POSCO DX said on the 2nd it will accelerate manufacturing artificial intelligence (AI) innovation by using a domestic Neural Processing Unit (NPU). By introducing an AI infrastructure based on a domestic NPU that replaces foreign GPUs, the company aims to cut expense and move to build an intelligent factory capable of on-site, immediate analysis and control.
POSCO DX signed a business agreement at its Pangyo office the same day with AI Semiconductor startup Mobilint to implement NPU-based AX. In Feb., POSCO DX invested a total of 3 billion won in Mobilint through its corporate venture capital (CVC) new technology investment fund.
An NPU is a semiconductor specialized for AI computation such as Deep Learning and Machine Learning, offering higher power efficiency than a GPU and optimization for inference, which can be expected to reduce infrastructure expense. It is particularly suitable for implementing Edge AI by embedding it directly into on-site equipment control systems, enabling real-time analysis and control without external data transfer.
POSCO DX plans to build an intelligent factory by applying Mobilint's NPU to its in-house industrial control system "POSMASTER," enabling immediate AI analysis and control at the equipment control stage. The two companies will also jointly promote technology development to maximize AI model algorithm performance in an NPU environment.
Mobilint has high-performance NPU technology capable of running large language models (LLMs) in edge environments, enabling real-time decision-making and control using on-site data.
POSCO DX plans to convert its GPU-based AI systems to NPUs and expand application to various industrial sites, including steel, secondary battery material production, and logistics.
Cho Seok-ju, head of the POSCO DX AX Convergence Research Institute, said, "With this collaboration, we have secured LLM-based Edge AI technology needed to build an intelligent factory," and added, "We will lead a paradigm shift in manufacturing AI through cooperation with domestic startups."