Performance Analysis of Machine Learning Algorithm on GD32 Microcontrollers

With the rapid development of 5G communication networks and edge computing (EC), more and more people prefer to deploy artificial intelligence (AI) algorithms on edge devices. The Train++ algorithm is one of the very few machine learning (ML) algorithms that can perform binary classification training and inference on MCUs. Due to the high performance and popularity of GD32 series MCUs, we conduct performance analysis of the Train++ algorithm using six datasets with different sample sizes and feature dimensions on two GD32 MCUs with different cores, also a STM32F103 MCU is considered as a comparison. Through our experiments, we discover that the average unit training time of the selected GD32F303 MCU on different datasets is 29.2% faster than that of STM32F103MCU, the average Flash occupancy rate is 24.4% lower than that of STM32F103MCU, and the average SRAM occupancy rate is 24.9% lower than that of STM32F103MCU. Therefore, the GD32F303 is more suitable for deployment in scenarios with certain performance requirements when compared with STM32 MCU. GD32E230 MCU can complete all training and inference processes at a price less than 40% of STM32F103. It is ideal for deployment in high-volume, high-density, and cost-constrained scenarios.