Thermo-Aware Anytime Inference on Raspberry Pi 5: Uncertainty-Aware Control and Full-Stack Reproducibility
Year: 2025
Type: preprint
Abstract: Edge devices must satisfy latency service-level objectives (SLOs) while operating under tight thermal constraints [2], [3]. We present THERMO-ANYTIME, a thermally aware runtime for Raspberry Pi 5 [4] that dynamically selects among anytime inference policies-combinations of early exits [1], numeric precision [7], and CPU threading-to stabilize tail latency under temperature drift. The controller relies on a temperature→latency lookup table (LUT) built from short warm segments; each policy p is modeled as L90(p, T) ≈ ap + bpT with an uncertainty buffer that inflates predictions by a temperature-dependent standard error. This formulation yields a compact constrained selection problem over a discrete policy set with hysteresis to prevent oscillation [25]. We release a complete, reproducible artifact: packaged ONNX models (FP32/INT8 for two exits and the final head), Pi-side orchestration and loggers (temperature and latency segments), the LUT builder, and post-processing scripts that regenerate figures and L A T E X tables directly from logs [23]. On a Pi 5, THERMO-ANYTIME maintains stable p90 latency within its thermal cap while adapting policies in response to workload variation; quantized exits deliver most of the latency-energy gains [?]. The end-to-end bundle (single-command execution with automated analysis) enables exact reproduction on commodity hardware and forms a foundation for future extensions-learned or model-predictive controllers, cross-backend (ONNX/TFLite) comparison [34], and multi-tenant thermal coordination on embedded AI platforms.
Author: Manu Nicholas Jacob
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4415716366