Rockchip RKNN

RKNN-Toolkit2, RKNN-Toolkit-Lite2 and the C runtime API.

RKNN-Toolkit2 runs on an x86 Linux host and converts ONNX, TFLite, PyTorch and other models to .rknn, with INT8 quantisation. On the board you run models with RKNN-Toolkit-Lite2 (Python) or the C API (librknnrt). Older RV1109/RV1126 SDKs use the original RKNN-Toolkit (v1).

Key points

Quantise with a representative dataset (a dataset.txt listing 20–200 real images).
Match mean/std normalisation in rknn.config() to what the model was trained with.
Use accuracy_analysis() to find layers that lose precision after quantisation.

References

RKNN guides

AI & NPUIntermediate

Convert YOLO to RKNN and Run It on the RK3588 NPU

Export YOLO to ONNX, quantise and convert it to .rknn with RKNN-Toolkit2 on an x86 host, then run it on the board with RKNN-Toolkit-Lite2, using all three RK3588 NPU cores.

10 min read · RK3588, RK3588S, RK3576

Spotted an error or have something to add? Suggest an improvement.

Building a Rockchip-based product?

From hardware design to Linux, Android and AI integration, get engineering support for your Rockchip project.

Request Engineering Support See what we do