RK182X Series NPU Overview (RK1828 Model)

This chapter covers the NPU architecture of the RK182X series (focusing on the RK1828 model), supported precisions, and the RKNN3 development flow with conversion / inference code.

The RK182X series uses the RKNN3 Toolkit for AI model development. All example code in this document is based on RKNN3; RKNN-Toolkit2 is for host SoCs such as the RK3588.

Overall block diagram

PC side                                   Device side
┌──────────────────────┐                 ┌──────────────────────┐
│  RKNN3-Toolkit (whl)  │   Transfer model │  RKNN3-Toolkit-Lite   │
│  1. export_llm.py     │  ────────────►  │  4. rkllm3-server    │
│  2. export_rknn.py    │                 │  5. rknn3_*_demo     │
│  3. .rknn/.weight/etc │                 │  6. Inference output │
└──────────────────────┘                 └──────────────────────┘

1. Supported Precisions

Precision modes supported by the RK182X NPU (based on RKNN3 Release Note V1.0.0):

Precision	Description
INT8	Peak compute 20 TOPS (confirmed by official datasheet)
FP16	High-precision inference / intermediate format before quantization
BF16	Brain Float 16 (a separate format from FP16)

2. RKNN3 Development Flow

Step	Where	Tool
1. Export ONNX	PC	`export_llm.py`
2. Convert to RKNN	PC	`rknn-toolkit3` (`from rknn.api import RKNN`)
3. Transfer model to device	PC → board	scp / adb push
4. Start inference service	Board	`rkllm3-server` / `rknn3_*_demo`
5. Application calls	Board	OpenAI-compatible API or Python API

3. Model Conversion Example (PC Side)

from rknn.api import RKNN

rknn = RKNN(verbose=True)

rknn.config(target_platform='rk1820',          # ← not 'rk1828'
            quantized_dtype='w4a16',
            quantized_algorithm='grq',
            quantized_method='group32')

rknn.load_llm(model='Qwen3-1.7B.onnx', config='Qwen3-1.7B.config.pkl')
rknn.build(do_quantization=True, dataset='datasets/CMMLU/dataset.txt')
rknn.export_rknn('Qwen3-1.7B.rknn')

4. Runtime Inference Example (C API, board side)

#include "rknn3_api.h"   // request the header file from Rockchip

rknn3_init(&ctx, &ext);   // context pointer + init_extend

rknn3_load_model_from_path(ctx, "model.rknn", "model.weight");
// Only 3 arguments; core_mask is not passed here

rknn3_config cfg = {0};
cfg.run_core_mask = 0xff;   // LLM/VLM use all 8 cores (0xff); some CNN models use 0x01
rknn3_model_init(ctx, &cfg);

rknn3_run(ctx, inputs, n_inputs, outputs, n_outputs);   // CNN inference
// LLM inference goes through rknn3_session_run(session, ...) (LLM-only session API)

rknn3_destroy(ctx);

5. NPU-Supported Operators

Based on RKNN3 Release Note V1.0.0:

Optimized implementations of operators such as exSDPA, exMatMul, Resize, and Transpose
For the complete operator list, refer to the official RKNN3 Toolkit documentation

6. Performance Data (RKNN3 V1.0.0)

Category	Capability
LLM Decode performance	Overall improvement >15%
Multi-core multi-model	Simultaneous inference supported (PCIe card has 8 cores; the model declares the required core_mask at compile time — `0xff` for LLM/VLM, `0x01` for some CNNs)
New formats	Supports mRoPE, Function Call, and YUV-format input
rkllm3 server	Supports embedding models and audio input

7. Verifying NPU Status

systemctl is-active rknn3.service && lspci -d 1d87:182a

Output:

active
0004:41:00.0 Processing accelerators: Rockchip Electronics Co., Ltd Device 182a (rev 01)

8. SDK Reference Documents

PDF documents:

ls /userdata/RK1820_RK1828_AI_SDK/docs/Release/

Output:

Rockchip_RK1820_RK1828_AI_SDK_Quick_Start_CN.pdf
Rockchip_RK1820_RK1828_AI_SDK_Quick_Start_EN.pdf
Rockchip_RK1820_RK1828_AI_SDK_Quick_Start_For_Android_CN.pdf
Rockchip_RK1820_RK1828_AI_SDK_Quick_Start_For_Android_EN.pdf
Rockchip_RK1820_RK1828_AI_SDK_RELEASE_CN.pdf
Rockchip_RK1820_RK1828_AI_SDK_RELEASE_EN.pdf

Markdown documents:

ls -la /userdata/RK1820_RK1828_AI_SDK/docs/RK1820_RK1828_AI_Release-Note_CN.md /userdata/RK1820_RK1828_AI_SDK/docs/RK1820_RK1828_AI_Release-Note_EN.md

Output:

-rw-r--r-- 1 linaro linaro 3905 2026年 1月28日 RK1820_RK1828_AI_Release-Note_CN.md
-rw-r--r-- 1 linaro linaro 4947 2026年 1月28日 RK1820_RK1828_AI_Release-Note_EN.md

Release Note content (excerpt):

cat /userdata/RK1820_RK1828_AI_SDK/docs/RK1820_RK1828_AI_Release-Note_CN.md | head -20

Output:

# 更新日志

## Release V1.0.0

1. Examples 应用示例
 - 新增 Docker 环境部署说明及示例，提升示例运行与复现效率
 - rockit_yolo_detect 示例兼容 RKNN2 与 RKNN3 接口
 - 新增工业异常检测、vi_yolo_venc 示例，并补充对应应用文档说明

2. 工具集更新
- RKNN-Console 新增多卡与 USB 设备支持，修复异常场景下的稳定性问题
- RKNN-SMI 新增 NPU 负载、电流、电压、功耗等查询接口，并优化 CLI 与日志机制
- 新增 rknn-test 自动化测试脚本，提升功能验证与稳定性验证效率

3. RKNN3更新
- 版本更新至 V1.0.0，覆盖 RKNN3 Toolkit、RKNN3 Runtime、RKNN3 Model Zoo
- 大幅提升 LLM/ViT 性能，LLM Decode 性能整体提升超过15%

Documentation directory structure:

ls /userdata/RK1820_RK1828_AI_SDK/docs/

Output:

Examples
Others
Release
Tools
RK1820_RK1828_AI_Release-Note_CN.md
RK1820_RK1828_AI_Release-Note_EN.md

9. FAQ

Symptom	Cause	Resolution
`rknn.config` reports unsupported `target_platform`	Written as `'rk1828'`	Change to `'rk1820'`
`rknn3_init` header file missing	`rknn3_api.h` not installed	Request it from Rockchip
`core_mask X is not match with npu core number N!`	`-c` mismatched with the model's core count	Use `0xff` for LLM, `0x01` for CNN
182a not visible in `lspci`	Card not seated / insufficient power	Re-seat the card + check dmesg
`rknn3.service` not active	systemd service failure	`sudo systemctl restart rknn3.service`

10. Next Steps

RKNN3 SDK Overview — SDK overview
RKNN3 Toolkit Installation — Toolkit installation
INT8 Quantized Inference — quantization details