Sipeed Maix-III AXera-Pi Linux AI Core Board

Sipeed Maix-III AXera-Pi Linux AI Development Board

In the past few years, many Linux AI chips come out Raspberry Pi and Jetson Nano these boards are really popular by their good performance, but now MAIX-III AXera-Pi comes out, equipped with the high-performance 3.6TOPs@INT8 chip ax620a, which will be your good choice and provide you a really good experience. 

We design a core module for this chip, which is convenient if you want to design the ext-board for yourself.

1. Introduction

In WORLD ARTIFICIAL INTELLIGENCE CONFERENCE 2022, this AI board is shown the first tima, like what's in the following picture.

And finally we launched this AI development board equipped with AX620A in 20221001. Its official name is MAIX-III AXera-Pi, while we like to call it M3AXPI in the following content.

So, why is this AI development board good?

• Need documents? Follow our guides, to start journey from embedded Linux development to AI model deployment.
• High power consumption? Just use one usb cable connecting with USB3.0 port to boot this device, no need of other power supply.
• Difficult to develop? Compile your own application on this board, or train your own AI model online and deploy it on board.
• Awful camera image? This chip is aimed for AI ISP night vision enhancement, we have adjust the camera to best state.
• Too expensive? This board sold about 100$, chiper than any other similar board.

2. Special column Documents

These documents are from Chinese club, read these via browser translator is a good idea.

3. Camera effect

3.1. Night vision enhancement

Left picture is original, right picture is enhanced

3.2. Exhibition darkroom

• The left screen shows the enhanced picture and the righr screen is the original one.

3.3. Actual AI effect

4. AI Ecology

4.1. Support many Operators

Screenshots of part operators. Click me to see whole operator support list

4.2. Many archived models

Before receiving this board, visit MaixHub to see your interested model, choose Platform axera-pi at the Lower left corner to find model can be deployed on AX-PI

Apart from these, many models are uploaded to github AXERA-TECH/ax-samples, and we also put sme models in sustem image, just burn the system image into SD card and boot it on board to experience these models. Up to now the supported models are as follows:

• Objects Classification
  • MobileNetv1
  • MobileNetv2
  • MobileOne-s0
  • ResNet18
  • ResNet50
  • Others......
    
    
    
• Objects Detection
  • PP-YOLOv3
  • YOLOv3
  • YOLOv3-Tiny
  • YOLOv4
  • YOLOv4-Tiny
  • YOLOv4-Tiny-3l
  • YOLOv5s
  • YOLOv5-Lite(original model)
  • YOLOv7-Tiny
  • YOLOv8s
  • YOLOX-S
  • YOLO-Fastest-XL
  • NanoDet
    
    
    
• Human Body Detection
  • YOLO-Fastest-Body

• Human Face Detection
  • scrfd
  • YOLOv5-Face(original model)

• Human Face Segmentation
  • Face_Parsing

• Obstacles Detection
  • Robot-Obstacle-Detect

• Object detection of Uav view
  • YOLOv5s_visdrone

• Monocular 3D Vehicle Detection
  • Monodlex

• Human body Key points
  • HRNet
  • AX-POSE-PPL
  • HandPose

• Human body Segmentation
  • PP-HumanSeg

• Semantics Segmentation
  • PP-Seg

• Classic CV
  • CropResize

• Pipeline example
  • NV12 -> CropResize -> NN(Classification)

Visit Deploy models to AX-Pi (Maix-III(M3) series) board to deploy models to AX-Pi.

4.3. Tranning your models

• [m3axpi] YOLOv5 train and convert model guide
• [m3axpi] YOLOv8 train and convert model guide

4.4. Tranning model online

Now MaixHub has been upgraded, with more friendly interface and more functions, here are steps to tranning model online.

4.5. Share your model

There is a model and codes or programs to load your model after you finish training, you can share these on Maixhub to others. Let's have fun and build the model Community together.

5. Cases

Updated in 20221205

6. Linux development

6.1. Run debian

We provide a linux image based on debian11, and you can boot this board from SD card after you burn the system image into SD card. We aim to make it easiler for user to use. Besides, it's really conveient to program AI development with Python API.

1. Burning image: WIKI AXera-Pi Burn image

2. Basic usage：WIKI AXera-Pi guide

3. Program with Python：ax_pipeline_api

4. Development Note：[maixpy3 axpi] How to port debian11 to ax620a (中文)

5. Development Note：[maixpy3 axpi] Edit and pack the compiled debian image (中文)

6.2. BSP SDK

We provide source code, api documents, sdk usages, visit sdk usage to get more.

6.2.1. ax-sample

Here are typical AI models provided by Axera for development, evaluation and verification, those who are good at AI development or want to test these models can have a try, and these are not involved with peripherals.

1. examples
2. PP-YOLOv3
3. PP-HumanSeg
4. PP-Seg

6.2.2. libmaix

A unified embedded development environment come by SIPEED, includes camera, screen, vision, image processing and pipelines-related deployment examples, is suitable for those who want to learn embedded Linux.

1. Test screen display
2. Screen displays camera content
3. camera + screen + yolov5 model

6.2.3. axpi_bsp_sdk

BSP development package for business, provided by Axera. It's the original development package for this chip, containing components like uboot、linux、msp 、sample、rtsp、ipcdemo, and this is progressively open source, you can get these codes for commercial evaluation, such as ipcdemo, but normally these codes are complex and highly coupled, those who are experienced enough and want to use this board for business cooperation can try this.

1. Development documents docs
2. Platform application ipcdmeo
3. Peripheral Verificationsample

6.2.4. ax-pipeline

AX-Pipeline is provided by Axera. With M3AXPI, this project is used for pepole to know the usages of ISP, image processing, NPU, codec and display, which is convenient for developers to quickly evaluate and develop their own multimedia applications.

1. Quick compilation Simple cross-platform compilation via cmake.
2. How to replace into my own trained yolov5 model
3. How to deploy my other models
4. How to adjust the image orientation
5. ModelZoo descriptions for models which are supported or to be supported

7. Hardware

7.1. Hardware function diagram

7.2. Core module parameters

7.3. Ext-board parameters

Hardware summary informations

• Datasheet
• Schematic
• Bit map
• 3D file
• Dimension
• GC4653 Camera DS

7.4. Camera comparison

• Low-end camera gc4653 400w basic experience

• High-end camera os04a10 400w enhanced night vision edition

For more information about Sipeed Maix-III AXera-Pi Linux AI Development Board, please contact us