A Software-Only Thermal Proxy for Edge AI Inference (Raspberry Pi 5)

Reproducible artifact for "CPU Utilization as a Software-Only Thermal Proxy for Multi-Tenant Edge AI Inference: A 13-Hour Characterization and Coupling Law on Raspberry Pi 5" (submitted to IEEE Embedded Systems Letters).

Summary

Thermal-aware control of edge inference usually assumes an instrumented power sensor, which most deployed single-board computers lack. From a 13-hour measurement campaign on Raspberry Pi 5 (Arm Cortex-A76), we show that CPU utilization alone, a fully software-observable signal, predicts SoC temperature through a compact, cross-validated coupling law:

T ≈ 45.6 °C + 0.175 °C/% · U     (R² = 0.93, slope 95% CI [0.1746, 0.1750], n = 232,136)

The slope is consistent across four independent experiment blocks, and the thermal response settles within one control interval, so the law is directly usable for real-time thermal control with no power instrumentation.

Dataset

data/ contains the de-duplicated real telemetry (Apache Parquet, 5 Hz), one collector file per experiment block: single-workload characterization, load sweep, two-tenant concurrency, and a 6-hour stability run. Columns include SoC temperature, whole-CPU utilization (one value for all four cores) and memory; the full column dictionary is data/README.md.

Data revision (Sep 2026). Earlier versions of these Parquet files carried voltage_v/current_a/power_w columns written by the collector as placeholders: no power sensor was attached, and forensics show the values were a random stub (constant 5.0 V, zero correlation with utilization or temperature, zero lag-1 autocorrelation). They contained no measurement and have been removed (scripts/strip_stub_power_columns.py) so the dataset cannot be mistaken for a power measurement. No analysis in the paper ever used them.

Scope note (stated in the paper): this deployment read no power sensor, so no board power is reported; the utilization law is offered as the software-only alternative. The CPU governor was pinned (performance), so the law characterizes the fixed-frequency regime.

Contents

paper/main_v2_rewrite.pdf   the paper
data/*.parquet              de-duplicated 232,136-sample telemetry (4 blocks)
scripts/paper8_deep.py      coupling-law fit, leave-one-block-out CV, dynamic time constant,
                            multi-tenant characterization
scripts/paper8_figs.py      regenerates the two figures
figures/                    coupling_law.pdf, slope_consistency.pdf
LICENSE                     MIT

State-of-the-art comparison (revision)

scripts/paper8_sota_comparison.py evaluates the standard temperature-prediction structures on this dataset under the same leave-one-block-out protocol as the law: a HotSpot-class lumped RC model driven by utilization, AR(1)/ARX forecasters on the temperature sensor, a multi-signal static regression, and an EWMA-filtered law. Result (mean held-out RMSE): every sensor-free, actuatable structure lands in the same 1.6-1.8 degC band as the two-constant law, while sensor-feedback forecasters are ~2.7x tighter at one 200 ms step but cannot predict the effect of an allocation change. Summary table: data/sota_comparison_summary.csv.

Reproduce

pip install -r requirements.txt
python scripts/check_headline.py   # README headline numbers, fails on drift
python scripts/paper8_deep.py   # every number in the paper
python scripts/paper8_figs.py   # the figures

Citation

See CITATION.cff (artifact DOI 10.5281/zenodo.21844859). The manuscript is under review at IEEE Embedded Systems Letters. MIT License.

Archived version

This artifact is archived on Zenodo. The concept DOI 10.5281/zenodo.21844859 always resolves to the latest release, and CITATION.cff carries the full metadata, which is what GitHub's "Cite this repository" button renders.