NPU Playground

A hands-on exploration of the Neural Processing Unit on the Luckfox Pico Mini B (Rockchip RV1103) using the RKNN Toolkit: setting up the cross-development environment, converting trained models to RKNN format, and running accelerated inference on the device.

Heat-index prediction pipeline

The flagship experiment is an end-to-end pipeline that predicts heat index on the NPU. The model is trained with quantization-aware training (QAT) and deployed as a fully quantized INT8 network, then profiled for inference throughput and memory-access behaviour while maintaining regression accuracy — a study in how much useful ML fits on a sub-$10 board.