Workload-Aware Performance Profiling of IoT Microcontrollers: A Cross-Architecture Benchmark Across Compression, Cryptography and Signal Processing

Abstract

This study compares active-computation energy and latency for twelve compression, security/integrity and signal-processing workloads on ESP32, RP2040 and STM32F446. Common scalar C kernels process fixed 2048-byte inputs on one application core under two operating profiles: maximum nominal CPU clocks of 240, 200 and 180 MHz, respectively, and a common nominal 160 MHz. A Nordic PPK2 supplies each target domain at 3.3 V and records the current and one execution marker at nominal 100 kS/s. External regulators are removed on ESP32/RP2040; Nucleo retains its LDO with JP6 open. Sixty separate power-on captures provide ten observations per board and profile for arithmetic means and sample standard deviations. Energy integrates the full RUN current without baseline subtraction. At maximum clocks, ESP32 has the lowest observed mean latency for ten workloads, RP2040 for LZ77 and STM32F446 for DCT. At 160 MHz, the counts become six, one and five; minimum mean energy is distributed across six, two and four workloads. Relative to the maximum-clock profiles, ESP32’s mean energy falls by 1.29–3.54%, whereas RP2040’s rises by 3.35–4.00%. STM32’s mean energy is lower in eleven workloads but 15.38% higher for FIR. Equal nominal CPU frequency does not equalize internal supply settings, memory clocks or software execution. These results support the workload-specific comparison of documented operating configurations. The small cross-profile energy changes remain descriptive because capture SD characterizes repeatability, while absolute accuracy and cross-session systematic errors are not fully quantified.