POSCO DX said on the 18th that it developed an unstructured data analysis platform for industrial sites using a domestically made artificial intelligence (AI) semiconductor.
According to POSCO DX, a Neural Processing Unit (NPU) is a semiconductor specialized for AI computations such as Deep Learning and Machine Learning. The company said it is more optimized for AI inference than a GPU, allowing infrastructure expense and power consumption to be reduced. It can also be used to implement "Edge AI," performing computations without going through remote data centers or servers by being installed directly in on-site equipment control systems.
POSCO DX integrated these advantages into its "Vision AI platform." The newly developed platform operates various video data collected from industrial sites as training data within a single environment and standardizes functions repeatedly needed to implement Vision AI services—such as AI model management, performance metrics, and application history—to provide them.
In particular, while GPUs are used for AI model training and development as before, a hybrid method was adopted that applies a domestically made NPU for real-time computation and decision-making in industrial sites. An environment was also created to use NPUs in the research and validation stages so that NPU-based inference performance and effectiveness can be confirmed in advance before applying them to actual industrial sites. Through this, the accuracy and processing speed of AI models, power efficiency, and operating expense can be verified in advance before on-site application, and the optimal AI models and NPU application plans suited to site characteristics can be derived.
Previously, POSCO DX pursued AI localization in various areas—such as fire monitoring at industrial sites, worker safety monitoring, and checking product loading status in logistics environments—based on cooperation with domestic NPU developers including DeepX and Mobilint. As a result, infrastructure build-out and operation expense was cut by about 50% compared with equivalent GPU inference performance, and power consumption fell by about 90%. POSCO DX plans to gradually convert sites where Vision AI technology is applied to a platform based on domestically made NPUs and expand the scope of application to a variety of industrial sites.
A POSCO DX official said, "The NPU-based Vision AI platform will serve as a foundation to implement AI optimized for industrial sites more efficiently," adding, "We will work to systematize the use of domestically made NPUs and improve inference capabilities at industrial sites, helping to invigorate the AI Semiconductor ecosystem."