Samsung cuts an 18-hour chip simulation to 41 seconds as Korea pushes physical AI with Nvidia

Korea’s industrial giants are moving AI from demos into engineering and production. Nvidia highlighted Samsung using GPU solvers and AI physics in semiconductor design, Kakao Mobility building toward level 4 autonomous driving on Nvidia’s vehicle platform, and LG Electronics training its CLOiD robot inside a real factory. Each bets that simulation plus AI shortens the loop between design, testing and deployment; if those loops shrink, product cycles in chips, cars and factories shrink too. It came alongside Nvidia’s Korean startup investments (startups) and agent-security news (secure agents).

Selected quantum chemistry runs on Samsung’s sMOL platform finished up to five times faster end to end with cuEST. In one thermal and mechanical case using PhysicsNeMo, combining AI-predicted material properties with a coarser grid and GPU acceleration cut a roughly 18-hour, 32-core CPU job to 41 seconds, about a 1,600x speedup. That changes how many design options engineers can afford to try.

The caveat: part of that speedup comes from the coarser grid and AI-estimated properties, not just faster hardware. That is legitimate but trades fidelity for speed, and the hard question is when the approximation is trusted enough to sign off a chip. Validation against fine-grid baselines matters more than stopwatch numbers. Nvidia also floated its Agent Toolkit to link Samsung’s simulation tools so engineering agents could run analyses and compare designs, described as a future direction, not a deployed system.

Kakao Mobility intends to use Nvidia Hyperion with DRIVE AGX Thor and the Halos safety stack, starting on Kia purpose-built vehicles and drawing on operating experience in Seoul’s Gangnam district. Level 4 is where public-safety stakes get real, and the operative word is “plans”; it should be judged on safety disclosures and permitted service areas, the same yardstick Nvidia invites with its safety-layer pitch (physical AI safety).

LG’s CLOiD work is the clearest picture of where factory robotics is going: a flywheel that collects real-world data, augments it with Cosmos, trains GR00T models, runs on Jetson AGX Thor and evaluates in simulation, the same stack Nvidia showed at CoRL (GR00T 2). Repeatable training loops are good engineering; scaled across factories they also mean fewer manual line roles, and companies should discuss workforce plans as plainly as workflows. More in our Nvidia coverage and robotics coverage.