Nvidia in Korean factories: artificial intelligence physics cuts chip simulation to 41 seconds
Listen to this article
Read by Anchor
Recent developments in South Korea reveal a qualitative shift that goes beyond generative text models, as high-performance computing moves directly onto factory floors and chip production lines and smart vehicles. The bet is no longer confined to isolated digital algorithms, but is now rooted in what is called physical artificial intelligence, which merges physical simulation and deep-learning models into real manufacturing hardware through partnerships led by Nvidia with major Korean industrial conglomerates.
Samsung Electronics leads this path by introducing artificial-intelligence physics algorithms into chip design, materials research and device engineering.The company recorded a qualitative leap in analyzing thermal and mechanical stresses of chips, as simulation time dropped by about 1,600 times, falling from roughly 18 hours on a 32-core central processor to just 41 seconds using GPU acceleration and the PhysicsNeMo platform.In parallel, Samsung’s sMOL platform succeeded in speeding up quantum-chemistry calculations by up to five times using the cuEST library, while the company is testing the cuEquivariance library to accelerate tensor operations for atomic AI models, employing the cuISS and cuDSS libraries to accelerate sparse linear systems in computer-aided design and optical processes.
This transformation extends to transport and industrial robotics, as Kakao Mobility moves toward building a level-4 autonomous driving system based on the Hyperion platform, DRIVE AGX Thor processors and Halos safety technologies, integrating its field-tested software in the Gangnam district with custom vehicles from Kia. In production lines,LG Electronics showcased its industrial robot CLOiD operating in a real manufacturing environment using the Isaac GR00T model and portable Jetson AGX Thor computing.It establishes what is known as the physical data loop that gathers operational data and artificially multiplies it via the Cosmos platform to train robots and reevaluate them periodically.
This shift carries clear strategic implications for those leading industrial and logistics projects in the Gulf and Egypt, where current investments in computing infrastructure largely focus on language models and text inference. The Korean experience shows that the real return on manufacturing lies in merging physical simulation with graphics units to shrink engineering cycles from days to a few seconds. In practice, this means that advanced research and manufacturing centers and mobility-solution developers in the region need to redirect computing capabilities toward 3D simulation environments and material modelling, and to equip engineering teams to work with predictive-physics algorithms rather than relying solely on traditional software engineering, especially in autonomous-transport projects and electronics localisation that require costly physical safety testing before field deployment.
These steps place manufacturing systems at a new stage where engineering agents automatically link simulation models with design-option testing. When thermal-stress testing time for a chip drops to seconds, the cost of error changes and the pace of engineering innovation accelerates, making on-site ownership of physical data loops the benchmark of industrial superiority in the next phase.