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.
MaixIII Core module | |
---|---|
CPU | 4 * Cortex-A7, with NEON and FPU |
NPU | 14.4Tops@int4,3.6Tops@int8 |
ISP | 4K@30fps |
Codec format | H.264, H.265 |
Video encode | 4K@30fps |
Video decode | 1080P@60fps |
Ethernet | Support RGMII / RMII Ethernet |
Display | MIPI DSI 4 Lane,up to 4K@30FPS |
DRAM | 2GB LPDDR4X 3733Mhz |
Storge | Optional 16GB EMMC (Boot from TF card by default) |
IO pins | DDR4 SODIMM 260P route all IO |
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.
Official Documents:
-
MAIX-III AXera-Pi:New generation of 3.6T visual AI board comes out
-
YOLOv8 object detection:Train and deploy model on Axera-Pi ①
-
YOLOv8 object detection:Train and deploy model on Axera-Pi ②
-
MMPose (Extra section) :Fall Detection and alert on AXera-Pi
Community Documents:
- Axera-Pi unbox experience
- HonestQiao —— unbox Axera-Pi
- Axera-Pi NPU experience
- MMYOLO object detection algorithm on AX620A
- Deploy yolov8 on Axera-Pi
- [m3axpi] YOLOv5 Train and deploy model
- [m3axpi] YOLOv8 Train and deploy model
- Prepare AX620A development enviroment on ubuntu22.04
- Train and deploy yolo5s on Axera-Pi
- [AXPI] Use RNDIS on m1/m2 MAC
- Kaldi - Real-time speech recognition on embedded device
Video reference:
- MAIX-III AXera-Pi unbox and start
- MAIX-III AXera-Pi development deployment
- MAIX-III AXera-Pi model transformer and deploy
- MAIX-III AXera-Pi video inference and push
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
- Human Face Segmentation
- Face_Parsing
- Obstacles Detection
- Robot-Obstacle-Detect
- Object detection of Uav view
- Monocular 3D Vehicle Detection
- Human body Key points
- Human body Segmentation
- Semantics Segmentation
- 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
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.
-
Burning image: WIKI AXera-Pi Burn image
-
Basic usage:WIKI AXera-Pi guide
-
Program with Python:ax_pipeline_api
-
Development Note:[maixpy3 axpi] How to port debian11 to ax620a (中文)
-
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.
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.
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.
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.
- Quick compilation Simple cross-platform compilation via cmake.
- How to replace into my own trained yolov5 model
- How to deploy my other models
- How to adjust the image orientation
- ModelZoo descriptions for models which are supported or to be supported
7. Hardware
7.1. Hardware function diagram
7.2. Core module parameters
MaixIII Core module | |
---|---|
CPU | 4 * Cortex-A7, with NEON and FPU |
NPU | 14.4Tops@int4,3.6Tops@int8 |
ISP | 4K@30fps |
Codec format | H.264, H.265 |
Video encode | 4K@30fps |
Video decode | 1080P@60fps |
Ethernet | Support RGMII / RMII Ethernet |
Display | MIPI DSI 4 Lane |
RAM | 2GB LPDDR4X 3733Mhz |
Storge | Optional 16GB EMMC (Boot from TF card by default) |
IO pins | DDR4 SODIMM 260P route all IO |
7.3. Ext-board parameters
Ext-board | |
---|---|
Camera | Maixmum supported 3 cameras (Default connecting one camera): 1 MIPI 4-LANE + 2 MIPI 2-LANE |
Display | Maixmum supported 4-LANE MIPI DSI |
SD card slot | Default boot from SD card |
Recording | 2 onboard MEMS MIC support dual MIC noise reduction |
Audio output | Support 3.5MM interface, used for stereo output or MIC input |
Network | Support Gigabit Ethernet (ETH) and 2.4GHZ WI-FI |
USB | 1 x USB2.0HS used for OTG or HOST 1 x USB-UART |
Others | 1 User key, 3 x System keys, 1 x RTC, 2 x LEDs |
Interface | Reserved CJTAG,used for connecting corresponding debugger, 2 x 12 pin headers, 4 x M2.5 screw positioning holes |
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
We have a 30-day return policy, which means you have 30 days after receiving your item to request a return.
To be eligible for a return, your item must be in the same condition that you received it, unworn or unused, with tags, and in its original packaging. You’ll also need the receipt or proof of purchase.
To start a return, you can contact us at alex@friendlyarmtr.com. If your return is accepted, we’ll send you a return shipping label, as well as instructions on how and where to send your package. Items sent back to us without first requesting a return will not be accepted.
You can always contact us for any return question at alex@friendlyarmtr.com.
Damages and issues
Please inspect your order upon reception and contact us immediately if the item is defective, damaged or if you receive the wrong item, so that we can evaluate the issue and make it right.
Exceptions / non-returnable items
Certain types of items cannot be returned, like perishable goods (such as food, flowers, or plants), custom products (such as special orders or personalized items), and personal care goods (such as beauty products). We also do not accept returns for hazardous materials, flammable liquids, or gases. Please get in touch if you have questions or concerns about your specific item.
Unfortunately, we cannot accept returns on sale items or gift cards.
Exchanges
The fastest way to ensure you get what you want is to return the item you have, and once the return is accepted, make a separate purchase for the new item.
Refunds
We will notify you once we’ve received and inspected your return, and let you know if the refund was approved or not. If approved, you’ll be automatically refunded on your original payment method. Please remember it can take some time for your bank or credit card company to process and post the refund too.
Please contact us for shipping charges and additional information.
There are no reviews for this product yet...