RK182X Series CNN Inference (RK1828 Model)

Overall block diagram

Host (RK3588)                            RK1828 (PCIe AI coprocessor)
──────────────────                       ─────────────────────
test.jpg (224×224)                       8-core NPU
   │                                          │
   ▼                                          │
rknn3_cnn_demo -m -w -i -c 1 -cl 10 -mm     │
   │                                          │
   ▼                                          │
PCIe Gen3 x4 ──────────────────────────────►  NPU inference (1 core)

-c (core_mask) must match the core count declared at model compile time, not the physical core count of the NPU: the NPU actually has 8 cores (RK1828 PCIe card, 1d87:182a), MobileNet V2 declares 1 core at compile time (use -c 1), and Qwen3-1.7B declares 8 cores (use -c 0xff).
The preinstalled mobilenet_v2 is FP16-only (for INT8, convert it yourself; see ch05 Model Conversion).

1. Prerequisites

# 1. RK182X NPU service healthy
systemctl is-active rknn3.service    # → active
ls -l /dev/pcie-rkep-*              # → /dev/pcie-rkep-0004:41:00.0

# 2. An RKNN model of MobileNet V2 (see ch05)

Verification:

systemctl is-active rknn3.service

Output:

active

ls -l /dev/pcie-rkep-*

Output:

crw------- 1 root root 10, 124  8月20日 16:56 /dev/pcie-rkep-0004:41:00.0

2. Deployed Models (FP16)

/userdata/RK1820_RK1828_AI_SDK/examples/gstreamer/MobilenetV2/
├── mobilenetv2-12.rknn     60,352 B    # FP16 model structure
├── mobilenetv2-12.weight   7,037,440 B # FP16 model weights
├── labels.txt              21 KB       # ImageNet 1000-class labels
└── config.json                          # Model configuration

Verifying the files:

ls -la /userdata/RK1820_RK1828_AI_SDK/examples/gstreamer/MobilenetV2/

Output:

-rw-r--r-- 1 linaro linaro     218  2026年 1月28日 config.json
-rw-r--r-- 1 linaro linaro   21675  2026年 1月28日 labels.txt
-rw-r--r-- 1 linaro linaro   60352  2026年 1月28日 mobilenetv2-12.rknn
-rw-r--r-- 1 linaro linaro 7037440  2026年 1月28日 mobilenetv2-12.weight

File sizes:

du -h /userdata/RK1820_RK1828_AI_SDK/examples/gstreamer/MobilenetV2/*.{rknn,weight}

Output:

60K	.../MobilenetV2/mobilenetv2-12.rknn
6.8M	.../MobilenetV2/mobilenetv2-12.weight

3. Inference Test (FP16, 1-core model with -c 1)

rknn3_cnn_demo \
  -m /userdata/RK1820_RK1828_AI_SDK/examples/gstreamer/MobilenetV2/mobilenetv2-12.rknn \
  -w /userdata/RK1820_RK1828_AI_SDK/examples/gstreamer/MobilenetV2/mobilenetv2-12.weight \
  -i /rockchip-test/npu2/model/cat_224x224.jpg \
  -c 1 \
  -cl 10 \
  -mm

The -c value must match the core count declared at model compile time: MobileNet V2 declares 1 core → -c 1 (verified working). A mismatch reports core_mask X is not match with npu core number N! (N is the model's declared core count).

Tool help:

rknn3_cnn_demo --help

Output:

Usage:
rknn3_cnn_demo [options]

Options:
  -m, --model_path  .rknn model path
  -w, --weight_path .weight model path
  -c, --core_mask  core mask
  -i, --image_path  core mask
  -cl, --cnt_loop  default:1
  -tl, --time_loop(hours)
  -id, --device_id
  --yuv  input image is yuv420sp nv12
  -mm, --enable_memory_monitor  enable memory monitoring (default: disabled)

Test image:

ls -la /rockchip-test/npu2/model/cat_224x224.jpg

Output:

-rwxr-xr-x 1 root root 58701  5月26日 09:09 /rockchip-test/npu2/model/cat_224x224.jpg

4. Parameter Reference

Parameter	Long option	Meaning
`-m`	`--model_path`	RKNN model path
`-w`	`--weight_path`	Weight path (mandatory for the API)
`-i`	`--image_path`	Input image (224×224 JPG)
`-c`	`--core_mask`	NPU core mask; must match the cores declared at compile time
`-cl`	`--cnt_loop`	Loop count
`-id`	`--device_id`	Device ID
`-mm`	`--enable_memory_monitor`	Enable memory monitoring (requires root)

5. Performance (1-core model with -c 1, average of 10 runs)

RunCount	TotalTime	AvgFPS	AvgCost	AvgMemoryUsage	AvgNpuUtilization
10	45.44 ms	220.064	4.54 ms	unavailable	unavailable

Precision	Single inference	FPS	Model size
FP16	4.54 ms	220	60KB + 6.8MB
INT8	—	—	—

-mm requires root (non-root users get Permission denied); inference itself does not need root:
sudo rknn3_cnn_demo ... -mm

6. Comparison: `-c` for CNN vs LLM

Model type	Cores declared by model	Recommended `-c`
CNN (MobileNet V2)	1	`1` / `0x01`
LLM (Qwen3-1.7B)	8	`0xff`

7. Obtaining Other Precision Versions

# Download the ONNX source model
cd /userdata/RK1820_RK1828_AI_SDK/rknn/rknn3-model-zoo/examples/mobilenet_v2/model
bash download_model.sh
# → downloads mobilenetv2-12.onnx

# Convert to INT8 RKNN with rknn3-toolkit (see ch05)

8. FAQ

Symptom	Cause	Fix
`core_mask X is not match with npu core number N!`	`-c` mismatched with model cores	MobileNetV2 uses `-c 1`, LLM uses `-c 0xff`
`Permission denied` (with -mm)	Non-root user lacks permission	Add `sudo`
`rknn_load_model failed! ret=-2`	Model architecture mismatch	Confirm it is the RK1828 version (not RK356X/RK3588)
`weight_path is null or empty!`	Missing `-w` parameter	Add `-w <path>.weight`

9. Next Steps

Model Conversion — convert to INT8 / other precisions
LLM Inference — 8-core models (-c 0xff)
NPU Overview — NPU hardware and drivers