Introduction of Edge Devices Used in Edge AI !

In recent years, the need for edge AI has increased, and along with that, edge devices for operating AI are diversifying. As you can see just by searching on the Internet, there is an extremely wide variety of products on the market, but it is not uncommon for people to come up with the most suitable device for what purpose. Is not it?

Therefore, in this article, I would like to introduce the trend of edge devices and their simple advantages, disadvantages, and suitable uses. Single-board computers and microcomputer boards are widely used as edge devices, and each has advantages and disadvantages. Read this article to help you decide which edge device is best for your application.

single board computer
microcomputer board
in conclusion

1. Single board computer

In this section, we will introduce edge devices that mainly use OSs such as Linux and Ubuntu. It is recommended to use it not only for simple data processing, but also for AI inference, or when there is a need to complete everything to some extent with this one.

Raspberry Pi 4

merit	Rich input/output (HDMI, USB port, GPIO pin, etc.) Because there are many users, there are abundant materials both in Japan and overseas, and it is easy to find out what you do not understand Mass sales of expansion modules such as camera modules and GPS modules External calculation boards are sold, and performance can be expanded to some extent Equipped with BLE and Wi-Fi, it can be used as a PC
Demerit	Slightly expensive with peripheral equipment Device performance and processing speed may not be sufficient
Suitable for	If you want to use arbitrary Linux software instead of PC or processing that involves screen display When you’re not exactly sure what you’re going to do, but just want to start developing

Raspberry Pi zero W

merit	Very small for a device running an OS such as Ubuntu (about the size of a Frisk case) Other than input/output, it has almost the same functions as Raspberry Pi 4 etc.
Demerit	processing speed is slow Insufficient memory to use AI model
Suitable for	If you want to use Raspberry Pi 4, but it is not possible due to size constraints

Beagle Bone Series

merit	Processing speed (CPU performance, memory performance) is higher than Raspberry Pi in the same price range
Demerit	Fixed screen output resolution and poor video encoding capability Less popular than Raspberry Pi (less documentation available)
Suitable for	Alternative to Raspberry Pi with no screen output or fewer restrictions on image processing

The Beagle Bone series is a product developed with Raspberry Pi in mind. In addition to this, products such as “Banana Pi” and “Orange Pi” are sold by various companies as products that respect Raspberry Pi.
However, some products are only sold on overseas EC sites, and wireless communication may not be possible because they have not acquired the Technical Conformity (Technical Standards Conformity Certification) mark, so they are not very practical. .

On the other hand, many of these respected products improve the shortcomings of the Raspberry Pi (mainly processing speed), making it an option in very limited situations.
In any case, in order not to violate the Radio Law, it is necessary to develop products that have acquired the Technical Conformity Mark.

NVIDIA Jetson series

merit	GPU is installed and image processing is very fast It is sold by NVIDIA, a GPU developer, and has a very complete framework for deep learning. Large number of users In addition to the large Jetson Xavier, the small and inexpensive Jetson Nano is on sale.
Demerit	relatively expensive
Suitable for	Few restrictions on size or price When advanced image processing is required When you need to do advanced deep learning

Ticker Board

merit	A board with the same size as the Raspberry Pi is equipped with a GPU as standard, and is strong in image processing. Stable operation (enhanced circuit protection function)
Demerit	More expensive than Raspberry Pi Slightly harder to obtain (more corporate than individual)
Suitable for	When stable operation is required in an environment where thermal runaway or instantaneous electrical shocks are likely to occur Commercialization of Raspberry Pi projects developed through PoC and prototyping

Summary of single board computers

If the only condition is system development using applications that run on a Linux-based OS, the Raspberry Pi series should be used for the time being. The Raspberry Pi is especially suited for prototyping purposes.
However, if the performance becomes unsatisfactory during development or if you are aiming for commercialization, it is recommended to try other boards.

In particular, in the development of deep learning models, which are gaining importance in recent years, it is often not efficient to use Raspberry Pi for industrial purposes.
Therefore, it is a good idea to check the operation on Raspberry Pi and then use the latest Jeston series to make it into a product.

2. Microcomputer board

This section introduces OS-less edge devices for lower-level processing. You might be thinking, “Isn’t this a low-level choice?” But when I say level, I mean layers of processing. The microcomputer board, which specializes in low-level processing, can perform specified processing at extremely high speed and read the values of connected sensors. For example, a microcomputer board is sufficient for the purpose of collecting sensor data in a factory.

The microcomputer board is generally used by connecting it to a PC or a single-board computer with the above-mentioned OS, not by itself.

Arduino series

merit	It is also used for educational purposes and has a large number of reference materials. There are many related products, including the faster Arduino Mega, the Due and the smaller Arduino Micro, and the LilyPad Arduino, which is intended to be attached to clothes. Abundant packages for program development Abundant expansion modules such as GPS and microphone
Demerit	The performance is relatively low, and the series with high performance is slightly large
Suitable for	When high performance is not required prototyping

mbed series

merit	Smaller than the Arduino series Excellent in clock (processing speed) and memory capacity
Demerit	Since it is developed in a browser, it cannot be developed without an Internet environment (can be used offline)
Suitable for	If you need more complex processing than Arduino When a small and highly functional microcomputer board is required

ESP series (microcomputer boards equipped with ESP32)

merit	Wi-Fi and Bluetooth connection possible from the beginning The same development environment as Arduino can be used, so projects can be reused.
Demerit	Slightly high power consumption There are many relatively new products, and information may be scarce
Suitable for	If you want to use communication such as Wi-Fi or Bluetooth You can use the same development environment as Arduino, so if you want to reuse existing Arduino products

M5Stack series

merit	Equipped with a unique large display as standard on a microcomputer board Special input devices such as buttons, microphones, and IMUs (such as accelerometers) can be used Equipped with ESP32 and can be developed in the same environment as Arduino Abundant expansion modules, easy to connect, friendly for beginners
Demerit	Large for a microcomputer board
Suitable for	When processing using information entered by the user, such as a button or microphone, is required for learning programming

M5 Stick V

merit	Equipped with a small display Built-in processing unit (KPU) for deep learning Relatively small
Demerit	It is difficult to expand such as adding sensors other than the attached connector
Suitable for	When image processing is required on an inexpensive and small device

Summary of microcomputer board

Until a few years ago, when it came to microcomputer boards, Arduino and mbed were the only two choices.
However, microcomputer boards equipped with a processor called ESP32 have begun to be sold, and for beginners, the M5Stack and M5Stick series, which are all-in-one sensors and microcomputer boards, have been released. Even though it’s for beginners, it’s a product that has a lot of fans because of its ease of use and the many things it can do. In particular, M5Stick V and its derivatives are becoming standard microcomputer boards for image processing on edge devices.

3. Conclusion

Did you find an edge device that might suit your needs?
Among the products introduced in this article, we recommend Raspberry Pi, Jetson, M5 Stick V, and their derivatives for building edge AI.
If the edge device only collects various sensor data and AI inference is performed on a separate PC, “Arduino”, “mbed” and their derivatives are suitable.
However, the processing speed, power consumption, heat generation, etc. are often unknown until you actually operate it, so please actually purchase the products introduced in this article and try various things.