Dear TVM Community,
We recently released a newly developed AIoT toolkit U-TOE and thrilled to invite you to try out. As specialist in the field of ML system, your insights are invaluable to us.
U-TOE is an open source toolkit that seamlessly combines the generic model compiler TVM with the open-source IoT operating system RIOT. It allows for evaluating resource consumption of model inference locally or remotely on various low-power boards like Raspberry Pi Pico, nrf52840dk, and Arduino Nano. It supports MCU architectures such as ARM Cortex-M and RISC-V.
We attached a preprint of our paper entitled On-Device Evaluation Toolkit for Machine Learning on Heterogeneous Low-Power System-on-Chip. It provides an overview of U-TOE’s design, measurement principle, and benchmarks of representative ML models.
We kindly invite you to explore U-TOE and test it on your preferred IoT devices. Your feedback on usability, performance, and compatibility with different ML frameworks and IoT boards would greatly contribute to ongoing development efforts.
If you have share any issues while trying out U-TOE, please do not hesitate to contact me. Your input will bring a bright future of this toolkit.
Thank you for considering our invitation. We truly appreciate your time and expertise.
I am a PhD student at Freie Universität Berlin in the field of TinyML/Machine Learning systems, advised by Prof. Dr. Emmanuel Baccelli (a co-founder of RIOT). We aim to explore not only inference on device, but also the capabilities of on-device learning/training. A more general perspective could be to integrate TVM support into mainline RIOT to empower the wider RIOT open source community with low-power AI programmability.
Looking forward to our journey!