Sipeed MAix: AI at the edge
AI is pervasive today, from consumer to enterprise applications. With the explosive growth of connected devices, combined with a demand for privacy/confidentiality, low latency, and bandwidth constraints, AI models trained in the cloud increasingly need to be run at the edge.
MAIX is Sipeed’s purpose-built module designed to run AI at the edge, we called it AIoT. It delivers high performance in a small physical and power footprint, enabling the deployment of high-accuracy AI at the edge, and the competitive price makes it possible to embed to any IoT devices. As you see, Sipeed MAIX is quite like Google edge TPU, but it acts as a master controller, not an accelerator like edge TPU, so it is more low cost and low power than AP+edge TPU solution.
MAix’s Advantage and Usage Scenarios:
- MAIX is not only hardware but also provide an end-to-end, hardware + software infrastructure for facilitating the deployment of customers’ AI-based solutions.
- Thanks to its performance, small footprint, low power, and low cost, MAIX enables the broad deployment of high-quality AI at the edge.
- MAIX isn’t just a hardware solution, it combines custom hardware, open software, and state-of-the-art AI algorithms to provide high-quality, easy to deploy AI solutions for the edge.
- MAIX can be used for a growing number of industrial use-cases such as predictive maintenance, anomaly detection, machine vision, robotics, voice recognition, and many more. It can be used in manufacturing, on-premise, healthcare, retail, smart spaces, transportation, etc.
- In hardware, MAIX have powerful KPU K210 inside, it offers many exciting features:
- 1st competitive RISC-V chip, also 1st competitive AI chip, newly released in Sep. 2018
- 28nm process, dual-core RISC-V 64bit IMAFDC, on-chip huge 8MB high-speed SRAM (not for XMR :D), 400MHz frequency (able to 800MHz)
- KPU (Neural Network Processor) inside, 64 KPU which is 576bit width, supports convolution kernels, any form of the activation function. It offers 0.25TOPS@0.3W,400MHz, when overclocking to 800MHz, it offers 0.5TOPS. It means you can do object recognition 60fps@VGA
- APU (Audio Processor) inside, support 8mics, up to 192KHz sample rate, hardcore FFT unit inside, easy to make a Mic Array (MAIX offer it too)
- Flexible FPIOA (Field Programmable IO Array), you can map 255 functions to all 48 GPIOs on the chip
- DVP camera and MCU LCD interface, you can connect an DVP camera, run your algorithm, and display on LCD
- Many other accelerators and peripherals: AES Accelerator, SHA256 Accelerator, FFT Accelerator (not APU’s one), OTP, UART, WDT, IIC, SPI, I2S, TIMER, RTC, PWM, etc.
Inherit the advantage of K210’s small footprint, Sipeed MAIX-I module, or called M1, integrate K210, 3-channel DC-DC power, 8MB/16MB/128MB Flash (M1w module add wifi chip esp8285 on it) into Square Inch Module. All usable IO breaks out as 1.27mm(50mil) pins, and the pin’s voltage is selectable from 3.3V and 1.8V.
MAIX’s development board (M1 dock & M1w dock, w means WiFi version)
Firstly, We make a prototype development board for M1, called M1 dock or Dan Dock, it is simple, small, cheap, but all functions include.
As many DIYer want to build their own work with a breadboard, Sipeed newly provide breadboard-friendly board for you, it called MAix BiT
- It is twice of M1 size, 1×2 inch size, breadboard-friendly, and also SMT-able,
- It integrates USB2UART chip, auto-download circuit, RGB LED, DVP Camera FPC connector(support small FPC camera and standard M12 camera), MCU LCD FPC connector(support our 2.4-inch QVGA LCD), TF card slot.
- MAix BiT is able to adjust core voltage! you can adjust from 0.8V~1.2V, overclock to 800MHz!
And the bigger and better MAIX’s development board
- It is 88x60mm, all pins out, with standard M12 lens DVP camera, and the Camera can be flipped from front to rear!
- It has onboard JTAG&UART based on STM32F103C8, so you can debug M1 without extra Jlink.
- It has lithium battery manager chip with power path management function, you can use the board with lithium battery and USB power without conflict~
- It has I2S Mic, Speaker, RGB LED, Mic array connector, thumbwheel, TF card slot, and so on.
- This suite includes a 2.8 inch LCD too and has a simple case for it.
MAIX supports original standalone SDK, FreeRTOS SDK base on C/C++.
It supports FPIOA, GPIO, TIMER, PWM, Flash, OV2640, LCD, etc. And it has modem, vi, SPIFFS on it, you can edit python directly or Sz/Rz file to board.
MAix’s Deep learning
MAIX supports a fixed-point model that the mainstream training framework trains, according to specific restriction rules, and have a model compiler to compile models to its own model format.
It supports tiny-Yolo, mobile net-v1, and, TensorFlow Lite! Many TensorFlow Lite model can be compiled and run on MAIX! And We will soon release the model shop, you can trade your model on it.
- Getting Started
- MAIX_ Tools
- Libraries – MAIX
- MicroPython Introduction
- Kendryte K210 FreeRTOS SDK V0.5.0
- Kendryte K210 Standalone SDK V0.5.2
- Kendryte K210 datasheet English ver.V0.1.5
- Kendryte Standalone SDK Programming Guide-EN-V0.3.0
- Kendryte FreeRTOS SDK Programming Guide-EN-V0.1.0
- Kendryte OpenOCD for win32 V0.1.3
- Kendryte OpenOCD for Ubuntu x86_64 V0.1.3
- RISC-V 64bit toolchain for Kendryte K210_win32 V8.2.0
- RISC-V 64bit toolchain for Kendryte K210_ubuntu_amd64 V8.2.0
- K-Flash V0.3.0
- Kendryte K210 Model Download Guide V0.1.0
- Kendryte K210 Face Detection Demo V0.1.0
- Cmake installation
- Windows CPP Build tools
- Supports the fixed-point model that the mainstream training framework trains according to specific restriction rules.
- There is no direct limit on the number of network layers, and each layer of convolutional neural network parameters can be configured separately, including the number of input and output channels, and the input and output line width and column height.
- Up to 8 channels of audio input data, ie 4 stereo channels
- Simultaneous scanning pre-processing and beamforming for sound sources in up to 16 directions
- 16-bit wide internal audio signal processing
- Support for 12-bit, 16-bit, 24-bit, and 32-bit input data widths
- Up to 192kHz sample rate
- Built-in FFT unit supports 512-point FFT of audio data
- Uses system DMAC to store output data in system memory
- The maximum supported neural network parameter size for real-time work is 5MiB to 5.9MiB
1 x Sipeed MAIX-I module w/o WiFi