$8 Edge AI: running YOLOv8 on a 0.5-TOPS Rockchip RV1106 NPU
Quantization-aware training matters more than post-training quant. INT8 RKNN cuts size 3× with <2% accuracy loss. Offline-first saves the day when 4G drops.
Why RV1106
Deployed a farm-monitoring station 80km from town with 4G dropping 3-4× a day. Cloud-based goes blind during outages. Owners need pest detection within minutes.
Rockchip RV1106 = $8 chip, 0.5-TOPS NPU, 256MB RAM. Enough for YOLOv8-nano if compressed properly.
Optimal pipeline
- Train YOLOv8n on an 18k-image dataset of 7 pest classes (FP32, GPU)
- Quantization-aware training for the last 50 epochs with real calibration data from station cameras
- Export ONNX → convert to RKNN INT8 (rknn-toolkit2)
- Test on real hardware — not just simulator
Compression results
| Size | Acc mAP@0.5 | FPS | |
|---|---|---|---|
| FP32 PT | 6.1 MB | 0.882 | 1.8 |
| FP16 ONNX | 3.2 MB | 0.879 | 4.1 |
| INT8 PTQ | 1.8 MB | 0.812 | 12 |
| INT8 QAT | 2.1 MB | 0.864 | 12 |