[data update] NPU development data based on 3568 development board is fully upgraded

With artificial intelligence , The era of big data is coming , Previously embedded processors CPU and GPU Gradually, it is difficult to meet the increasing demand , Especially in deep learning . In response to growing demand ,NPU Was born .

NPU The full English name is Neural Process Unit, Neural network processor .

To meet the needs of artificial intelligence , The processor of Ruixin micro is also gradually integrated NPU, be called RKNPU.

RKNPU After several generations of development , The first generation introduced RKNPU It's from RK3399pro and RK1808 At the beginning , Compared with traditional CPU and GPU, It has greatly improved the deep learning and operation ability . The next in RV1109 and RV1126 The second generation is used on NPU, Promoted NPU Utilization ratio . The third generation NPU Apply to RK3566 and RK3568 On , With brand new NPU Self research framework .

Ruixin micro RK3568 Chip built in NPU, yes RKNPU Representative products of the third generation . The third generation RKNPU The features are as follows ：

* Equipped with a new self-developed architecture

* Supports integers 8、 Integers 16 Convolution operation

* Built in NPU Support INT8/INT16 Mixing operation

* Inference tools support ：TensorFlow, Caffe, Tflite, Pytorch, Onnx NN, Android NN wait

* the height is 1 TOPS Neural network acceleration processing performance

be based on iTOP-RK3568 The new implementation of the development board NPU Self research framework , It can be applied to the following scenarios through algorithm blessing ：

at present ,iTOP-RK3568 The development board is selling well , be based on RK3568 Development board 《iTOP-3568 Development board NPU User manual 》 Also synchronize with the new .

Manual contents display

The first 1 Chapter Hello ！NPU

1.1 NPU The birth of ！

1.2 First time to know RKNPU

The first 2 Chapter Get ready RKNPU development environment

2.1 development environment

2.1 Software architecture

2.2 SDK explain

The first 3 Chapter Give Way NPU Run

3.1 stay Linux The system uses NPU

3.1.1 Set the cross compiler

3.1.2 Modify the compiler tool path

3.1.3 to update RKNN Model

3.1.4 compile demo

3.1.5 Development board running demo

3.2 stay Android The system uses NPU

3.2.1 Download the tools required for compilation

3.2.2 Modify the compiler tool path

3.2.3 to update RKNN Model

3.2.4 compile demo

3.2.5 Development board running demo

The first 4 Chapter Experience RKNN_DEMO

4.1 rknn_ssd Target detection test

4.2 rkn_api_demo test

4.3 Multimedia input test

The first 5 Chapter Model transformation

5.1 RKNN-Toolkit2 Introduce

5.2 install RKNN-Toolkit2

5.2.1 adopt docker Image mount

5.2.2 adopt pip install install

5.3 RKNN-Toolkit2 Use

5.3.1 The model runs in the simulator

5.3.2 The model runs on RK3568 Development board

The first 6 Chapter Other model transformations

6.1 Use tensorflow frame

6.2 Use caffe frame

6.3 Use tflite frame

6.4 Use onnx frame

6.5 Use darknet frame

6.6 Use pytorch frame

The first 7 Chapter Use RKNN-Toolkit-lite2

7.1 Main function description

7.2 Debian Change download source

7.3 install RKNN Toolkit Lite2

7.4 Run the test program