On-device artificial intelligence (AI) semiconductor corporations DeepX said on the 14th that it will link Ultralytics YOLO (YOLO), PaddlePaddle, and Raspberry Pi with its Neural Processing Unit (NPU). It plans to build an open physical AI ecosystem that allows developers to develop AI models and then carry them through to proof of concept (PoC) and mass production of industrial products.
Physical AI is a technology in which AI performs perception, judgment, and control on devices in the real world, such as robots, intelligent cameras, industrial equipment, and smart factories. DeepX supports connecting the AI model ecosystem with ultra-low-power NPU so that developers can apply models created in existing environments to industrial sites.
DeepX signed a strategic alliance in May with Ultralytics, which develops real-time vision AI models. Developers can run YOLO-based models on Raspberry Pi and on robots, intelligent cameras, and industrial equipment. After model training, it reduces separate conversion and optimization steps, making it easier to expand into Edge AI systems.
It is also working with the open-source Deep Learning framework PaddlePaddle. After forming a partnership in Aug. last year, DeepX applied PaddlePaddle's lightweight AI model "PP-OCR 5th generation" to its M.2 form factor AI accelerator "DX-M1." This environment can be used in areas that require real-time AI inference, such as optical character recognition (OCR), robots, drones, and Smart City.
The hardware development environment was linked with Raspberry Pi. In June, DeepX released an AI acceleration processor module for Raspberry Pi 5th generation. Developers can validate Ultralytics- or PaddlePaddle-based models on Raspberry Pi and then expand to industrial cameras, robots, edge gateways, and smart factory equipment.
DeepX plans to establish a structure in which developers build AI applications on Raspberry Pi, and corporations and research institutes verify them as PoC and then mass-produce them as industrial products. To that end, it will link AI models, developer hardware, NPU, software development kit (SDK), and industrial reference models, and expand hands-on examples and on-site verification programs for corporations.
DeepX Chief Executive Kim Nok-won said, "In the era of physical AI, a good chip alone is not enough," and added, "By connecting developers, AI models, and hardware, we will create an open execution platform that allows global developers and corporations to easily implement physical AI."