diff --git a/docs/hardware/en/lichee/th1520/lpi4a/3_images.md b/docs/hardware/en/lichee/th1520/lpi4a/3_images.md
index eb9ce9af0e..aabb96e725 100644
--- a/docs/hardware/en/lichee/th1520/lpi4a/3_images.md
+++ b/docs/hardware/en/lichee/th1520/lpi4a/3_images.md
@@ -39,12 +39,12 @@ The relevant files are in the `20230803_tempfix.zip` compressed package.
 (images of 0721 and later versions can be used normally, no need to replace it with the file here)
 
 ```shell
-sudo ./fastboot flash ram ./images/u-boot-with-spl-ddr16g.bin
+sudo ./fastboot flash ram ./images/u-boot-with-spl-lpi4a-16g.bin
 sudo ./fastboot reboot
 sleep 1
-sudo ./fastboot flash uboot ./images/u-boot-with-spl-ddr16g.bin
+sudo ./fastboot flash ram ./images/u-boot-with-spl-lpi4a-16g.bin
 # If there is no device tree corresponding to 16G ddr in the boot.ext4 you use, you need to burn the boot.ext4 corresponding to 16G ddr
-sudo ./fastboot flash boot ./images/boot_16gddr.ext4
+sudo ./fastboot flash boot ./images/boot.ext4
 ```
 
 ### Debian
@@ -57,9 +57,9 @@ Download Links:
 Mega Cloud Storage：[click me](https://mega.nz/folder/phoQlBTZ#cZeQ3qZ__pDvP94PT3_bGA)
 ISCAS mirror: [click me](https://mirror.iscas.ac.cn/revyos/extra/images/lpi4a/)
 
-1. LPI4A_20230721.zip
+1. LPI4A_20230912.zip
 
-    - Release date: July 21, 2023
+    - Release date: September 12, 2023
     - Kernel: 5.10
     - Root file system: ext4, 4.3GB
     - Main pre-installed packages:
@@ -76,11 +76,11 @@ ISCAS mirror: [click me](https://mirror.iscas.ac.cn/revyos/extra/images/lpi4a/)
        - There is a problem with the sound quality of HDMI audio playing under chromium
     - Instructions for use:
        - This zip file only supports HDMI display, when burning, please note that the suffix in the file name corresponds to the relevant parameters of the device.
-         - boot, 8gddr/16gddr in the u-boot file name corresponds to 8g/16g memory respectively.
+         - u-boot with 16g suffix is u-boot that supports 16g memory, without suffix is u-boot that supports 8g memory.
 
-2. LPI4A_20230721_mipi.zip
+2. LPI4A_20230912_MIPI.zip
 
-    - Release date: July 21, 2023
+    - Release date: September 12, 2023
     - Kernel: 5.10
     - Root file system: ext4, 4.3GB
     - Main pre-installed packages:
@@ -102,8 +102,7 @@ ISCAS mirror: [click me](https://mirror.iscas.ac.cn/revyos/extra/images/lpi4a/)
        - There is a problem with the sound quality of HDMI audio playing under chromium
     - Instructions for use:
       - This zip file is the image of the MIPI screen, please note that the suffix in the file name corresponds to the relevant parameters of the device when burning.
-         - 8gddr/16gddr in the boot and u-boot file names correspond to 8g/16g memory respectively.
-         - mipi_720p/mipi_1080p in the boot file name corresponds to 800x1280 MIPI screen (with touch)/1200x1920 mipi screen respectively.
+         - u-boot with 16g suffix is u-boot that supports 16g memory, without suffix is u-boot that supports 8g memory.
 
 <!-- 1. LPi4A_Test_0425
    1. Release date: April 25, 2023
diff --git a/docs/hardware/en/lichee/th1520/lpi4a/4_burn_image.md b/docs/hardware/en/lichee/th1520/lpi4a/4_burn_image.md
index 6d7d5d4b67..d214625012 100644
--- a/docs/hardware/en/lichee/th1520/lpi4a/4_burn_image.md
+++ b/docs/hardware/en/lichee/th1520/lpi4a/4_burn_image.md
@@ -107,20 +107,20 @@ After putting the board into burning mode, you can use fastboot from `burn_tool.
 Let´s take linux as an example:
 Note that you need to mark the fastboot binary as executable first via `chmod +x fastboot`
 
-`u-boot-with-spl-ddr8g.bin` and `u-boot-with-spl-ddr16g.bin` are u-boot firmware, please refer to the image description for specific differences.
+`u-boot-with-spl-lpi4a.bin` and `u-boot-with-spl-lpi4a-16g.bin` are u-boot firmware, please refer to the image description for specific differences.
 ```bash
-sudo ./fastboot flash ram ./images/u-boot-with-spl-ddr16g.bin
+sudo ./fastboot flash ram ./images/u-boot-with-spl-lpi4a-16g.bin
 sudo ./fastboot reboot
-sleep 10
-sudo ./fastboot flash uboot ./images/u-boot-with-spl-ddr16g.bin
-sudo ./fastboot flash boot ./images/boot_16gddr.ext4
-sudo ./fastboot flash root ./images/rootfs-thead-image-linux_sing.ext4.ext4
+sleep 1
+sudo ./fastboot flash uboot ./images/u-boot-with-spl-lpi4a-16g.bin
+sudo ./fastboot flash boot ./images/boot_sing.ext4
+sudo ./fastboot flash root ./images/rootfs-sing.ext4
 ```
 
 The first three lines will check and create the partitions on the flash. If you skip this step,
 burning the rootfs will be very slow later on.
 
-`boot_8gddr.ext4`, `boot_8gddr_mipi_720p.ext4`, `boot_8gddr_mipi_1080p.ext4`, `boot_16gddr.ext4`, `boot_16gddr_mipi_720p.ext4` and `boot_16gddr_mipi_1080p.ext4` for boot For specific differences, please refer to the image description. They mainly include the following:
+`boot-sing.ext4` and `boot-dual.ext4` for boot For specific differences, please refer to the image description. They mainly include the following:
 ```bash
 fw_dynamic.bin        #opensbi
 Image                 #kernel image
@@ -131,7 +131,7 @@ light-lpi4a.dtb       # ddr8G dtb
 light-lpi4a-ddr16g.dtb # ddr16G dtb
 ```
 
-`rootfs-thead-image-linux_sing.ext4` and `rootfs-thead-image-linux_mipi.ext4` are the root file system, and the default is Debian system. For the specific differences between the two, please refer to the image description.
+`rootfs-sing.ext4` and `rootfs-mipi.ext4` are the root file system, and the default is Debian system. For the specific differences between the two, please refer to the image description.
 
 Log output you typically see while burning an image:
 
diff --git a/docs/hardware/en/lichee/th1520/lpi4a/6_peripheral.md b/docs/hardware/en/lichee/th1520/lpi4a/6_peripheral.md
index 4e84c90ed0..7514cd7467 100644
--- a/docs/hardware/en/lichee/th1520/lpi4a/6_peripheral.md
+++ b/docs/hardware/en/lichee/th1520/lpi4a/6_peripheral.md
@@ -2,6 +2,11 @@
 title: Peripheral Use
 keywords: Linux, Lichee, TH1520, SBC, RISCV, Peripheral
 update:
+  - date: 2023-09-12
+    version: v1.3
+    author: ztd
+    content:
+      - Update NPU user guide
   - date: 2023-08-18
     version: v1.2
     author: ztd
@@ -865,7 +870,68 @@ When playing video with Chromium browser, the state of GPU is as follows:
 
 ## NPU
 
-TODO
+> **Note**: To use the NPU driver, you need to upgrade to the [20230912](https://mega.nz/folder/phoQlBTZ#cZeQ3qZ__pDvP94PT3_bGA) version image.
+
+NPU-related driver initialization commands 
+```shell
+sudo npu_init
+```
+LicheePi4A contains a 1GHz NPU supporting 4TOPS@INT8 general purpose NNA computing power. The Wiki contains examples of NPU support as follows:
+ 
+|Example Name|Example Functionality|Usage Model|Interface|HHB Version| 
+|---|---|---|---|---|
+|Mobilenetv2 do image classification|image classification|mobilenetv2|c/c++|2.4 and above|
+|YOLOv5 do object detection|object detection|YOLOv5|python|2.4 and above|
+|RTMPose do pose estimation|pose estimation|RTMPose|python|2.6 and above|
+
+In order to cross-compile the models in the above examples into executables on LicheePi4A, we first need to set up the HHB development environment on our own computer.
+> Recommended environment: Ubuntu 20.04 system, Docker version 20.10.21.It is recommended to use the Docker image to set up the environment.
+
+### Installation
+
+First, we need to install Docker on our own computer. Uninstall any existing Docker version: 
+```shell
+sudo apt-get remove docker docker-engine docker.io containerd runc
+```
+
+Install the basic software Docker depends on: 
+```shell
+sudo apt-get update
+sudo apt-get install apt-transport-https ca-certificates curl gnupg-agent software-properties-common
+Add the official source: ```shell
+curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -
+sudo add-apt-repository "deb [arch=amd64] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable"
+```
+
+Install Docker:
+```shell
+sudo apt-get update
+sudo apt-get install docker-ce docker-ce-cli containerd.io
+```
+After installation, get the Docker image of the HHB environment:
+```shell
+docker pull hhb4tools/hhb:2.4.5
+```
+
+Once the image is pulled, enter the Docker image using the following command:
+```shell
+docker run -itd --name=your.hhb2.4 -p 22 "hhb4tools/hhb:2.4.5"
+docker exec -it your.hhb2.4 /bin/bash
+```
+
+Once you are inside the Docker image, you can use the following command to verify the HHB version: 
+```shell
+hhb --version
+```
+
+Once you are in the Docker image, you need to configure the cross-compile environment. Note that you must use the toolchain here, otherwise the compiled binaries will not run on LicheePi4A. 
+```shell
+export PATH=/tools/Xuantie-900-gcc-linux-5.10.4-glibc-x86_64-V2.6.1-light.1/bin/:$PATH
+```
+
+At this point, the HHB environment is preliminarily built. You can try the following NPU examples:
+[Mobilenetv2 for image classification](https://wiki.sipeed.com/hardware/eh/lichee/th1520/lpi4a/8_application.html#MobilenertV2)
+[YOLOv5 for object detection](https://wiki.sipeed.com/hardware/eh/lichee/th1520/lpi4a/8_application.html#Yolov5n)
 
 ## Other
 Contributions are welcome~ You can get ￥5~150 ($1~20) coupon if your contribution is accepted!
\ No newline at end of file
diff --git a/docs/hardware/en/lichee/th1520/lpi4a/8_application.md b/docs/hardware/en/lichee/th1520/lpi4a/8_application.md
index ba2d60ec1f..6a1707ac2c 100644
--- a/docs/hardware/en/lichee/th1520/lpi4a/8_application.md
+++ b/docs/hardware/en/lichee/th1520/lpi4a/8_application.md
@@ -2,6 +2,11 @@
 title: Typical Application
 keywords: Linux, Lichee, TH1520, SBC, RISCV, application
 update:
+  - date: 2023-09-12
+    version: v1.2
+    author: ztd
+    content:
+      - Add some NPU applications
   - date: 2023-07-21
     version: v1.1
     author: ztd
@@ -30,7 +35,7 @@ The feasibility of running the 7B model on an entry-level C906 core was also bri
 
 ## YOLOX Target Detection
 
-This tutorial is an example of how to deploy the [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) model to accomplish target detection on the LPi4A (LicheePi 4A) development board platform.
+This tutorial is an example of how to deploy the [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) model to accomplish target detection on the LPi4A (LicheePi 4A) development board platform(CPU inference).
 Included in the tutorial:
 - Installing the Python environment on the LPi4A development board
 - Executing the model using the source code from the YOLOX project.
@@ -184,6 +189,540 @@ After the example executes normally, the result image soccer.jpg will be generat
 
 ![yolox_detection_soccer_output.jpg](./../../../../zh/lichee/th1520/lpi4a/assets/application/yolox_detection_soccer_output.jpg)
 
+## MobilenertV2 
+
+This tutorial is an example of how to deploy the mobilenetv2 model for image classification on the LicheePi4A platform.
+ The tutorial includes:
+ - Compiling the onnx model into a binary available on LicheePi4A using HHB
+ - Preprocessing the mobilenetv2 model using opencv c++ version on LicheePi4A
+ - Differences between using CPU and NPU on LicheePi4A
+
+### NPU
+
+#### Setup
+
+After setting up the NPU environment according to the [peripheral](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/6_peripheral.html#NPU) documentation, we can go to the Docker image of HHB environment.
+
+ First, we need to get the model of this tutorial, which is downloaded to the example directory `/home/example/th1520_npu/onnx_mobilenetv2_c++`: [mobilenetv2-12.onnx](https://github.com/onnx/models/blob/main/vision/classification/mobilenet/model/mobilenetv2-12.onnx)
+
+To get the libraries for the optimized version of opencv used in this tutorial, go to [github](https://xuantie.t-head.cn/community/download?id=4112956065753141248) and download them to the directory `/home/example/th1520_npu/`.
+```shell
+cd /home/example/th1520_npu/
+git clone https://github.com/zhangwm-pt/prebuilt_opencv.git
+```
+
+#### Compile
+
+**HHB Compiling Model:**
+To cross-compile the ONNX model into an executable program on the NPU, you need to use the hhb command. Note that the NPU only supports 8-bit or 16-bit fixed point operations, which in this case are int8 asymmetric quantization. To compile, you need to go to the directory `/home/example/th1520_npu/onnx_mobilenetv2_c++`:
+```shell
+cd /home/example/th1520_npu/onnx_mobilenetv2_c++
+hhb -D --model-file mobilenetv2-12.onnx --data-scale 0.017 --data-mean "124 117 104"  --board th1520  --postprocess save_and_top5 --input-name "input" --output-name "output" --input-shape "1 3 224 224" --calibrate-dataset persian_cat.jpg  --quantization-scheme "int8_asym"
+```
+
+The options are:
+- -D: Specifies that the HHB process will run until the executable is generated
+- --model-file: Specifies the mobilenet model that has been downloaded in the current directory
+- --data-mean: Specifies the mean value
+- --data-scale: Specifies the scale value
+- --board: Specifies that the target platform is th1520
+- --input-name: The input name of the model
+- --output-name: The output name of the model
+- --input-shape: The input size of the model
+- --postprocess: Saves the output results and prints the top5 results
+- --calibrate-dataset: Specifies the calibration images required for quantization
+- --quantization-scheme: Specifies the quantization scheme as int8 asymmetric
+
+After the command is executed, the hhb_out subdirectory will be generated in the current directory, which includes hhb_runtime, model.c and other files:
+- hhb. bm: The HHB model file, including the quantized weight data and other information
+- hhb_runtime: The executable on the th1520 platform, compiled from the c file in the directory
+- main. c: A temporary file, the reference entry of the sample program
+- model. c: A temporary file, the model structure file, related to the model structure
+- model.params: A temporary file, the weight values
+- io. c: A temporary file, the auxiliary functions for reading and writing files
+- io. h: A temporary file, the auxiliary functions for reading and writing files declaration
+- process. c: A temporary file, the image preprocessing function
+- process. h: A temporary file, the image preprocessing function declaration
+
+For more detailed HHB options, refer to the command line options in the [HHBUserManual](https://www.yuque.com/za4k4z/kvkcoh/shl1nybhel6iviwt).
+
+**g++ Compiling**
+```shell
+riscv64-unknown-linux-gnu-g++ main.cpp -I../prebuilt_opencv/include/opencv4 -L../prebuilt_opencv/lib   -lopencv_imgproc   -lopencv_imgcodecs -L../prebuilt_opencv/lib/opencv4/3rdparty/ -llibjpeg-turbo -llibwebp -llibpng -llibtiff -llibopenjp2    -lopencv_core -ldl  -lpthread -lrt -lzlib -lcsi_cv -latomic -static -o mobilenetv2_example
+```
+
+The mobilenetv2_example file is generated in the example directory after the compilation command is executed correctly.
+
+#### Execute
+
+After the cross-compilation is complete, you can copy the files needed for program execution to the directory of the development board. The scp command can be used:
+```shell
+scp -r ../onnx_mobilenetv2_c++ sipeed@your_ip:~
+```
+
+First confirm whether the driver in the development board is loaded:
+```shell
+lsmod
+```
+If there are 'img_mem', 'vha' and 'vha_info' in the output of the three modules, the NPU driver is loaded successfully, if not, you can run the following command to manually load it:
+```shell
+sudo npu_init
+```
+
+Refer to [YOLOX](https://wiki.sipeed.com/hardware/en/lichee/th1520/lpi4a/8_application.html#YOLOX-Target-Detection) Install and configure the python virtual environment:
+```shell
+sudo apt update
+sudo apt install wget git vim
+
+wget https://github.com/T-head-Semi/csi-nn2/releases/download/v2.4-beta.1/c920.tar.gz
+tar xf c920.tar.gz
+cp c920/lib/* /usr/lib/riscv64-linux-gnu/ -rf
+
+sudo apt install python3-pip
+sudo apt install python3.11-venv
+
+cd /root
+python3 -m venv ort
+source /root/ort/bin/activate
+
+git clone -b python3.11 https://github.com/zhangwm-pt/prebuilt_whl.git
+cd prebuilt_whl
+
+pip install numpy-1.25.0-cp311-cp311-linux_riscv64.whl
+pip install opencv_python-4.5.4+4cd224d-cp311-cp311-linux_riscv64.whl
+pip install kiwisolver-1.4.4-cp311-cp311-linux_riscv64.whl
+pip install Pillow-9.5.0-cp311-cp311-linux_riscv64.whl
+pip install matplotlib-3.7.2.dev0+gb3bd929cf0.d20230630-cp311-cp311-linux_riscv64.whl
+pip install pycocotools-2.0.6-cp311-cp311-linux_riscv64.whl
+pip3 install loguru-0.7.0-py3-none-any.whl
+pip3 install torch-2.0.0a0+gitc263bd4-cp311-cp311-linux_riscv64.whl
+pip3 install MarkupSafe-2.1.3-cp311-cp311-linux_riscv64.whl
+pip3 install torchvision-0.15.1a0-cp311-cp311-linux_riscv64.whl
+pip3 install psutil-5.9.5-cp311-abi3-linux_riscv64.whl
+pip3 install tqdm-4.65.0-py3-none-any.whl
+pip3 install tabulate-0.9.0-py3-none-any.whl
+```
+
+Run the just-compiled example in the appropriate directory on the development board:
+```shell
+./mobilenetv2_example
+```
+
+After the execution, the terminal will prompt the stages of the execution:
+1. Preprocessing
+2. Model execution
+3. Post-processing
+
+The files used in the execution of mobilenetv2_example are:
+- persian_cat.jpg: the input image
+- input_img.bin: the intermediate results generated by the input image during the preprocessing stage
+- hhb_out/hhb_runtime: the file used in the model execution stage, generated by HHB on the x86 host
+- hhb_out/hhb. bm: the file used in the model execution stage, generated by HHB on the x86 host
+- input_img.bin_output0_1_1000.txt: the output file of the model execution stage, including 1000 results of the model execution output
+
+#### Results
+
+```shell
+(ort) root@lpi4a:/home/sipeed/onnx_mobilenetv2_c++# ./mobilenetv2_example
+ ********** preprocess image **********
+ ********** run mobilenetv2 **********
+INFO: NNA clock:792000 [kHz]
+INFO: Heap :ocm (0x18)
+INFO: Heap :anonymous (0x2)
+INFO: Heap :dmabuf (0x2)
+INFO: Heap :unified (0x5)
+Run graph execution time: 7.87149ms, FPS=127.04
+
+=== tensor info ===
+shape: 1 3 224 224 
+data pointer: 0x857ca0
+
+=== tensor info ===
+shape: 1 1000 
+data pointer: 0x3fc9abe000
+The max_value of output: 16.053827
+The min_value of output: -8.026914
+The mean_value of output: -0.001889
+The std_value of output: 9.203342
+ ============ top5: ===========
+283: 16.053827
+281: 14.165141
+287: 11.709850
+285: 11.615416
+282: 11.332113
+ ********** postprocess result **********
+ ********** probability top5: ********** 
+n02123394 Persian cat
+n02123045 tabby, tabby cat
+n02127052 lynx, catamount
+n02124075 Egyptian cat
+n02123159 tiger cat
+```
+
+### CPU
+Replace the HHB compile command in the NPU step above with:
+```shell
+hhb -D --model-file mobilenetv2-12.onnx --data-scale 0.017 --data-mean "124 117 104"  --board c920  --postprocess save_and_top5 --input-name "input" --output-name "output" --input-shape "1 3 224 224"
+```
+
+g++ compilation command replaced with:
+```shell
+riscv64-unknown-linux-gnu-g++ main.cpp -I../prebuilt_opencv/include/opencv4 -L../prebuilt_opencv/lib   -lopencv_imgproc   -lopencv_imgcodecs -L../prebuilt_opencv/lib/opencv4/3rdparty/ -llibjpeg-turbo -llibwebp -llibpng -llibtiff -llibopenjp2    -lopencv_core -ldl  -lpthread -lrt -lzlib -lcsi_cv -latomic -static -o mobilenetv2_example
+```
+
+Then send the compiled binary file to the development board to run. The reference results are as follows:
+```shell
+(ort) root@lpi4a:/home/sipeed/onnx_mobilenetv2_c++# ./mobilenetv2_example 
+ ********** preprocess image **********
+ ********** run mobilenetv2 **********
+Run graph execution time: 79.77252ms, FPS=12.54
+
+=== tensor info ===
+shape: 1 3 224 224 
+data pointer: 0x259240
+
+=== tensor info ===
+shape: 1 1000 
+data pointer: 0x1c5200
+The max_value of output: 16.843750
+The min_value of output: -7.414062
+The mean_value of output: 0.001131
+The std_value of output: 9.056762
+ ============ top5: ===========
+283: 16.843750
+281: 13.789062
+287: 12.257812
+282: 10.898438
+285: 10.765625
+ ********** postprocess result **********
+ ********** probability top5: ********** 
+n02123394 Persian cat
+n02123045 tabby, tabby cat
+n02127052 lynx, catamount
+n02123159 tiger cat
+n02124075 Egyptian cat
+```
+
+## Yolov5n 
+
+### NPU
+
+#### Setup
+
+Refer to [Peripheral](https://wiki.sipeed.com/hardware/en/lichee/th1520/lpi4a/6_peripheral.html#NPU)After the NPU usage environment is set up, the document enters the Docker image of the HHB environment.
+
+First get the model for this tutorial and download it to the sample directory `/home/example/th1520_npu/yolov5n` :
+```shell
+git clone https://github.com/ultralytics/yolov5.git
+cd yolov5
+pip3 install ultralytics
+python3 export.py --weights yolov5n.pt --include onnx
+```
+
+#### Compile
+
+**HHB Compile**
+To cross-compile the ONNX model into an executable program on the NPU, you need to use the hhb command. Note that only 8-bit or 16-bit fixed-point operations are supported on the NPU, which is defined as int8 asymmetric quantization in this example. When compiling, you need to go to the directory where the sample is stored `/home/example/th1520_npu/yolov5n`:
+```shell
+cd /home/example/th1520_npu/yolov5n
+hhb -D --model-file yolov5n.onnx --data-scale-div 255 --board th1520 --input-name "images" --output-name "/model.24/m.0/Conv_output_0;/model.24/m.1/Conv_output_0;/model.24/m.2/Conv_output_0" --input-shape "1 3 384 640" --calibrate-dataset kite.jpg  --quantization-scheme "int8_asym"
+```
+
+Option description:
+- -D: specify the HHB process to generate the executable file until the end
+- --model-file: specify the yolov5 model that has been downloaded in the current directory
+- --data-mean: specify the mean value
+- --data-scale: specify the scale value
+- --board: specify the target platform as th1520
+- --input-name: the input name of the model
+- --output-name: the output name of the model
+- --input-shape: the input size of the model
+- --postprocess: save the output results and print the top5 results
+- --calibrate-dataset: specify the calibration images required for quantization
+- --quantization-scheme: specify the quantization scheme as int8 asymmetric
+
+After the command is executed, the hhb_out subdirectory will be generated in the current directory, which includes hhb_runtime, model. c and other files:
+- hhb. bm: HHB model file, including the weight data after quantization
+- hhb_runtime: the executable file on the th1520 platform, compiled from the c file in the directory
+- main. c: temporary file, the reference entry of the sample program
+- model. c: temporary file, model structure file, related to the model structure
+- model. params: temporary file, weight value
+- io. c: temporary file, auxiliary functions for reading and writing files
+- io. h: temporary file, auxiliary function declaration for reading and writing files
+- process. c: temporary file, image preprocessing function
+- process. h: temporary file, image preprocessing function declaration
+
+For more detailed HHB options, refer to the command line options in the [HHBUserManual](https://www.yuque.com/za4k4z/kvkcoh/shl1nybhel6iviwt).
+
+**g++ compile**
+```shell
+riscv64-unknown-linux-gnu-gcc yolov5n.c -o yolov5n_example hhb_out/io.c hhb_out/model.c -Wl,--gc-sections -O2 -g -mabi=lp64d -I hhb_out/ -L /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/lib/ -lshl -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/decode/install/lib/rv -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/runtime/riscv_linux -lprebuilt_runtime -ljpeg -lpng -lz -lstdc++ -lm -I /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/include/ -mabi=lp64d -march=rv64gcv0p7_zfh_xtheadc -Wl,-unresolved-symbols=ignore-in-shared-libs
+```
+
+The yolov5n_example file is generated in the example directory after the compilation command is executed correctly.
+
+#### Excute
+
+After the cross-compilation is complete, you can copy the files needed for program execution to the directory of the development board. The scp command can be used:
+```shell
+scp -r ../yolov5n sipeed@your_ip:~
+```
+
+First confirm whether the driver in the development board is loaded:
+```shell
+lsmod
+```
+If there are 'img_mem', 'vha' and 'vha_info' in the output of the three modules, the NPU driver is loaded successfully, if not, you can run the following command to manually load it:
+```shell
+sudo npu_init
+```
+
+Refer to [YOLOX](https://wiki.sipeed.com/hardware/en/lichee/th1520/lpi4a/8_application.html#YOLOX-Target-Detection) Install and configure the python virtual environment:
+```shell
+sudo apt update
+sudo apt install wget git vim
+
+wget https://github.com/T-head-Semi/csi-nn2/releases/download/v2.4-beta.1/c920.tar.gz
+tar xf c920.tar.gz
+cp c920/lib/* /usr/lib/riscv64-linux-gnu/ -rf
+
+sudo apt install python3-pip
+sudo apt install python3.11-venv
+
+cd /root
+python3 -m venv ort
+source /root/ort/bin/activate
+
+git clone -b python3.11 https://github.com/zhangwm-pt/prebuilt_whl.git
+cd prebuilt_whl
+
+pip install numpy-1.25.0-cp311-cp311-linux_riscv64.whl
+pip install opencv_python-4.5.4+4cd224d-cp311-cp311-linux_riscv64.whl
+pip install kiwisolver-1.4.4-cp311-cp311-linux_riscv64.whl
+pip install Pillow-9.5.0-cp311-cp311-linux_riscv64.whl
+pip install matplotlib-3.7.2.dev0+gb3bd929cf0.d20230630-cp311-cp311-linux_riscv64.whl
+pip install pycocotools-2.0.6-cp311-cp311-linux_riscv64.whl
+pip3 install loguru-0.7.0-py3-none-any.whl
+pip3 install torch-2.0.0a0+gitc263bd4-cp311-cp311-linux_riscv64.whl
+pip3 install MarkupSafe-2.1.3-cp311-cp311-linux_riscv64.whl
+pip3 install torchvision-0.15.1a0-cp311-cp311-linux_riscv64.whl
+pip3 install psutil-5.9.5-cp311-abi3-linux_riscv64.whl
+pip3 install tqdm-4.65.0-py3-none-any.whl
+pip3 install tabulate-0.9.0-py3-none-any.whl
+```
+
+Run the just-compiled example in the appropriate directory on the development board:
+```shell
+python3 inference.py 
+```
+
+After the execution, the terminal will prompt the stages of the execution:
+1. Preprocessing: fill and scale the original image to 384 * 640
+2. Model execution and post-processing: perform model inference and do post-processing such as nms
+3. Frame: draw the detection results on the 384 * 640 size graph and output the new image
+
+Files used in the execution of inference.py:
+- kite.jpg: the input image
+- image_preprocessed.bin: the intermediate results generated from the input image in the preprocessing stage
+- yolov5n_example: the file used in the model execution stage, compiled by gcc on an x86 host
+- hhb_out/hhb.bm: the file used in the model execution stage, generated by HHB on an x86 host
+- detect.txt: the output file of the post-processing stage, including the four targets detected in the image
+- kite_result.jpg: the output image, which adds the detection frame to the results of the input graph
+
+#### Results
+
+```shell
+(ort) root@lpi4a:/home/sipeed/yolov5n_npu# python3 inference.py 
+ ********** preprocess image **********
+ ******* run yolov5 and postprocess *******
+INFO: NNA clock:792000 [kHz]
+INFO: Heap :ocm (0x18)
+INFO: Heap :anonymous (0x2)
+INFO: Heap :dmabuf (0x2)
+INFO: Heap :unified (0x5)
+Run graph execution time: 11.68591ms, FPS=85.57
+detect num: 4
+id:     label   score           x1              y1              x2              y2
+[0]:    0       0.895277        273.492188      161.245056      359.559814      330.644257
+[1]:    0       0.887368        79.860062       179.181244      190.755692      354.304474
+[2]:    0       0.815214        222.054550      224.477600      279.828979      333.717285
+[3]:    33      0.563324        67.625580       173.948883      201.687988      219.065765
+ ********** draw bbox **********
+[273.492188, 161.245056, 359.559814, 330.644257, 0.895277, 0]
+[79.860062, 179.181244, 190.755692, 354.304474, 0.887368, 0]
+[222.05455, 224.4776, 279.828979, 333.717285, 0.815214, 0]
+[67.62558, 173.948883, 201.687988, 219.065765, 0.563324, 33]
+```
+
+> The sample image is from the web.
+
+![yolov5n_detection_soccer_output](./../../../../zh/lichee/th1520/lpi4a/assets/application/yolov5n_detection_soccer_output.jpg)
+
+### CPU
+Replace the HHB compile command in the NPU step above with:
+```shell
+hhb -D --model-file yolov5n.onnx --data-scale-div 255 --board c920 --input-name "images" --output-name "/model.24/m.0/Conv_output_0;/model.24/m.1/Conv_output_0;/model.24/m.2/Conv_output_0" --input-shape "1 3 384 640"
+```
+
+gcc compile command replaced with:
+```shell
+riscv64-unknown-linux-gnu-gcc yolov5n.c -static -o yolov5n_example hhb_out/io.c hhb_out/model.c -Wl,--gc-sections -O2 -g -mabi=lp64d -I hhb_out/ -L /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/c920/lib/ -lshl -static -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/decode/install/lib/rv -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/runtime/riscv_linux -lprebuilt_runtime -ljpeg -lpng -lz -lstdc++ -lm -I /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/c920/include/ -mabi=lp64d -march=rv64gcv0p7_zfh_xtheadc
+```
+
+Then send the compiled binary file to the development board to run. The reference results are as follows:
+```shell
+(ort) root@lpi4a:/home/sipeed/yolov5n_cpu# python3 inference.py 
+ ********** preprocess image **********
+ ******* run yolov5 and postprocess *******
+Run graph execution time: 387.34067ms, FPS=2.58
+detect num: 4
+id:     label   score           x1              y1              x2              y2
+[0]:    0       0.901887        274.524475      158.559036      359.169312      332.431702
+[1]:    0       0.879545        80.073883       184.767792      190.130157      349.906281
+[2]:    0       0.845192        219.378418      221.662415      283.860413      333.798584
+[3]:    33      0.666908        67.099136       174.128189      202.971451      220.213608
+ ********** draw bbox **********
+[274.524475, 158.559036, 359.169312, 332.431702, 0.901887, 0]
+[80.073883, 184.767792, 190.130157, 349.906281, 0.879545, 0]
+[219.378418, 221.662415, 283.860413, 333.798584, 0.845192, 0]
+[67.099136, 174.128189, 202.971451, 220.213608, 0.666908, 33]
+```
+
+## BERT
+
+### CPU
+
+#### Setup
+
+Refer to [Peripheral](https://wiki.sipeed.com/hardware/en/lichee/th1520/lpi4a/6_peripheral.html#NPU)After the NPU usage environment is set up, the document enters the Docker image of the HHB environment.
+
+The model used in this tutorial is from the google bert repository and has been converted to the onnx version of the BERT model, which can be downloaded to the `/home/example/c920/bert_small` directory with the following command:
+```shell
+cd home/example/c920/bert_small
+wget https://github.com/zhangwm-pt/bert/releases/download/onnx/bert_small_int32_input.onnx
+```
+
+#### Compile
+
+**HHB Compile**
+To cross-compile the ONNX model into an executable program on the NPU, you need to use the hhb command. Note that only 8-bit or 16-bit fixed-point operations are supported on the NPU, which is defined as int8 asymmetric quantization in this example. To compile, you need to first go to the directory where the sample is located `/home/example/c920/bert_small` :
+```shell
+cd /home/example/c920/bert_small
+hhb --model-file bert_small_int32_input.onnx --input-name "input_ids;input_mask;segment_ids" --input-shape '1 384;1 384;1 384' --output-name "output_start_logits;output_end_logits" --board c920 --quantization-scheme "float16" --postprocess save_and_top5 -D --without-preprocess
+```
+
+The options are:
+- -D: Specifies that the HHB process will run until the executable is generated
+- --model-file: Specifies that the bert model has been downloaded in the current directory
+- --data-mean: Specifies the mean value
+- --data-scale: Specifies the scale value
+- --board: Specifies that the target platform is th1520
+- --input-name: The input name of the model
+- --output-name: The output name of the model
+- --input-shape: The input size of the model
+- --postprocess: Saves the output results and prints the top 5 results
+- --calibrate-dataset: Specifies the calibration images required for quantization
+- --quantization-scheme: Specifies that the quantization scheme is int8 asymmetric
+
+After the command is executed, the hhb_out subdirectory will be generated in the current directory, which includes hhb_runtime, model.c and other files:
+- hhb. bm: The HHB model file, including the quantized weight data and other information
+- hhb_runtime: The executable on the th1520 platform, compiled from the c file in the directory
+- main. c: A temporary file, the reference entry of the sample program
+- model. c: A temporary file, the model structure file, related to the model structure
+- model.params: A temporary file, the weight values
+- io. c: A temporary file, the auxiliary functions for reading and writing files
+- io. h: A temporary file, the auxiliary functions for reading and writing files declaration
+- process. c: A temporary file, the image preprocessing function
+- process. h: A temporary file, the image preprocessing function declaration
+
+
+
+For more detailed HHB options, refer to the command line options in the [HHBUserManual](https://www.yuque.com/za4k4z/kvkcoh/shl1nybhel6iviwt).
+
+
+After the compilation command is executed correctly, an executable file is generated in the example directory. You can copy the example directory to the development board and run it.
+
+```shell
+scp -r ../bert_small sipeed@your_ip:~
+```
+
+#### Excute
+
+The preceding steps to execute the example program are the same as YOLOv5n, which will not be described here.
+After the preceding steps are correct, you can run commands in the example directory
+```shell
+python3 inference.py
+```
+
+#### Results
+
+The reference input in this example comes from the SQuAD dataset, a reading comprehension dataset consisting of questions posed by a set of Wikipedia articles, where the answer to each question is a piece of text from the corresponding reading article or question.
+The input for this example is as follows. The article describes the game of a rugby game and asks the question of who participated in the game.
+
+```
+[Context]:  Super Bowl 50 was an American football game to determine the champion of the National Football League (NFL) for the 2015 season. The American Football Conference (AFC) champion Denver Broncos defeated the National Football Conference (NFC) champion Carolina Panthers 24–10 to earn their third Super Bowl title. The game was played on February 7, 2016, at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California. As this was the 50th Super Bowl, the league emphasized the "golden anniversary" with various gold-themed initiatives, as well as temporarily suspending the tradition of naming each Super Bowl game with Roman numerals (under which the game would have been known as "Super Bowl L"), so that the logo could prominently feature the Arabic numerals 50.
+
+[Question]:  Which NFL team represented the AFC at Super Bowl 50?
+```
+Based on the reading comprehension results, the expected output would be the Denver Broncos
+
+```shell
+(ort) root@lpi4a:/home/sipeed/bert_small_cpu# python3 inference.py
+ ********** preprocess test **********
+[Context]:  Super Bowl 50 was an American football game to determine the champion of the National Football League (N
+FL) for the 2015 season. The American Football Conference (AFC) champion Denver Broncos defeated the National Footba
+ll Conference (NFC) champion Carolina Panthers 24–10 to earn their third Super Bowl title. The game was played on Fe
+bruary 7, 2016, at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California. As this was the 50th Sup
+er Bowl, the league emphasized the "golden anniversary" with various gold-themed initiatives, as well as temporarily
+ suspending the tradition of naming each Super Bowl game with Roman numerals (under which the game would have been k
+nown as "Super Bowl L"), so that the logo could prominently feature the Arabic numerals 50.
+[Question]:  Which NFL team represented the AFC at Super Bowl 50?
+ ******* run bert *******
+Run graph execution time: 1713.15491ms, FPS=0.58
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x183d60
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x185380
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x1869a0
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x2a8610
+The max_value of output: 3.826172
+The min_value of output: -9.968750
+The mean_value of output: -8.412353
+The std_value of output: 5.128320
+ ============ top5: ===========
+ 46: 3.826172
+ 57: 3.142578
+ 39: 1.303711
+ 38: 1.179688
+ 27: 0.624512
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x2a8300
+The max_value of output: 3.617188
+The min_value of output: -9.625000
+The mean_value of output: -7.798176
+The std_value of output: 4.820137
+ ============ top5: ===========
+ 47: 3.617188
+ 58: 3.482422
+ 32: 2.523438
+ 29: 1.541992
+ 41: 1.473633
+ ********** postprocess **********
+[Answer]:  Denver Broncos
+```
+
 ## Docker
 
 First install the required packages
diff --git a/docs/hardware/zh/lichee/th1520/lpi4a/3_images.md b/docs/hardware/zh/lichee/th1520/lpi4a/3_images.md
index 9d2eaeb5be..8f258c2f78 100644
--- a/docs/hardware/zh/lichee/th1520/lpi4a/3_images.md
+++ b/docs/hardware/zh/lichee/th1520/lpi4a/3_images.md
@@ -37,12 +37,12 @@ root 账户默认没有设置密码。
 （0721及以后版本的镜像内存能正常使用，无需替换为此处的文件）
 
 ```shell
-sudo ./fastboot flash ram ./images/u-boot-with-spl-ddr16g.bin
+sudo ./fastboot flash ram ./images/u-boot-with-spl-lpi4a-16g.bin
 sudo ./fastboot reboot
 sleep 1
-sudo ./fastboot flash uboot ./images/u-boot-with-spl-ddr16g.bin
+sudo ./fastboot flash uboot ./images/u-boot-with-spl-lpi4a-16g.bin
 # 若自己使用的 boot.ext4 中没有 16G ddr 对应的设备树，则需要再烧录 16G ddr 对应的 boot.ext4
-sudo ./fastboot flash boot ./images/boot_16gddr.ext4
+sudo ./fastboot flash boot ./images/boot.ext4
 ```
 
 ### Debian
@@ -55,9 +55,9 @@ sudo ./fastboot flash boot ./images/boot_16gddr.ext4
 Mega 云盘：[点我](https://mega.nz/folder/phoQlBTZ#cZeQ3qZ__pDvP94PT3_bGA)
 ISCAS 镜像站：[点我](https://mirror.iscas.ac.cn/revyos/extra/images/lpi4a/)
 
-1. LPI4A_20230721.zip
+1. LPI4A_20230912.zip
 
-   - 发布日期：2023年7月21日
+   - 发布日期：2023年9月12日
    - 内核：5.10
    - 根文件系统：ext4, 4.3GB
    - 主要预装软件包：
@@ -74,11 +74,11 @@ ISCAS 镜像站：[点我](https://mirror.iscas.ac.cn/revyos/extra/images/lpi4a/
       - HDMI 音频在 chromium 下播放音质有问题
    - 使用说明：
       - 该压缩包为仅支持 HDMI 显示，烧录时请注意文件名中的后缀和设备相关参数对应
-         - boot，u-boot 文件名中的 8gddr/16gddr 分别对应 8g/16g 内存
+         - u-boot 带 16g 后缀的为支持 16g 内存的u-boot，不带后缀的为支持 8g 内存的 u-boot
 
-2. LPI4A_20230721_mipi.zip
+2. LPI4A_20230912_MIPI.zip
 
-   - 发布日期：2023年7月21日
+   - 发布日期：2023年9月12日
    - 内核：5.10
    - 根文件系统：ext4, 4.3GB
    - 主要预装软件包：
@@ -100,8 +100,7 @@ ISCAS 镜像站：[点我](https://mirror.iscas.ac.cn/revyos/extra/images/lpi4a/
       - HDMI 音频在 chromium 下播放音质有问题
    - 使用说明：
       - 该压缩包为支持 MIPI 屏幕的镜像，烧录时请注意文件名中的后缀和设备相关参数对应
-         - boot，u-boot 文件名中的 8gddr/16gddr 分别对应 8g/16g 内存
-         - boot 文件名中的 mipi_720p/mipi_1080p 分别对应800x1280的 MIPI 屏幕（带触摸）/1200x1920的mipi屏幕
+         - u-boot 带 16g 后缀的为支持 16g 内存的u-boot，不带后缀的为支持 8g 内存的 u-boot
 
 ### OpenWRT
 
diff --git a/docs/hardware/zh/lichee/th1520/lpi4a/4_burn_image.md b/docs/hardware/zh/lichee/th1520/lpi4a/4_burn_image.md
index 65ad2bbb35..2a0727ce20 100644
--- a/docs/hardware/zh/lichee/th1520/lpi4a/4_burn_image.md
+++ b/docs/hardware/zh/lichee/th1520/lpi4a/4_burn_image.md
@@ -97,21 +97,21 @@ b. 打开设备管理器出现“USB download gadget”设备。
 这时侯我们可以通过fastboot下载并启动u-boot镜像的命令，来进入到u-boot的fastboot烧录模式（相比Brom阶段，会有更大下载buffer，速度会更快）
 下面的指令会检查并格式化分区，请务必执行，否则后面烧录 rootfs 会很慢。
 
-`u-boot-with-spl-ddr8g.bin` 和 `u-boot-with-spl-ddr16g.bin` 为u-boot 固件，具体差异请参考镜像说明。
+`u-boot-with-spl-lpi4a.bin` 和 `u-boot-with-spl-lpi4a-16g.bin` 为u-boot 固件，具体差异请参考镜像说明。
 ```bash
-sudo ./fastboot flash ram ./images/u-boot-with-spl-ddr16g.bin
+sudo ./fastboot flash ram ./images/u-boot-with-spl-lpi4a-16g.bin
 sudo ./fastboot reboot
 sleep 1
 ```
 
 分别烧录下面三个镜像：启动引导镜像-uboot，启动分区-boot，操作系统根分区-root
 ```bash
-sudo ./fastboot flash uboot ./images/u-boot-with-spl-ddr16g.bin
-sudo ./fastboot flash boot ./images/boot_16gddr.ext4
-sudo ./fastboot flash root ./images/rootfs-thead-image-linux_sing.ext4.ext4
+sudo ./fastboot flash uboot ./images/u-boot-with-spl-lpi4a-16g.bin
+sudo ./fastboot flash boot ./images/boot_sing.ext4
+sudo ./fastboot flash root ./images/rootfs-sing.ext4
 ```
 
-`boot_8gddr.ext4`，`boot_8gddr_mipi_720p.ext4`，`boot_8gddr_mipi_1080p.ext4`，`boot_16gddr.ext4`，`boot_16gddr_mipi_720p.ext4` 和 `boot_16gddr_mipi_1080p.ext4` 为 boot 分区，具体差异请参考镜像说明。它们主要包含以下内容： 
+`boot-sing.ext4` 和 `boot-dual.ext4` 为 boot 分区，具体差异请参考镜像说明。它们主要包含以下内容： 
 ```bash
 fw_dynamic.bin        #opensbi
 Image                 #kernel image
@@ -122,7 +122,7 @@ light-lpi4a.dtb       # ddr8G dtb
 light-lpi4a-ddr16g.dtb # ddr16G dtb
 ```
 
-`rootfs-thead-image-linux_sing.ext4` 和 `rootfs-thead-image-linux_mipi.ext4` 为根文件系统，默认为 Debian 系统。两者的具体差异请参考镜像说明。
+`rootfs-sing.ext4` 和 `rootfs-mipi.ext4` 为根文件系统，默认为 Debian 系统。两者的具体差异请参考镜像说明。
 
 烧录镜像的典型 log 输出如下：
 
diff --git a/docs/hardware/zh/lichee/th1520/lpi4a/6_peripheral.md b/docs/hardware/zh/lichee/th1520/lpi4a/6_peripheral.md
index 6df2042a7a..79082caa06 100644
--- a/docs/hardware/zh/lichee/th1520/lpi4a/6_peripheral.md
+++ b/docs/hardware/zh/lichee/th1520/lpi4a/6_peripheral.md
@@ -2,6 +2,11 @@
 title: 外设使用
 keywords: Linux, Lichee, TH1520, SBC, RISCV, Peripheral
 update:
+  - date: 2023-09-12
+    version: v1.3
+    author: ztd
+    content:
+      - Update NPU user guide
   - date: 2023-08-18
     version: v1.2
     author: ztd
@@ -850,7 +855,73 @@ sudo watch cat /sys/kernel/debug/pvr/status
 
 ## NPU
 
-TODO
+> **注意**：使用 NPU 驱动，需要升级到[20230912](https://pan.baidu.com/e/1xH56ZlewB6UOMlke5BrKWQ)版本镜像。
+
+NPU 相关驱动初始化命令
+```shell
+sudo npu_init
+```
+
+LicheePi4A 板载了一颗支持 4TOPS@INT8 通用 NNA 算力，主频 1GHz 的 NPU。Wiki 中包含的支持 NPU 的示例如下：
+
+|示例名称|示例功能|使用模型|接口|HHB版本|
+|---|---|---|---|---|
+|Mobilenetv2 做图像分类|图像分类|mobilenetv2|c/c++|2.4及以上|
+|YOLOv5 做目标检测|目标检测|YOLOv5|python|2.4及以上|
+|RTMPose 做姿态估计|姿态估计|RTMPose|python|2.6及以上|
+
+为了将上述示例中的模型交叉编译为 LicheePi4A 上的可执行程序，我们首先需要在自己的电脑上搭建 HHB 开发环境。
+> 推荐环境：ubuntu20.04 系统，Docker 使用20.10.21版本。推荐使用 Docker 镜像来搭建环境。
+
+### 安装
+首先要在自己的电脑上安装 Docker，先卸载可能存在的 Docker 版本：
+```shell
+sudo apt-get remove docker docker-engine docker.io containerd runc
+```
+
+安装Docker依赖的基础软件：
+```shell
+sudo apt-get update
+sudo apt-get install apt-transport-https ca-certificates curl gnupg-agent software-properties-common
+```
+
+添加官方源：
+```shell
+curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -
+sudo add-apt-repository "deb [arch=amd64] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable"
+```
+
+安装 Docker：
+```shell
+sudo apt-get update
+sudo apt-get install docker-ce docker-ce-cli containerd.io
+```
+
+安装完毕后，获取 HHB 环境的 Docker 镜像：
+```shell
+docker pull hhb4tools/hhb:2.4.5
+```
+
+拉取镜像完毕后，使用下面的命令进入 Docker 镜像：
+```shell
+docker run -itd --name=your.hhb2.4 -p 22 "hhb4tools/hhb:2.4.5"
+docker exec -it your.hhb2.4 /bin/bash
+```
+
+进入 Docker 镜像后，可使用下面的命令确认 HHB 版本：
+```shell
+hhb --version
+```
+
+进入 Docker 镜像中后，还需要配置交叉编译环境。注意必须要使用这这里的工具链，否则编译出的二进制文件无法在 LicheePi4A 上运行。
+```shell
+export PATH=/tools/Xuantie-900-gcc-linux-5.10.4-glibc-x86_64-V2.6.1-light.1/bin/:$PATH
+```
+
+至此，HHB 环境初步搭建完成。可以尝试以下的 NPU 示例：
+
+[Mobilenetv2 做图像分类](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/8_application.html#MobilenertV2)
+[YOLOv5 做目标检测](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/8_application.html#Yolov5n)
 
 ## 其它
 欢迎投稿～ 投稿接受后可得￥5～150（$1~20）优惠券！
diff --git a/docs/hardware/zh/lichee/th1520/lpi4a/8_application.md b/docs/hardware/zh/lichee/th1520/lpi4a/8_application.md
index 2704712863..9223e29d39 100644
--- a/docs/hardware/zh/lichee/th1520/lpi4a/8_application.md
+++ b/docs/hardware/zh/lichee/th1520/lpi4a/8_application.md
@@ -2,6 +2,11 @@
 title: 典型应用
 keywords: Linux, Lichee, TH1520, SBC, RISCV, application
 update:
+  - date: 2023-09-12
+    version: v1.2
+    author: ztd
+    content:
+      - Add some NPU applications
   - date: 2023-07-20
     version: v1.1
     author: ztd
@@ -30,7 +35,7 @@ llama 是 META 开源的大语言模型，[llama.cpp](https://github.com/ggergan
 
 ## YOLOX 目标检测
 
-本教程是一个如何在 LPi4A（LicheePi 4A） 开发板平台上部署 [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) 模型完成目标检测的示例。
+本教程是一个如何在 LPi4A（LicheePi 4A） 开发板平台上部署 [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) 模型完成目标检测的示例（CPU推理）。
 教程中包括了：
 - 在 LPi4A 开发版上安装 Python 环境
 - 使用 YOLOX 项目中的源码执行模型
@@ -184,6 +189,540 @@ python3 onnx_inference.py -m yolox_s.onnx -i soccer.jpg -o outdir -s 0.3 --input
 
 ![yolox_detection_soccer_output.jpg](./assets/application/yolox_detection_soccer_output.jpg)
 
+## MobilenertV2 
+
+本教程是一个如何在 LicheePi4A 平台上部署 mobilenetv2 模型完成图像分类的示例。
+
+教程中包括了：
+- 使用 HHB 编译 onnx 模型为 LicheePi4A 上可用的二进制
+- 在 LicheePi4A 上使用 opencv c++ 版本做 mobilenetv2 模型的预处理
+- 在 LicheePi4A 上使用 CPU 和 NPU 的差异
+
+### NPU
+
+#### 环境配置
+
+参考[外设使用](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/6_peripheral.html#NPU)文档搭建好 NPU 使用相关环境后，进入到 HHB 环境的 Docker 镜像中。
+
+首先获取本节教程的模型，下载到示例目录 `/home/example/th1520_npu/onnx_mobilenetv2_c++` 下：
+[mobilenetv2-12.onnx](https://github.com/onnx/models/blob/main/vision/classification/mobilenet/model/mobilenetv2-12.onnx)
+
+并获取本次教程所使用的优化版本 opencv 所需的库文件，前往 [github仓库下载](https://xuantie.t-head.cn/community/download?id=4112956065753141248)下载到上一级目录 `/home/example/th1520_npu/` 下。
+```shell
+cd /home/example/th1520_npu/
+git clone https://github.com/zhangwm-pt/prebuilt_opencv.git
+```
+
+#### 编译
+
+**HHB 编译模型：**
+将 ONNX 模型交叉编译成 NPU 上可执行的程序，需要使用 hhb 命令。注意，NPU 上仅支持8位或者16位定点运算，本示例中指定为 int8 非对称量化。编译时需要先进入到示例所在目录 `/home/example/th1520_npu/onnx_mobilenetv2_c++`：
+```shell
+cd /home/example/th1520_npu/onnx_mobilenetv2_c++
+hhb -D --model-file mobilenetv2-12.onnx --data-scale 0.017 --data-mean "124 117 104"  --board th1520  --postprocess save_and_top5 --input-name "input" --output-name "output" --input-shape "1 3 224 224" --calibrate-dataset persian_cat.jpg  --quantization-scheme "int8_asym"
+```
+
+选项说明：
+- -D ：指定 HHB 流程执行到生成可执行文件的阶段为止
+- --model-file ：指定当前目录中已经下载好的 mobilenet 模型
+- --data-mean ：指定均值
+- --data-scale ：指定缩放值
+- --board ：指定目标平台为 th1520
+- --input-name： 模型的输入名
+- --output-name：模型的输出名
+- --input-shape：模型的输入大小
+- --postprocess：保存输出结果，并且打印 top5 结果
+- --calibrate-dataset：指定量化时所需的校准图片
+- --quantization-scheme：指定量化方式为 int8 非对称
+
+命令执行完成后，会在当前目录生成 hhb_out 子目录，里面的包括了 hhb_runtime，model.c 等多个文件：
+- hhb.bm：HHB 的模型文件，包括了量化后的权重数据等信息
+- hhb_runtime：th1520 平台上的可执行文件，由目录中的c文件编译而成
+- main.c：临时文件，示例程序的参考入口
+- model.c：临时文件，模型结构文件，与模型结构相关
+- model.params：临时文件，权重数值
+- io.c：临时文件，读写文件的辅助函数
+- io.h：临时文件，读写文件的辅助函数声明
+- process.c：临时文件，图像预处理函数
+- process.h：临时文件，图像预处理函数声明
+
+更详细的 HHB 选项说明可以参考 [HHB用户手册](https://www.yuque.com/za4k4z/oxlbxl/keyg70qggt5n3fpa)中的命令行选项说明。
+
+**g++编译示例：**
+```shell
+riscv64-unknown-linux-gnu-g++ main.cpp -I../prebuilt_opencv/include/opencv4 -L../prebuilt_opencv/lib   -lopencv_imgproc   -lopencv_imgcodecs -L../prebuilt_opencv/lib/opencv4/3rdparty/ -llibjpeg-turbo -llibwebp -llibpng -llibtiff -llibopenjp2    -lopencv_core -ldl  -lpthread -lrt -lzlib -lcsi_cv -latomic -static -o mobilenetv2_example
+```
+
+编译命令正确执行完成后会在示例目录生成 mobilenetv2_example 文件。
+
+#### 执行
+
+交叉编译完成后，即可将程序执行所需的文件复制到开发板的目录中。可以使用 scp 命令：
+```shell
+scp -r ../onnx_mobilenetv2_c++ sipeed@你的开发板ip:~
+```
+
+先确认开发板中驱动是否加载：
+```shell
+lsmod
+```
+若在输出中有 `img_mem`，`vha` 和 `vha_info` 这三个模块，NPU驱动即加载成功，若没有则可以先运行下面的命令手动加载一下：
+```shell
+sudo npu_init
+```
+
+参考 [YOLOX章节](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/8_application.html#YOLOX-%E7%9B%AE%E6%A0%87%E6%A3%80%E6%B5%8B) 安装并配置好 python 虚拟环境：
+```shell
+sudo apt update
+sudo apt install wget git vim
+
+wget https://github.com/T-head-Semi/csi-nn2/releases/download/v2.4-beta.1/c920.tar.gz
+tar xf c920.tar.gz
+cp c920/lib/* /usr/lib/riscv64-linux-gnu/ -rf
+
+sudo apt install python3-pip
+sudo apt install python3.11-venv
+
+cd /root
+python3 -m venv ort
+source /root/ort/bin/activate
+
+git clone -b python3.11 https://github.com/zhangwm-pt/prebuilt_whl.git
+cd prebuilt_whl
+
+pip install numpy-1.25.0-cp311-cp311-linux_riscv64.whl
+pip install opencv_python-4.5.4+4cd224d-cp311-cp311-linux_riscv64.whl
+pip install kiwisolver-1.4.4-cp311-cp311-linux_riscv64.whl
+pip install Pillow-9.5.0-cp311-cp311-linux_riscv64.whl
+pip install matplotlib-3.7.2.dev0+gb3bd929cf0.d20230630-cp311-cp311-linux_riscv64.whl
+pip install pycocotools-2.0.6-cp311-cp311-linux_riscv64.whl
+pip3 install loguru-0.7.0-py3-none-any.whl
+pip3 install torch-2.0.0a0+gitc263bd4-cp311-cp311-linux_riscv64.whl
+pip3 install MarkupSafe-2.1.3-cp311-cp311-linux_riscv64.whl
+pip3 install torchvision-0.15.1a0-cp311-cp311-linux_riscv64.whl
+pip3 install psutil-5.9.5-cp311-abi3-linux_riscv64.whl
+pip3 install tqdm-4.65.0-py3-none-any.whl
+pip3 install tabulate-0.9.0-py3-none-any.whl
+```
+
+在开发板相应目录下运行刚刚编译好的示例：
+```shell
+./mobilenetv2_example
+```
+
+执行完成后，会在终端上提示执行到的各个阶段：
+1. 预处理
+2. 模型执行
+3. 后处理
+
+mobilenetv2_example 执行会使用到的文件：
+- persian_cat.jpg：输入图片
+- input_img.bin：预处理阶段，根据输入图片生成的中间结果
+- hhb_out/hhb_runtime：模型执行阶段使用的文件，由x86主机上 HHB 生成
+- hhb_out/hhb.bm：模型执行阶段使用的文件，由x86主机上 HHB 生成
+- input_img.bin_output0_1_1000.txt：模型执行阶段的输出文件，包括了模型执行输出的 1000 个结果数值
+
+#### 参考结果
+
+```shell
+(ort) root@lpi4a:/home/sipeed/onnx_mobilenetv2_c++# ./mobilenetv2_example
+ ********** preprocess image **********
+ ********** run mobilenetv2 **********
+INFO: NNA clock:792000 [kHz]
+INFO: Heap :ocm (0x18)
+INFO: Heap :anonymous (0x2)
+INFO: Heap :dmabuf (0x2)
+INFO: Heap :unified (0x5)
+Run graph execution time: 7.87149ms, FPS=127.04
+
+=== tensor info ===
+shape: 1 3 224 224 
+data pointer: 0x857ca0
+
+=== tensor info ===
+shape: 1 1000 
+data pointer: 0x3fc9abe000
+The max_value of output: 16.053827
+The min_value of output: -8.026914
+The mean_value of output: -0.001889
+The std_value of output: 9.203342
+ ============ top5: ===========
+283: 16.053827
+281: 14.165141
+287: 11.709850
+285: 11.615416
+282: 11.332113
+ ********** postprocess result **********
+ ********** probability top5: ********** 
+n02123394 Persian cat
+n02123045 tabby, tabby cat
+n02127052 lynx, catamount
+n02124075 Egyptian cat
+n02123159 tiger cat
+```
+
+### CPU
+将上述 NPU 步骤中的 HHB 编译命令替换为：
+```shell
+hhb -D --model-file mobilenetv2-12.onnx --data-scale 0.017 --data-mean "124 117 104"  --board c920  --postprocess save_and_top5 --input-name "input" --output-name "output" --input-shape "1 3 224 224"
+```
+
+g++ 编译后处理命令替换为：
+```shell
+riscv64-unknown-linux-gnu-g++ main.cpp -I../prebuilt_opencv/include/opencv4 -L../prebuilt_opencv/lib   -lopencv_imgproc   -lopencv_imgcodecs -L../prebuilt_opencv/lib/opencv4/3rdparty/ -llibjpeg-turbo -llibwebp -llibpng -llibtiff -llibopenjp2    -lopencv_core -ldl  -lpthread -lrt -lzlib -lcsi_cv -latomic -static -o mobilenetv2_example
+```
+
+再将编译的到的二进制文件发送到开发板上运行即可。参考结果如下：
+```shell
+(ort) root@lpi4a:/home/sipeed/onnx_mobilenetv2_c++# ./mobilenetv2_example 
+ ********** preprocess image **********
+ ********** run mobilenetv2 **********
+Run graph execution time: 79.77252ms, FPS=12.54
+
+=== tensor info ===
+shape: 1 3 224 224 
+data pointer: 0x259240
+
+=== tensor info ===
+shape: 1 1000 
+data pointer: 0x1c5200
+The max_value of output: 16.843750
+The min_value of output: -7.414062
+The mean_value of output: 0.001131
+The std_value of output: 9.056762
+ ============ top5: ===========
+283: 16.843750
+281: 13.789062
+287: 12.257812
+282: 10.898438
+285: 10.765625
+ ********** postprocess result **********
+ ********** probability top5: ********** 
+n02123394 Persian cat
+n02123045 tabby, tabby cat
+n02127052 lynx, catamount
+n02123159 tiger cat
+n02124075 Egyptian cat
+```
+
+## Yolov5n 
+
+### NPU
+
+#### 环境配置
+
+参考[外设使用](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/6_peripheral.html#NPU)文档搭建好 NPU 使用相关环境后，进入到 HHB 环境的 Docker 镜像中。
+
+首先获取本节教程的模型，下载到示例目录 `/home/example/th1520_npu/yolov5n` 下：
+```shell
+git clone https://github.com/ultralytics/yolov5.git
+cd yolov5
+pip3 install ultralytics
+python3 export.py --weights yolov5n.pt --include onnx
+```
+
+#### 编译
+
+**HHB 编译模型：**
+将 ONNX 模型交叉编译成 NPU 上可执行的程序，需要使用 hhb 命令。注意，NPU 上仅支持8位或者16位定点运算，本示例中指定为 int8 非对称量化。编译时需要先进入到示例所在目录 `/home/example/th1520_npu/yolov5n`：
+```shell
+cd /home/example/th1520_npu/yolov5n
+hhb -D --model-file yolov5n.onnx --data-scale-div 255 --board th1520 --input-name "images" --output-name "/model.24/m.0/Conv_output_0;/model.24/m.1/Conv_output_0;/model.24/m.2/Conv_output_0" --input-shape "1 3 384 640" --calibrate-dataset kite.jpg  --quantization-scheme "int8_asym"
+```
+
+选项说明：
+- -D ：指定 HHB 流程执行到生成可执行文件的阶段为止
+- --model-file ：指定当前目录中已经下载好的 yolov5 模型
+- --data-mean ：指定均值
+- --data-scale ：指定缩放值
+- --board ：指定目标平台为 th1520
+- --input-name： 模型的输入名
+- --output-name：模型的输出名
+- --input-shape：模型的输入大小
+- --postprocess：保存输出结果，并且打印 top5 结果
+- --calibrate-dataset：指定量化时所需的校准图片
+- --quantization-scheme：指定量化方式为 int8 非对称
+
+命令执行完成后，会在当前目录生成 hhb_out 子目录，里面的包括了 hhb_runtime，model.c 等多个文件：
+- hhb.bm：HHB 的模型文件，包括了量化后的权重数据等信息
+- hhb_runtime：th1520 平台上的可执行文件，由目录中的c文件编译而成
+- main.c：临时文件，示例程序的参考入口
+- model.c：临时文件，模型结构文件，与模型结构相关
+- model.params：临时文件，权重数值
+- io.c：临时文件，读写文件的辅助函数
+- io.h：临时文件，读写文件的辅助函数声明
+- process.c：临时文件，图像预处理函数
+- process.h：临时文件，图像预处理函数声明
+
+更详细的 HHB 选项说明可以参考 [HHB用户手册](https://www.yuque.com/za4k4z/oxlbxl/keyg70qggt5n3fpa)中的命令行选项说明。
+
+**g++编译示例：**
+```shell
+riscv64-unknown-linux-gnu-gcc yolov5n.c -o yolov5n_example hhb_out/io.c hhb_out/model.c -Wl,--gc-sections -O2 -g -mabi=lp64d -I hhb_out/ -L /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/lib/ -lshl -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/decode/install/lib/rv -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/runtime/riscv_linux -lprebuilt_runtime -ljpeg -lpng -lz -lstdc++ -lm -I /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/th1520/include/ -mabi=lp64d -march=rv64gcv0p7_zfh_xtheadc -Wl,-unresolved-symbols=ignore-in-shared-libs
+```
+
+编译命令正确执行完成后会在示例目录生成 yolov5n_example 文件。
+
+#### 执行
+
+交叉编译完成后，即可将程序执行所需的文件复制到开发板的目录中。可以使用 scp 命令：
+```shell
+scp -r ../yolov5n sipeed@你的开发板ip:~
+```
+
+先确认开发板中驱动是否加载：
+```shell
+lsmod
+```
+若在输出中有 `img_mem`，`vha` 和 `vha_info` 这三个模块，NPU驱动即加载成功，若没有则可以先运行下面的命令手动加载一下：
+```shell
+sudo npu_init
+```
+
+参考 [YOLOX章节](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/8_application.html#YOLOX-%E7%9B%AE%E6%A0%87%E6%A3%80%E6%B5%8B) 安装并配置好 python 虚拟环境：
+```shell
+sudo apt update
+sudo apt install wget git vim
+
+wget https://github.com/T-head-Semi/csi-nn2/releases/download/v2.4-beta.1/c920.tar.gz
+tar xf c920.tar.gz
+cp c920/lib/* /usr/lib/riscv64-linux-gnu/ -rf
+
+sudo apt install python3-pip
+sudo apt install python3.11-venv
+
+cd /root
+python3 -m venv ort
+source /root/ort/bin/activate
+
+git clone -b python3.11 https://github.com/zhangwm-pt/prebuilt_whl.git
+cd prebuilt_whl
+
+pip install numpy-1.25.0-cp311-cp311-linux_riscv64.whl
+pip install opencv_python-4.5.4+4cd224d-cp311-cp311-linux_riscv64.whl
+pip install kiwisolver-1.4.4-cp311-cp311-linux_riscv64.whl
+pip install Pillow-9.5.0-cp311-cp311-linux_riscv64.whl
+pip install matplotlib-3.7.2.dev0+gb3bd929cf0.d20230630-cp311-cp311-linux_riscv64.whl
+pip install pycocotools-2.0.6-cp311-cp311-linux_riscv64.whl
+pip3 install loguru-0.7.0-py3-none-any.whl
+pip3 install torch-2.0.0a0+gitc263bd4-cp311-cp311-linux_riscv64.whl
+pip3 install MarkupSafe-2.1.3-cp311-cp311-linux_riscv64.whl
+pip3 install torchvision-0.15.1a0-cp311-cp311-linux_riscv64.whl
+pip3 install psutil-5.9.5-cp311-abi3-linux_riscv64.whl
+pip3 install tqdm-4.65.0-py3-none-any.whl
+pip3 install tabulate-0.9.0-py3-none-any.whl
+```
+
+在开发板相应目录下运行刚刚编译好的示例：
+```shell
+python3 inference.py 
+```
+
+执行完成后，会在终端上提示执行到的各个阶段：
+1. 预处理：将原图填充缩放到 384 * 640 的大小
+2. 模型执行和后处理：执行模型推理，并做 nms 等后处理
+3. 画框：将检测结果画在 384 *640 尺寸的图上，并输出新图片
+
+inference.py 执行会使用到的文件：
+- kite.jpg：输入图片
+- image_preprocessed.bin：预处理阶段，根据输入图片生成的中间结果
+- yolov5n_example：模型执行阶段使用的文件，由x86主机上 gcc 编译生成
+- hhb_out/hhb.bm：模型执行阶段使用的文件，由x86主机上 HHB 生成
+- detect.txt：后处理阶段的输出文件，包括了图片中检测出来的 4 个目标
+- kite_result.jpg：输出图片，将检测框加入到输入图上的结果
+
+#### 参考结果
+
+```shell
+(ort) root@lpi4a:/home/sipeed/yolov5n_npu# python3 inference.py 
+ ********** preprocess image **********
+ ******* run yolov5 and postprocess *******
+INFO: NNA clock:792000 [kHz]
+INFO: Heap :ocm (0x18)
+INFO: Heap :anonymous (0x2)
+INFO: Heap :dmabuf (0x2)
+INFO: Heap :unified (0x5)
+Run graph execution time: 11.68591ms, FPS=85.57
+detect num: 4
+id:     label   score           x1              y1              x2              y2
+[0]:    0       0.895277        273.492188      161.245056      359.559814      330.644257
+[1]:    0       0.887368        79.860062       179.181244      190.755692      354.304474
+[2]:    0       0.815214        222.054550      224.477600      279.828979      333.717285
+[3]:    33      0.563324        67.625580       173.948883      201.687988      219.065765
+ ********** draw bbox **********
+[273.492188, 161.245056, 359.559814, 330.644257, 0.895277, 0]
+[79.860062, 179.181244, 190.755692, 354.304474, 0.887368, 0]
+[222.05455, 224.4776, 279.828979, 333.717285, 0.815214, 0]
+[67.62558, 173.948883, 201.687988, 219.065765, 0.563324, 33]
+```
+
+> 示例图片来源于网络。
+
+![yolov5n_detection_soccer_output](./assets/application/yolov5n_detection_soccer_output.jpg)
+
+### CPU
+将上述 NPU 步骤中的 HHB 编译命令替换为：
+```shell
+hhb -D --model-file yolov5n.onnx --data-scale-div 255 --board c920 --input-name "images" --output-name "/model.24/m.0/Conv_output_0;/model.24/m.1/Conv_output_0;/model.24/m.2/Conv_output_0" --input-shape "1 3 384 640"
+```
+
+gcc 编译后处理命令替换为：
+```shell
+riscv64-unknown-linux-gnu-gcc yolov5n.c -static -o yolov5n_example hhb_out/io.c hhb_out/model.c -Wl,--gc-sections -O2 -g -mabi=lp64d -I hhb_out/ -L /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/c920/lib/ -lshl -static -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/decode/install/lib/rv -L /usr/local/lib/python3.8/dist-packages/hhb/prebuilt/runtime/riscv_linux -lprebuilt_runtime -ljpeg -lpng -lz -lstdc++ -lm -I /usr/local/lib/python3.8/dist-packages/hhb/install_nn2/c920/include/ -mabi=lp64d -march=rv64gcv0p7_zfh_xtheadc
+```
+
+再将编译的到的二进制文件发送到开发板上运行即可。参考结果如下：
+```shell
+(ort) root@lpi4a:/home/sipeed/yolov5n_cpu# python3 inference.py 
+ ********** preprocess image **********
+ ******* run yolov5 and postprocess *******
+Run graph execution time: 387.34067ms, FPS=2.58
+detect num: 4
+id:     label   score           x1              y1              x2              y2
+[0]:    0       0.901887        274.524475      158.559036      359.169312      332.431702
+[1]:    0       0.879545        80.073883       184.767792      190.130157      349.906281
+[2]:    0       0.845192        219.378418      221.662415      283.860413      333.798584
+[3]:    33      0.666908        67.099136       174.128189      202.971451      220.213608
+ ********** draw bbox **********
+[274.524475, 158.559036, 359.169312, 332.431702, 0.901887, 0]
+[80.073883, 184.767792, 190.130157, 349.906281, 0.879545, 0]
+[219.378418, 221.662415, 283.860413, 333.798584, 0.845192, 0]
+[67.099136, 174.128189, 202.971451, 220.213608, 0.666908, 33]
+```
+
+## BERT
+
+### CPU
+
+#### 环境配置
+
+参考[外设使用](https://wiki.sipeed.com/hardware/zh/lichee/th1520/lpi4a/6_peripheral.html#NPU)文档搭建好 NPU 使用相关环境后，进入到 HHB 环境的 Docker 镜像中。
+
+首先获取模型，本教程中使用的模型来自 google bert 仓库，已转换成 onnx 版本的 BERT 模型，可以用如下命令下载到 `/home/example/c920/bert_small` 目录下:
+```shell
+cd home/example/c920/bert_small
+wget https://github.com/zhangwm-pt/bert/releases/download/onnx/bert_small_int32_input.onnx
+```
+
+#### 编译
+
+**HHB 编译模型：**
+将 ONNX 模型交叉编译成 NPU 上可执行的程序，需要使用 hhb 命令。注意，NPU 上仅支持8位或者16位定点运算，本示例中指定为 int8 非对称量化。编译时需要先进入到示例所在目录 `/home/example/c920/bert_small`：
+```shell
+cd /home/example/c920/bert_small
+hhb --model-file bert_small_int32_input.onnx --input-name "input_ids;input_mask;segment_ids" --input-shape '1 384;1 384;1 384' --output-name "output_start_logits;output_end_logits" --board c920 --quantization-scheme "float16" --postprocess save_and_top5 -D --without-preprocess
+```
+
+选项说明：
+- -D ：指定 HHB 流程执行到生成可执行文件的阶段为止
+- --model-file ：指定当前目录中已经下载好的 bert 模型
+- --data-mean ：指定均值
+- --data-scale ：指定缩放值
+- --board ：指定目标平台为 th1520
+- --input-name： 模型的输入名
+- --output-name：模型的输出名
+- --input-shape：模型的输入大小
+- --postprocess：保存输出结果，并且打印 top5 结果
+- --calibrate-dataset：指定量化时所需的校准图片
+- --quantization-scheme：指定量化方式为 int8 非对称
+
+命令执行完成后，会在当前目录生成 hhb_out 子目录，里面的包括了 hhb_runtime，model.c 等多个文件：
+- hhb.bm：HHB 的模型文件，包括了量化后的权重数据等信息
+- hhb_runtime：th1520 平台上的可执行文件，由目录中的c文件编译而成
+- main.c：临时文件，示例程序的参考入口
+- model.c：临时文件，模型结构文件，与模型结构相关
+- model.params：临时文件，权重数值
+- io.c：临时文件，读写文件的辅助函数
+- io.h：临时文件，读写文件的辅助函数声明
+- process.c：临时文件，图像预处理函数
+- process.h：临时文件，图像预处理函数声明
+
+更详细的 HHB 选项说明可以参考 [HHB用户手册](https://www.yuque.com/za4k4z/oxlbxl/keyg70qggt5n3fpa) 中的命令行选项说明。
+
+
+编译命令正确执行完成后会在示例目录生成可执行文件，将该示例目录拷贝至开发板中即可运行。
+
+```shell
+scp -r ../bert_small sipeed@你的ip:~
+```
+
+#### 执行
+
+执行示例程序的前置步骤同 YOLOv5n，此处不在赘述。
+前置步骤准确无误后，即可在示例目录执行命令
+```shell
+python3 inference.py
+```
+
+#### 参考结果
+
+本示例中的参考输入来自 SQuAD 数据集，SQuAD 是一个阅读理解数据集，由一组维基百科文章提出的问题组成，其中每个问题的答案都是来自相应阅读文章或问题的一段文本。
+本示例的输入如下，文章内容描述了一次橄榄球比赛的赛况，提出的问题是谁参加了比赛。
+
+```
+[Context]:  Super Bowl 50 was an American football game to determine the champion of the National Football League (NFL) for the 2015 season. The American Football Conference (AFC) champion Denver Broncos defeated the National Football Conference (NFC) champion Carolina Panthers 24–10 to earn their third Super Bowl title. The game was played on February 7, 2016, at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California. As this was the 50th Super Bowl, the league emphasized the "golden anniversary" with various gold-themed initiatives, as well as temporarily suspending the tradition of naming each Super Bowl game with Roman numerals (under which the game would have been known as "Super Bowl L"), so that the logo could prominently feature the Arabic numerals 50.
+
+[Question]:  Which NFL team represented the AFC at Super Bowl 50?
+```
+根据阅读理解的结果，预期输出将是 Denver Broncos
+
+```shell
+(ort) root@lpi4a:/home/sipeed/bert_small_cpu# python3 inference.py
+ ********** preprocess test **********
+[Context]:  Super Bowl 50 was an American football game to determine the champion of the National Football League (N
+FL) for the 2015 season. The American Football Conference (AFC) champion Denver Broncos defeated the National Footba
+ll Conference (NFC) champion Carolina Panthers 24–10 to earn their third Super Bowl title. The game was played on Fe
+bruary 7, 2016, at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California. As this was the 50th Sup
+er Bowl, the league emphasized the "golden anniversary" with various gold-themed initiatives, as well as temporarily
+ suspending the tradition of naming each Super Bowl game with Roman numerals (under which the game would have been k
+nown as "Super Bowl L"), so that the logo could prominently feature the Arabic numerals 50.
+[Question]:  Which NFL team represented the AFC at Super Bowl 50?
+ ******* run bert *******
+Run graph execution time: 1713.15491ms, FPS=0.58
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x183d60
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x185380
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x1869a0
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x2a8610
+The max_value of output: 3.826172
+The min_value of output: -9.968750
+The mean_value of output: -8.412353
+The std_value of output: 5.128320
+ ============ top5: ===========
+ 46: 3.826172
+ 57: 3.142578
+ 39: 1.303711
+ 38: 1.179688
+ 27: 0.624512
+
+=== tensor info ===
+shape: 1 384 
+data pointer: 0x2a8300
+The max_value of output: 3.617188
+The min_value of output: -9.625000
+The mean_value of output: -7.798176
+The std_value of output: 4.820137
+ ============ top5: ===========
+ 47: 3.617188
+ 58: 3.482422
+ 32: 2.523438
+ 29: 1.541992
+ 41: 1.473633
+ ********** postprocess **********
+[Answer]:  Denver Broncos
+```
+
 ## Docker
 
 首先安装所需要的软件包
diff --git a/docs/hardware/zh/lichee/th1520/lpi4a/assets/application/usb_cam.py b/docs/hardware/zh/lichee/th1520/lpi4a/assets/application/usb_cam.py
new file mode 100644
index 0000000000..801a2c9045
--- /dev/null
+++ b/docs/hardware/zh/lichee/th1520/lpi4a/assets/application/usb_cam.py
@@ -0,0 +1,30 @@
+import cv2
+
+# 打开摄像头
+cap = cv2.VideoCapture(0)
+
+# 检查是否成功打开摄像头
+if not cap.isOpened():
+    print("无法打开摄像头")
+    exit()
+
+# 循环读取摄像头的每一帧图像
+while True:
+    # 从摄像头中获取一帧图像
+    ret, frame = cap.read()
+
+    # 检查是否读取成功
+    if not ret:
+        print("无法获取图像")
+        break
+
+    # 显示图像
+    cv2.imshow("USB Camera", frame)
+
+    # 等待按下 ESC 键退出
+    if cv2.waitKey(1) == 27:
+        break
+
+# 释放摄像头并关闭所有窗口
+cap.release()
+cv2.destroyAllWindows()
diff --git a/docs/hardware/zh/lichee/th1520/lpi4a/assets/application/yolov5n_detection_soccer_output.jpg b/docs/hardware/zh/lichee/th1520/lpi4a/assets/application/yolov5n_detection_soccer_output.jpg
new file mode 100644
index 0000000000..47b846eb8c
Binary files /dev/null and b/docs/hardware/zh/lichee/th1520/lpi4a/assets/application/yolov5n_detection_soccer_output.jpg differ