Improving Classifier Latency at the Edge through ARM Helium

Abstract

The increasing diffusion of intelligent devices at the network edge has led to a growing demand for efficient on-device inference, capable of overcoming the limitations of traditional cloud-centric computing paradigms. This work investigates the acceleration of decision tree–based inference on resource-constrained edge platforms by exploiting ARM Helium vector extensions , which bring the Single Instruction Multiple Data (SIMD) paradigm to the Cortex-M class of processors. A dedicated SIMD-based kernel was implemented and tested on the NUCLEO-STM32N657 board across three UCI datasets (AI4I, Dry Bean, Avila). Results show up to ∼ 15% latency reduction over the non-SIMD baseline , confirming that ARM Helium effectively exploits data-level parallelism to enhance inference efficiency on lightweight microcontrollers. Overall, this study provides experimental evidence that vector extensions represent a key enabler for bringing advanced machine learning capabilities to low-power embedded systems, bridging the gap between traditional micro-controller efficiency and modern AI acceleration at the edge.