Thermo–PID–Interface (TPI): Dual-Loop Control of Information Heat and Physical Heat for Stable Large-Model Inference on Edge Hardware
Nz7 (Independent Researcher) · Zenodo (CERN European Organization for Nuclear Research) · 2025
This work introduces the Thermo–PID–Interface (TPI), a dual-loop controlframework that couples the internal information–state dynamics of a neurallanguage model with the physical thermal behavior of an edge device.By defining an information–heat metric ω and stabilizing both ω and the SoCtemperature Θ through cascaded PID control, TPI enables stable pseudo-7B/13B/30Binference workloads on a Raspberry Pi 5. Experiments reveal a clear thermalfeasibility boundary: 30B-equivalent inference converges at 60°C but becomesphysically infeasible at 50°C. This study provides empirical evidence thatinformation processing exhibits thermodynamic signatures and that inferencecapability on edge devices is fundamentally constrained by thermal limits.