In 2025, Artificial Intelligence (AI) and Edge Computing are about to reach a groundbreaking milestone. DeepX has launched a revolutionary product — the DeepX DX-M1 AI Inference Accelerator Card. It adopts the PCIe Gen3 M.2 M-Key interface and offers exceptional computing performance of up to 25 TOPS, along with features such as high accuracy, low power consumption, and low-temperature operation.
By combining the Orange Pi 5 Plus (Rockchip RK3588) with the DeepX AI accelerator card, this solution offers a highly cost-effective foundation, providing industrial computers with plug-and-play convenience, further emphasizing its practicality. The design concept is “Platform Upgrade: Embracing the AI Intelligent Era.”It helps developers easily step into the field of artificial intelligence, opening a new chapter for AI applications.
Figure 1. Diagram illustrating the advantages of the multi-object detection solution based on the DeepX AI accelerator card
At the same time, it provides abundant software resources and comprehensive support for mainstream deep learning frameworks (such as TensorFlow, PyTorch, and ONNX), enabling both beginners and professionals to quickly master operational skills and easily complete AI model deployment, allowing more people to get involved in artificial intelligence application development and making innovation more accessible.
Among them, Edge Computing and Neural Processing Units (NPUs) will be the key drivers of this wave, allowing smart devices to be more seamlessly integrated into people's daily lives.
With the continued advancement of AI technology, more innovative applications are gradually being implemented, completely reshaping our daily lives in the following areas:
- Smart Surveillance: Enhancing urban safety and monitoring systems through real-time target detection, behavior analysis, and intrusion alerts.
- Smart Retail: Improving the shopping experience and optimizing sales strategies through customer behavior analysis and intelligent shelf management.
- Medical Imaging Analysis: Assisting doctors in disease diagnosis (such as tumor detection) using AI to improve medical accuracy.
- Industrial Quality Inspection: Automating defect detection to enhance quality control and increase production efficiency in manufacturing.
- Advanced Driver Assistance Systems (ADAS): Applying AI vision technology to analyze road conditions, detect pedestrians, and other factors to enhance driving safety.
- Agricultural Monitoring: Monitoring crop health, detecting pests and diseases, optimizing farm management, and improving yield and quality.
- Autonomous Vehicles: Using sensors and AI algorithms to navigate and avoid obstacles, suitable for scenarios such as logistics and inspections.
- Robotic Arms: Combined with AI and vision systems to accurately perform tasks such as assembly and welding, thereby improving production efficiency and product quality.
- Automated Food Delivery Systems: Delivery platforms like Uber Eats are deploying delivery robots developed by Serve Robotics in multiple cities across the U.S. These robots are equipped with AI technology, allowing them to autonomously navigate to customer locations and provide efficient food delivery services.
- Drone Image Analysis: AI-powered vision systems enable drones to detect terrain objects, assess disasters, and inspect infrastructure, offering efficient data collection and analysis.
Figure 2: Diagram of AI Application Directions, Source: Generated by ChatGPT
Additionally, the exceptional computational performance of DeepX can be utilized to easily create multi-streamer object detection applications. Whether using a standard USB camera or a network streaming source, it can be seamlessly integrated into various smart scenarios. Since the current camera input limit of the Orange Pi 5 Plus is approximately 16 channels of 480p streaming, utilizing a more advanced NVR as the main platform can better unleash the full potential of the DX-M1. This enables even greater performance, making it ideal for applications such as merchandise theft prevention, vehicle security management, and crowd statistics in small shopping malls, parking lots, exhibition centers, and more.
Figure 3: Illustration of a Multi-Object Detection Solution Application, Source: Generated by ChatGPT
DeepX: The Trailblazer Leading the New Wave of Intelligence
In the wave of edge computing, DeepX — a South Korean startup with over 240 smart technology patents [link] — has emerged as a leader in the field. The company was honored with three Innovation Awards in Embedded Technologies and Robotics at CES 2024 .[link] , as well as major recognitions in areas such as computer integration. It was even recognized by the Consumer Technology Association (CTA) in the United States as a "must-visit company," becoming a focal point in the global market [link].
The DeepX DX-M1 AI accelerator card has arrived with impressive power, delivering a computing performance of 5 TOPS per watt. Equipped with robust edge computing capabilities (25 TOPS) and the exclusive IQ8™ (Intelligent Quantization Integer 8) technology, it achieves GPU-level accuracy while maintaining the extreme efficiency of INT8 inference. This enables developers to enjoy FP32-level precision with INT8 performance, delivering unparalleled AI accuracy [link].
With 4GB of dedicated DDR memory onboard, the module operates independently without drawing from the host system's resources—dramatically simplifying system integration. Its exceptional performance has made it an essential solution in areas such as smart surveillance, healthcare, and manufacturing, standing out as the ultimate choice for edge AI applications.
Figure 4: Schematic Diagram of DeepX AI Chip Specifications
Continuous software optimization and updates: Delivering the best user experience
DeepX not only delivers outstanding hardware performance but also offers a comprehensive and developer-friendly software ecosystem that provides full-spectrum support. This ecosystem includes detailed Quick Start Guides, a powerful Software Development Kit(SDK), an extensive Model Zoo, and a wide range of sample applications. These resources help developers efficiently integrate and optimize AI model deployment, significantly reducing development time while enhancing application performance—making AI innovation more accessible than ever.
The software stack is composed of three core components: the Quantizer, the DX-COM Compiler, and the DX-RT Runtime, which together power the entire DeepX AI SoC product line.Working in synergy, these components form a highly efficient AI computing platform that supports a broad spectrum of application scenarios. The following diagram illustrates the overall architecture:
Figure 5: Illustration of the DeepX Development Environment, Source: Official Document
Outstanding AI performance
Using the DeepX DX-M1 chip to run the currently popular YOLOv5s (640x640) object detection algorithm can easily achieve around 330 frames per second.
Figure 6: DeepX DX-M1 Chip Performance Data Sheet
The table below presents the accuracy analysis of the GPU and DX-M1, where green text represents Full Precision for the GPU, and blue text represents IQ8 for the DeepX NPU.
Figure 7: DeepX DX-M1 vs GPU Accuracy Comparison Table, Source: Official Document
Rich AI applications
DeepX is committed to accelerating the adoption of artificial intelligence by offering a wide range of AI examples and educational resources. These materials provide step-by-step guidance to help developers master the implementation of AI applications. Through hands-on examples, developers not only gain a deeper understanding of core AI technologies but also learn how to flexibly apply them to real-world scenarios—driving innovation and accelerating the deployment of AI solutions.
Figure 8: Illustration of DeepX in Practical Application
Video Introduction:
- AIoT Booster for AI NVR
- Yolov8 AI Accuracy Test: DEEPX DX-M1 vs GPGPU Showdown
- The coolest innovation in AI hardware! 🧊
- Embedded World 2025: Interview with DeepX CEO Lok Won Kim
Reference website:
- DeepX official website
- DeepX DX-M1 Product Brief
- DeepX develops websites
- Orange Pi 5 Plus website
- To obtain the DeepX DX-M1, please contact the WPI representative.
►场景应用图
►展示板照片
►方案方块图
►核心技术优势
1. Adopting IQ8™ (Intelligent Quantization Integer 8) technology for GPU-level accuracy: DeepX's IQ8™ technology requires dataset calibration to easily achieve the extreme efficiency of INT8 while maintaining FP32-level accuracy, delivering AI accuracy comparable to that of a GPU. 2. Does not occupy system memory: The module operates with an independent memory architecture, featuring a built-in 4GB storage capacity that does not rely on the main system's RAM resources. This effectively reduces the overall system performance burden. 3. Optimized data flow to minimize data movement: With an optimized data transfer flow design, the memory is integrated into the accelerator card, significantly reducing the number of data transfers within the main system. This results in a notable increase in processing speed and a substantial reduction in latency. 4. Cost-effective and low-power consumption solution: By pairing the main platform Orange Pi 5 Plus with the DeepX DX-M1 AI chip, the system can be seamlessly upgraded to a higher-level AI platform capable of running approximately 480 frames per second (YOLOv5s) for object detection. Additionally, the DX-M1 delivers 5 TOPS/W performance, with the entire multi-object detection solution consuming only about 14W of power. 5. A new concept for multi-channel applications: With the rapid development of edge computing technology, combining it with localized application scenarios may create an innovative and cost-effective solution. By using readily available cameras, a smart industrial host, and the DeepX M.2 accelerator card, diverse application needs can be met. Furthermore, the front-end cameras can be flexibly replaced and configured based on actual requirements, enhancing the system's adaptability and flexibility, thereby offering more possibilities for various industries.
►方案规格
1. The main platform development board is based on the RockChip RK3588 platform, featuring four Cortex-A76 processors and four Cortex-A55 processors. It is equipped with a high-performance Arm Mali-G610 GPU and a Neural Processing Unit (NPU), offering a powerful core architecture. 2. The I/O Board development board provides robust peripheral configurations, such as Gigabit Ethernet, HDMI (High-Definition Multimedia Interface), USB Type A/C 3.0 interfaces, M.2 E-Key interface, M.2 M-Key interface, and an extended 40-pin header that supports common signals like UART, I2C, SPI, and CAN. 3. The DeepX DX-M1 chip delivers powerful AI computing capabilities (25 TOPS) and adopts a PCIe Gen3 M.2 2280 M-Key interface design. It features the high-performance DX-M1 chip, achieving an impressive 5 TOPS per watt for exceptional low-power computing. Additionally, it includes 4GB of DRAM to ensure high efficiency and stability in module access. This accelerator card fully supports Linux and Windows operating systems and provides developers with abundant software resources for easy integration and application.