This research empirically examines embedded development tools viable for
on-device TinyML implementation. The research evaluates various development
tools with various abstraction levels on resource-constrained IoT devices, from
basic hardware manipulation to deployment of minimalistic ML training. The
analysis encompasses memory usage, energy consumption, and performance metrics
during model training and inference and usability of the different solutions.
Arduino Framework offers ease of implementation but with increased energy
consumption compared to the native option, while RIOT OS exhibits efficient
energy consumption despite higher memory utilization with equivalent ease of
use. The absence of certain critical functionalities like DVFS directly
integrated into the OS highlights limitations for fine hardware control.

本文经验性地研究了适用于设备上的 TinyML 实现的嵌入式开发工具。研究评估了不同抽象级别的开发工具在资源受限的物联网设备上的使用情况，从基本的硬件操作到最小化的机器学习训练部署。分析包括模型训练和推理过程中的内存使用、能源消耗和性能指标以及不同解决方案的可用性。Arduino 框架提供了易于实现的方式，但相比原生选项增加了能源消耗，而 RIOT OS 表现出高效的能源消耗，尽管内存利用率较高，其使用方便性与 Arduino 框架相当。操作系统中没有直接集成某些关键功能，比如 DVFS，凸显了对精细硬件控制的限制。