken from the paper's statements.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.