Useful for selecting devices in the early stages of development! A thorough explanation of the differences between the Qualcomm Dev-Kit designed specifically for edge AI and general-purpose kits

Performance requirements for edge AI

First, let's take camera applications as an example to explain why edge AI is attracting attention. For example, in-vehicle cameras and automatic inspection equipment in factories need to use AI to immediately analyze images captured by the camera and issue an alarm if an abnormality or danger is detected. For these applications, it is too slow to send data to the cloud and wait for the results, so edge AI is required, with the camera device itself able to perform image recognition processing.

Image of Edge AI camera application

The key factors in developing edge AI are real-time performance, the accuracy of AI inference, and ease of maintenance so that it can be used stably for a long period of time. For example, real-time performance means that a result can be produced within just a few milliseconds after a camera detects a person or object. If this is slow, accidents or defective products can be overlooked.
The accuracy of AI inference is also important. AI inference is the process of using the knowledge that AI has learned to judge new data. For example, in object recognition, which distinguishes people and objects from images, false positives or oversights directly affect the safety and quality of the workplace. The accuracy of AI models also varies greatly depending on the performance of the hardware used and the type of AI accelerator used.
Furthermore, when considering mass production and long-term operation, it is important that the hardware is durable and has a stable supply. Recently, there has been a demand for a system called OTA (Over-the-Air) that can automatically update software and AI models via the internet. This makes it possible to fix defects and add functions without going to the site.

Features of the general-purpose Raspberry Pi kit

Raspberry Pi 5 (Source: Raspberry Pi)

The Raspberry Pi 5 is a small, palm-sized computer used for a wide range of purposes, including education, prototyping, and IoT. The latest Raspberry Pi 5 boasts two to three times the processing power of the previous generation, and is capable of 4K video output and PCIe expansion. Its affordable price and easy availability, along with users around the world, make it easy to find information online when you need it, a major attraction.
However, the Raspberry Pi alone is not good at advanced AI inference processing. Therefore, to achieve edge AI processing such as image recognition with the Raspberry Pi, the IMX500 CAMERA MODULE with an AI accelerator is added. The IMX500 uses AI to analyze images taken with the camera on the spot and send only the necessary information to the Raspberry Pi. This makes it easy to run object recognition AI on the Raspberry Pi.

Appearance of Raspberry Pi5 (left) + IMX500 (right)

However, because the Raspberry Pi was originally developed for educational and prototyping purposes, some caution is required when using it for industrial purposes or mass production. For example, it may be difficult to ensure a stable supply over the long term, and it may not have the robustness or security features required in industrial settings.

Features of the Qualcomm Dev-Kit, a device designed specifically for AI

Qualcomm Dragonwing RB3 Gen 2 Development Kit (Source: Qualcomm)

The Qualcomm Dev-Kit (RB3 Gen2) is a development kit designed for industrial and IoT devices. Its core processor, the QCS6490, is equipped with eight CPU cores (Kryo 670), a high-performance GPU (Adreno 643), and an NPU (6th-generation AI engine) specifically designed for AI. The NPU is a dedicated circuit for efficiently and quickly performing AI calculations, and the RB3 Gen2 achieves a maximum of 12 TOPS (12 trillion calculations per second).

QCS6490 Simplified Block Diagram

The advantage of such specially designed devices is that they can handle computationally intensive processes such as AI object recognition and voice recognition in real time.
Please note that the Qualcomm Dev-kit is for evaluation purposes and cannot be used for mass production. Sales for mass production will be as chips or SOM (System On Module).

Features of Qualcomm Dev-Kit (RB3 Gen2)

Qualcomm Dev-kit (RB3 Gen2) Overview

The Qualcomm Dev-kit (RB3 Gen2) supports multiple operating systems, including Linux, Android, Ubuntu, and Windows. It also offers a wide range of interfaces, including USB, PCIe, MIPI, and Ethernet, and supports high-speed, low-latency communications via Wi-Fi 6E and Bluetooth 5.2. Furthermore, integration with Qualcomm AI Hub makes it easy to develop, train, and deploy AI models, facilitating a smooth transition from prototype to mass-production design.
Use cases include robotics, industrial automation, smart security, AI vision, and smart devices. For example, it is ideal for the high-precision, low-latency AI processing required in industrial settings, such as factory anomaly detection and safety management, autonomous control of logistics robots, and real-time video analysis from surveillance cameras. In addition, its long-term supply guarantee, over-the-air software updates, and comprehensive security features mean it can be used safely in both commercial and industrial applications.

Qualcomm Dev-kit: Vision Kit (Source: Qualcomm)

Vision Kit Features

camera

• Main Camera (IMX577 D-PHY 12MP): Used as the primary input for AI vision and image recognition
• Tracking Camera (OV9282 D-PHY 1MP): Supports motion and position detection and target tracking

Memory (RAM)

• 6 GB LPDDR4x

Storage (onboard)

• 128GB UFS Flash

Storage (external)

• 1x MicroSD Card Slot, PCIe expansion for NVMe

USB

• 1x USB 3.0 Type-C, 2x USB 3.0 Type-A, 1x USB 3.0 on high-speed expansion
• 1x USB 2.0 with OTG

Audio

• 4x DMIC, 2x digital audio amplifiers
• I2S/Soundwire/DMIC expansion on low-speed connectors

Sensors

• IMU (ICM-42688)
• Pressure sensor (ICP-10111)
• Mag sensor/compass (AK09915)

Qualcomm Dev-kit (RB3 Gen 2) Block Diagram (Source: Qualcomm)

Different object recognition performances depending on AI model

Next, we'll look at the difference in performance when running camera-based object recognition AI using the general-purpose Raspberry Pi 5 + IMX500 kit and the AI-specific Qualcomm Dev-Kit (RB3 Gen2). Compared to the general-purpose kit, the AI-specific kit can run AI models that require more complex processing. We used a lightweight AI model called Mobilenet SSD with the general-purpose Raspberry Pi 5 + IMX500 kit, and a high-precision AI model called YOLOX with the AI-specific Qualcomm Dev-Kit.
The main difference between the two kits is how efficiently they can process AI. While the Raspberry Pi 5 relies on the AI-enabled camera IMX500, it is limited in its ability to run complex AI models and perform high-precision processing. On the other hand, the Qualcomm Dev-Kit is equipped with a powerful NPU, allowing it to process more complex and high-precision AI models in real time.

What are Mobilenet SSD and YOLOX? Explaining the differences between object recognition AI models

When assessing the capabilities of edge AI, the type of AI model used is a very important point. Here, we will provide a beginner-friendly explanation of the two object recognition AI models used in this comparison, Mobilenet SSD and YOLOX.

Mobilenet SSD

Mobilenet SSD is a lightweight object recognition AI model designed to run easily on devices with limited computing resources, such as smartphones and small computers.
Mobilenet is a type of neural network packed with ingenuity for efficiently processing images by breaking them down into individual features. It combines this with a mechanism called SSD (Single Shot MultiBox Detector), which can find multiple objects in an image at once. The greatest features of this model are its fast processing speed and low power consumption. Even small devices like the Raspberry Pi can detect people and objects in real time. However, because the network structure is simple to reduce the amount of calculations, the accuracy of detecting complex backgrounds and small objects, or distinguishing between many types of objects simultaneously, may be slightly inferior to more powerful models.

Few objects can be recognized in a single image

Object detection image using Mobilenet SSD

YOLOX

YOLOX is a highly accurate and fast object recognition AI model that has emerged in recent years. YOLO stands for "You Only Look Once," and its unique feature is that it analyzes the entire image at once and determines what is in it. YOLOX is an evolved version of this model, offering even greater accuracy and flexibility than the previous YOLO series.
YOLOX can accurately detect objects even in complex images or scenes with many objects in them. For example, it is suitable for applications such as recognizing multiple parts simultaneously on a factory line or distinguishing between people, cars, and signs in a traffic surveillance camera. However, it requires more computational power than Mobilenet SSD, and requires more powerful hardware to run (such as a Qualcomm Dev-Kit with an NPU).

Many objects can be recognized in a single image

YOLOX object detection image

MobilenetSSD vs YOLOX: Which Should You Choose?

To summarize briefly, Mobilenet SSD is a model that is "lightweight and fast, but has limitations in accuracy and the types of objects it can handle," while YOLOX is a model that "can detect more objects with high accuracy, but requires powerful hardware to operate." Mobilenet SSD is suitable for prototyping, educational purposes, and simple applications where power consumption and cost need to be kept low. On the other hand, high-performance models like YOLOX are suitable for industrial applications where high accuracy and versatility are required.
As you can see, the optimal choice of AI model will depend on the location and purpose of use. When considering the combination of hardware and AI model, be sure to keep this difference in mind.

Running the object recognition demo with Raspberry Pi5 + IMX500

This article explains the steps to run the Mobilenet SSD object recognition demo using Raspberry Pi5 + IMX500.

Equipment used

• Raspberry Pi 5 (4GB)
• Raspberry Pi AI Camera (IMX500)
• microSD card (with Raspberry Pi OS 64bit installed)
• Internet connection (Wi-Fi or wired)
• POWER SUPPLIES adapter (5V/5A)
• Monitor and Micro HDMI CABLES
• USB mouse/USB KEY BOARD

Steps

1) Replace CABLES on the IMX500 module. Remove the FPC CABLES that came with the module and replace it with the tapered CABLES included in the IMX500 package.

IMX500 CABLES for Raspberry Pi 5 (MIPI-CSI)

2) Connect the IMX500 CABLES to the CSI camera CONNECTORS on the Raspberry Pi.

CSI CONNECTORS location on Raspberry Pi5

3) Execute the following command in Terminal on the Raspberry Pi 5 to install the IMX500 firmware. (Wi-Fi settings on the Raspberry Pi 5 must be completed beforehand.)

sudo apt update && sudo apt full-upgrade
    sudo apt install imx500-all
    

4) After the installation is complete, restart the Raspberry Pi 5 with the following command.

sudo reboot

5) Demonstrate the object recognition model using the Picamera2 library.
Picamera2 is a Python library for using the camera functions on Raspberry Pi. Using this library, you can easily acquire camera images, take photos and videos, and display previews from a Python program.
First, install picamera2 by running the following command in Terminal. The installation will take a few minutes.

sudo apt install python3-opencv python3-munkres
    git clone https://github.com/raspberrypi/picamera2.git

6) After installing Picamera2, go to the picamera2/examples/imx500 folder.

7) Run rpicam hello with the following command to run the object recognition demo using Mobilenet SSD.

rpicam-hello -t 0s --post-process-file /usr/share/rpi-camera-
    assets/imx500_mobilenet_ssd.json --viewfinder-width 1920 --viewfinder-
    height 1080 --framerate 30

Object recognition example using Raspberry Pi5 + IMX500

This completes the setup of Raspberry Pi5 + IMX500.

Running the object recognition demo on the Qualcomm Dev-kit (RB3 Gen2)

This article explains how to run the YOLOX object recognition demo using the Qualcomm Dev-kit (RB3 Gen2).

Equipment used

• RB3 Gen2 Vision Kit (Qualcomm Linux OS pre-installed)
• Internet connection (Wi-Fi or wired)
• POWER SUPPLIES adapter (12V/3A)
• Monitor and HDMI CABLES
• USB mouse/USB KEY BOARD

Steps

1) Connect AC ADAPTERS to the +12VDC CONNECTORS.

+12VDC POWER SUPPLIES CONNECTORS location

2) Connect the mouse and KEY BOARD to the USB ports.

3) Press and hold the Power button for 3-4 seconds to start the system.

RB3 Gen2 POWER SUPPLIES and

4) Connect the HDMI port on the RB3 Gen2 to the display using an HDMI CABLES.

HDMI port location

5) After the system starts up, make sure that the Q mark appears in the upper left corner of the home screen displayed on the display.

Q mark

6) Click the Q mark in the upper left corner of the home screen to display the Experience Application Demo menu. From the Wi-Fi tab, configure Wi-Fi settings (SSID and password) and confirm that the Wi-Fi status is Online.

7) Select Object detection in the Applications settings and press the Start button.

Experience Application Demo menu screen

8) Click Start and the AI model will be installed automatically. There is no need to run an installation command in Terminal like with Raspberry Pi. Object recognition will begin after a few minutes.

Object recognition example with RB3 Gen2

This completes the setup of the Qualcomm Dev-kit.

Actual demo results of a general-purpose kit and an AI-specific Qualcomm Dev-kit

We compared the results of evaluating object recognition AI using the general-purpose Raspberry Pi 5 + IMX500 kit and the Qualcomm Dev-kit, which is specifically designed for AI. Both are pre-trained models, and real-time object recognition at 30 FPS was possible simply by installing and running them.
The object recognition results for both systems varied greatly depending on the performance of the AI model used. The actual object recognition results are as follows:

Comparison of object recognition results between Qualcomm Dev-kit and Raspberry Pi5 + IMX500

The Qualcomm Dev-kit with YOLOX installed was able to accurately recognize multiple objects on a single screen, including a plastic bottle, a person, a chair, and a table.
On the other hand, the Raspberry Pi 5 + IMX500 with Mobilenet SSD installed could only accurately identify people. Furthermore, the Mobilenet SSD misidentified a plastic bottle as a cup, and briefly recognized a chair, but then misidentified it as a suitcase.
These results confirm that the object recognition performance of a camera varies greatly depending on the performance of the AI model used. These results demonstrate the superiority of the Qualcomm Dev-kit, which is specifically designed for AI and can run advanced AI models like YOLOX.

System load when running YOLOX on NPU

For reference, let's check the system load when YOLOX is running. The system load displayed using the top command is shown below. The YOLOX task is PID 1862.
The actual memory capacity (RES) consumed to run YOLOX was 336,544KB (336.5MB). Since the total memory capacity was 5,377.3MB, YOLOX only consumed 6.1% of the memory. We confirmed that the Qualcomm Dev-kit, which is equipped with an NPU, had sufficient system memory resources to run YOLOX.

YOLOX system load displayed with the TOP command

Key points for selecting hardware in the early stages of edge AI development

We have briefly summarized the key points to consider when selecting hardware in the early stages of edge AI development.
For prototyping or educational purposes, the Raspberry Pi is a strong choice as it is readily available and there is a wealth of information available. It is ideal for learning how AI works and creating simple prototypes.
On the other hand, when considering actual product development, mass production, and long-term operation, dedicated edge AI devices like Qualcomm products are more suitable. They fully meet the requirements of the business world, including high-performance NPUs that speed up AI inference and improve recognition accuracy, over-the-air software updates, guaranteed long-term supply, robust design, and security features.
In the early stages of edge AI development, it is important to carefully select devices based on this information.

Introducing Qualcomm's lineup of edge AI products

Qualcomm has a flexible portfolio of edge AI products to suit a variety of use cases. The QCS6490 used in the Qualcomm Dev-kit used here is the high-performance HIGH model.

Qualcomm's Edge AI Product Lineup

We have compiled Qualcomm's product lineup for edge AI. Please use this as a reference when selecting a device.

• Dragonwing Series: An edge AI-enabled SoC developed for industrial and commercial IoT. It combines Qualcomm's long-standing connectivity technology and advanced AI functions with low power consumption.
• Dragonwing Q8 Series: This system delivers low-power, high-performance edge AI processing. Equipped with Wi-Fi 7 connectivity and an optimized AI architecture, it supports the development of a wide range of IoT applications, including autonomous mobile robots, industrial drones, and cameras.
• Dragonwing Q6 Series: Combines high-performance computing, edge AI processing, connectivity, and power efficiency. The QCS5430 and QCS6490 are pin-compatible and feature up to 8-core CPUs and powerful AI engines (NPU and DSP). Ideal for applications requiring high computing power, such as robotics and smart vision.
• Dragonwing Q5 Series: A power-efficient SoC with AI capabilities designed for industrial and commercial IoT devices, ideal for handhelds, TABLETS, kiosks, scanners, POS, and more.

Introducing the Dragonwing Q8 Series and Q5 Series Development Kit

The Dragonwing Q8 Development Kit has been released by Thundercomm as the C8550 Development Kit, and is a high-performance development platform for IoT equipped with the QCS8550 processor.

Thundercomm C8550 (QCS8550) Development Kit (Source: Thundercomm)

The Dragonwing Q5 Development Kit is an RB3 Gen2 Lite Development Kit equipped with the QCS5430 processor. It supports multiple operating systems including Linux, Android, Ubuntu, and Windows, and offers a wide range of connectivity options, including USB, Ethernet, camera/display port, GPIO for various protocols, Wi-Fi 6E, and Bluetooth 5.2.

RB3 Gen2 Lite Development Kit (Source: Qualcomm)

Summary

To successfully introduce edge AI, it is important to select the optimal hardware to suit the application and site requirements. While a general-purpose kit such as Raspberry Pi + IMX500 is convenient for prototyping and training, for mass production and industrial use, a device designed specifically for AI, such as a Qualcomm product, is an option, as it combines real-time performance, high precision, scalability, and commercial quality.

Related Info

• Technology Column: Qualcomm Processors Core GPU and DSP (Hexagon)