- Establishes integrated management platform for video data and AI models to enhance AI inference capabilities and development productivity
- Applies a hybrid structure of NPU inference and GPU learning... Promoting the localization of the industrial AI ecosystem 

POSCO DX (CEO Shim Min-suk) is taking steps to achieve AI technology self-reliance in the industrial sector and reduce AI infrastructure costs by developing an unstructured industrial data analysis platform utilizing domestic AI semiconductors.

The NPU, a representative domestic AI semiconductor, is a semiconductor specialized for AI computations such as deep learning and machine learning. Optimized for AI inference compared to GPUs, it can reduce infrastructure costs and power consumption. In particular, by being directly integrated into field equipment control systems, it is suitable for implementing ‘Edge AI,’ which bypasses remote AI data centers or servers. This offers the advantage of ensuring security and real-time performance by enabling immediate inference without data leakage in manufacturing sites where security is critical. 

POSCO DX has incorporated these technological advantages into its 'Vision AI Platform.' The newly developed platform operates diverse video data collected from industrial sites as training data within a single environment, and standardizes functions repeatedly required for Vision AI service implementation—such as AI model management, performance metrics, and application history—to provide them as common components. This is expected to shorten the execution period for new projects and enable the rapid expansion of services to various projects and industrial sites. 

The most significant feature of this platform is its adoption of a hybrid approach that utilizes GPUs for AI model training and development, as before, while applying domestically produced NPUs for real-time computation and decision-making in industrial settings. By establishing an environment where NPUs can be utilized even during the research and verification phases, the platform enables the pre-verification of NPU-based inference performance and effectiveness before actual industrial deployment. This allows for the verification of AI model accuracy, processing speed, power efficiency, and operating costs prior to field application, thereby facilitating the identification of optimal AI models and NPU application strategies tailored to specific site characteristics. 

Previously, POSCO DX has been pursuing the localization of AI in various fields, such as industrial fire monitoring, worker safety monitoring, and product loading status verification in logistics environments, based on cooperation with domestic NPU developers like DEEPX and Mobilint. As a result, it has confirmed that infrastructure construction and operation costs are reduced by approximately 50% compared to GPUs with equivalent inference performance, and power consumption is reduced by approximately 90%. Based on this, POSCO DX plans to gradually transition sites where Vision AI technology is applied to domestic NPU-based platforms and expand the scope of application to various industrial sites. 

A POSCO DX official stated, “The NPU-based Vision AI platform will serve as a foundation for more efficiently implementing AI optimized for industrial sites,” adding, “We will contribute to the systematization of domestic NPU utilization and the enhancement of inference capabilities in industrial settings, thereby stepping forward to revitalize the AI semiconductor ecosystem.” 

Meanwhile, POSCO Group is actively pursuing the realization of a safe and comfortable workplace based on technology by expanding intelligent factories utilizing physical AI in manufacturing sites to respond promptly to the industrial paradigm shift, including AX. In addition, at the CEO Investor Day held on July 2, the company announced plans to commercialize physical AI for the process industry, based on its experience in equipment automation and intelligence accumulated in the steel industry and a vast amount of field data.