.png)
The era of NPUs that surpass GPGPUs in performance, cost-effectiveness, and power efficiency is coming!
- 태현 박
- Feb 3
- 3 min read
The three-day "AIoT International Exhibition and Conference" will be held from Wednesday, October 30 to Friday, November 1, 2024, in Hall D and the conference room on the 3rd floor of COEX. DeepX, a partner of Daewon CTS, will participate in this event at its exhibition booth (E101).
Heralding the Era of Intelligent Edge
At this exhibition booth, DeepX will demonstrate simultaneous processing of 100 CCTV channel data using an HP Z8 Fury G5 workstation equipped with the DX-H1 server product. The message this booth displays is: "There's no need to use GPUs in edge devices anymore!"
It's often believed that GPUs are necessary to run AI algorithms smoothly on edge devices or servers without performance concerns. However, this misconception stems from a lack of accurate information about NPUs.
DeepX's M.2 2280-format AI semiconductor DX-M1 and high-performance data center semiconductor DX-H1 provide the performance required by edge devices and edge servers. The reason DeepX NPU is attracting global attention is because it provides outstanding performance as well as amazing cost-effectiveness and power-to-performance ratio. Let's take a look at the DX-H1 installed in the edge server for this demonstration as an example to understand why DeepX NPU is attracting attention as a company leading the intelligent edge era.
The DX-H1 is a PCI-type card. Installing this card in a workstation or server creates a powerful edge server. Compared to GPGPU cards, the DX-H1 outperforms popular AI models like Yolo7m, Ylov8, and PIDNet. Based solely on performance, this result raises the question of whether GPGPUs are truly necessary in edge environments. Considering the price and cost-effectiveness, the conclusion is that NPUs are ideal for edge environments.
Assuming 100 channels are processed with a single card, the cost of installing the DX-H1 in a server is approximately 10 times lower than that of a GPGPU. In terms of power efficiency, the DX-H1 consumes less power than the GPGPU. Comparing the DX-H1 and the GPGPU based on Resnet-50, the FPS/W difference is approximately 20 times. This power-to-weight ratio also contributes to reducing carbon emissions. According to information disclosed by DeepX, using a single DX-H1 can reduce carbon emissions by 9,824 tons per year.
The meeting of Daewon CTS' AI infrastructure and NPU technology.
A must-see demo at the DeepX exhibition booth is a control screen that processes 100 channels of data simultaneously using a workstation equipped with two DX-H1s capable of 200 TOPS performance. The workstation deploys the YOLO v7 AI model, which performs AI-based analysis of the video being recorded by 100 channels in real time. This system has been optimized by Daewon CTS. Daewon CTS optimizes the server and workstation configuration by considering the purpose and use of the control, as well as the size and performance requirements of the model for intelligent control. For reference, Daewon CTS is supporting the design and construction of edge servers equipped with NPUs in collaboration with DeepX and K2S, and has recently expanded its support to SuperMicro, Dell, and HP environments.
Deploying edge servers or workstations configured by Daewon CTS on-site eliminates the need to connect to servers operating in public or private clouds. Deploying servers or workstations equipped with the DX-H1 card in the control center enables seamless AI-based intelligent analysis. So, how can the DeepX demo be utilized in the real world? Consider the following application areas.
Intrusion detection: Detects and alerts unauthorized entry into a specific area. Primarily used for security purposes.
Fire detection: Provides immediate warnings when signs of fire are detected, enabling early fire prevention. This is useful in high-risk locations such as warehouses and factories.
Access control: Controls access to secure areas by automatically verifying entry authorization. This contributes to increased security by preventing unauthorized entry.
Pothole & Crack Detection: Helps maintain road safety by detecting potholes and cracks on roads. Effective for traffic accident prevention and maintenance.
Hand detection: Detect human hand movements to enhance safety on the job site or recognize specific movements to perform automated tasks.
Fall detection: Detects falls and collapses, enabling rapid response. This is useful in hospitals and nursing homes.
Safety helmet detection: Helps ensure compliance with safety regulations by verifying whether workers are wearing hard hats on the job site.
Railroad and obstacle detection: Detects rail and surrounding obstacles to prevent train accidents and enhance railway safety.
People counting: Counting the number of people in a specific area can be used for crowd analysis or security enhancement.
Meanwhile, Daewon CTS is building and operating a demo system that recognizes objects using AI using the DX-M1 in the form of M.2 2280 and Raspberry Pi.
Considering the introduction and operation of edge devices or servers equipped with NPUs? We recommend visiting the DeepX booth at the AIoT International Exhibition and Conference or the Dewon CTS headquarters.
Comments