Radxa Zero 3 NPU with Ubuntu 22.04

---

Installation.

• Get a 16 GB (minimal) SD card holding the image.
• Download the Radxa_Zero3_NPU_Ubuntu22.img.xz image (2.9 GByte) from our Sync site.
• Flash the image on the SD card with the Imager or balenaEtcher.
• Insert the SD card in your Rock 5 and enjoy.
• Username: radxa
• no password: radxa

Showstopper.

• The NPU only works on the 2 and 4 GB models.

---

Model performance benchmark(FPS)

All models, with C++ examples, can be found on the SD image.

demo	model_name	inputs_shape	dtype	Radxa Zero3
yolov5	yolov5s_relu	[1, 3, 640, 640]	INT8	14.8
	yolov5n	[1, 3, 640, 640]	INT8	19.5
	yolov5s	[1, 3, 640, 640]	INT8	11.7
	yolov5m	[1, 3, 640, 640]	INT8	5.7
yolov6	yolov6n	[1, 3, 640, 640]	INT8	18.0
	yolov6s	[1, 3, 640, 640]	INT8	8.1
	yolov6m	[1, 3, 640, 640]	INT8	4.5
yolov7	yolov7-tiny	[1, 3, 640, 640]	INT8	16.1
	yolov7	[1, 3, 640, 640]	INT8	3.4
yolov8	yolov8n	[1, 3, 640, 640]	INT8	18.2
	yolov8s	[1, 3, 640, 640]	INT8	8.9
	yolov8m	[1, 3, 640, 640]	INT8	4.4
yolox	yolox_s	[1, 3, 640, 640]	INT8	10.0
	yolox_m	[1, 3, 640, 640]	INT8	4.8
ppyoloe	ppyoloe_s	[1, 3, 640, 640]	INT8	9.2
	ppyoloe_m	[1, 3, 640, 640]	INT8	5.0
yolov5_seg	yolov5n-seg	[1, 3, 640, 640]	INT8	1.04
	yolov5s-seg	[1, 3, 640, 640]	INT8	0.87
	yolov5m-seg	[1, 3, 640, 640]	INT8	0.71
yolov8_seg	yolov8n-seg	[1, 3, 640, 640]	INT8	0.91
	yolov8s-seg	[1, 3, 640, 640]	INT8	0.87
	yolov8m-seg	[1, 3, 640, 640]	INT8	0.7
RetinaFace	RetinaFace_mobile320	[1, 3, 320, 320]	INT8	88.5
	RetinaFace_resnet50_320	[1, 3, 320, 320]	INT8	11.8
PPOCR-Det	ppocrv4_det	[1, 3, 480, 480]	INT8	15.1
PPOCR-Rec	ppocrv4_rec	[1, 3, 48, 320]	FP16	17.3

• Due to the pixel-wise filling and drawing, segmentation models are relatively slow

---

Cooling.

You must cool your Zero3. It will get very hot without a heatsink.
We used a heatsink with two fans designed for the Raspberry Pi Zero, and it works fine.
Even with the NPU running 24/7 at 1.8 GHz, it never gets warmer than 42°C (107°F).

Use the thermal pad properly.
It should fill the space between the chip and the cooling element effectively.
If there is any gap, the heat flow will not be optimal, resulting in a much hotter CPU.
The delivered pad will come with two plastic protective sheets. These must be removed before applying the pad.
If there is still a small gap (the CPU of the Radxa is slightly thinner than the Raspberry Pi), cut some small slices from the pad and stack them.
The actual CPU core is located at the centre of the chip, where the heat is generated.

---

Tips.

• If you need extra space delete the opencv and the opencv_contrib folder from the SD card. They are no longer needed since all libraries are stored in the /usr/ directory.
• Use a tool like GParted sudo apt-get install gparted to expand the image to larger SD cards. We recommend a minimum of 64 GB. Deep learning requires a lot of space.
  

---

Pre-installed frameworks.

• OpenCV 4.10.0
• NPU rknpu2 1.5.2
• NPU model zoo 2.0.0
• NPU model zoo models 2.0.0

---

Thanks.

A more than special thanks to Joshua Riek for all the hard work on the Ubuntu OS.

---