NVIDIA and Korean Industry Leaders Bring Physical AI into Production
Summary:NVIDIA is expanding physical-AI deployments with Samsung Electronics, Kakao Mobility, and LG Electronics across semiconductor engineering, Level 4 autonomy, and manufacturing robotics. Samsung is using CUDA-domain libraries and PhysicsNeMo to accelerate sparse linear systems, quantum chemistry, tensor operations, and chip-scale thermal and mechanical analysis. In one reported material-homogenization case, GPU acceleration and coarse-grid modeling reduced solve time from roughly 18 hours to 41 se
Physical AI Enters Three Industrial Pipelines
NVIDIA’s work in Korea is not a single model launch but a set of engineering collaborations built around accelerated computing, AI simulation, and edge inference. Samsung Electronics is applying the stack to semiconductor materials and device engineering, Kakao Mobility to Level 4 autonomy, and LG Electronics to manufacturing robotics. Together, the projects cover chip design, vehicle development, and robot deployment, linking physical AI from design validation to production.
Unlike conventional generative-AI applications, physical systems must satisfy geometric, material, thermal, and dynamic constraints. AI can predict effective material properties, generate training data, or navigate a design space, but the results still require high-fidelity solvers, hardware-in-the-loop testing, and real-world validation. GPUs therefore do not simply replace CPUs; they make it possible to iterate across expensive numerical simulation, learned models, and deployed hardware.
Samsung: Rebuilding Semiconductor Simulation on CUDA
Samsung Electronics and NVIDIA are collaborating across materials research, process development, device engineering, and chip-scale thermal and mechanical analysis. In technology computer-aided design, Samsung uses cuISS for sparse linear systems and GPU acceleration for process and optical simulations. cuDSS provides direct sparse-solver capabilities. These workloads often contain irregular matrix structures, so parallel efficiency depends heavily on nonzero distribution, matrix fill-in, and the degree to which linear-algebra operations can be batched efficiently.
For materials computation, Samsung’s sMOL platform uses cuEST for selected quantum-chemistry calculations, reporting up to a fivefold end-to-end speedup. cuEquivariance is being explored for tensor operations in the training and inference of atomistic AI models. Its practical value lies in exploiting symmetry constraints: specialized kernels can avoid redundant contractions or group-representation transformations that generic matrix multiplication would perform inefficiently. The reported gains nevertheless apply to selected workloads and should not be generalized to every molecular system or model size.
What Makes the 1,600× Speedup Possible
Samsung is also using PhysicsNeMo for chip-scale thermal and mechanical analysis. In one AI material-homogenization case, the team combined AI-predicted effective material properties, coarser grids, and GPU acceleration. A fine-grid solve that took about 18 hours on a 32-core CPU fell to 41 seconds with a GPU-accelerated coarse-grid solve, yielding a reported speedup of roughly 1,600×.
That figure does not come from GPU peak compute alone. It results from stacking three approximations: AI predicts effective properties more cheaply, the coarser grid reduces degrees of freedom and solver iterations, and the GPU accelerates the remaining large-scale linear solve. In engineering simulation, such model-based approximation can matter more than raw hardware throughput. The trade-off is equally important: coarse grids and predicted properties may miss local hotspots, interface effects, or nonlinear transitions. Convergence studies, error bounds, and reference solutions remain necessary before these results can support manufacturing decisions.
Kakao Mobility: Building a Closed-Loop L4 Platform
Kakao Mobility plans to combine its proprietary autonomous-driving software and experience operating in Seoul’s Gangnam district with NVIDIA’s vehicle computing and safety technologies to develop a Level 4 system. Its reference architecture for sensors and compute will use Hyperion, with DRIVE AGX Thor providing vehicle compute and Halos supporting safety across development and validation.
The central issue is not the compute capacity of a single automotive SoC, but vertical integration across sensing, vehicle computing, simulation, and safety assurance. An L4 stack must process camera, radar, and lidar data; perform perception, prediction, planning, and control; and demonstrate safety within a defined operational design domain. DRIVE AGX Thor handles high-throughput inference and real-time onboard tasks, while Halos moves functional safety, reliability, and validation earlier in the development cycle so that model capability and production requirements remain aligned.
Initial work is planned around Kia purpose-built vehicles, integrating Kakao Mobility software with NVIDIA technologies in support of future Level 4 mobility services in Korea. Planned adoption, however, is not equivalent to completed road-level validation. The metrics that matter include failure modes under severe weather and rare events, minimal-risk fallback behavior, simulation-to-reality domain gaps, shadow-mode operation, and the recurring cost of fleet-wide regression testing.
LG: Building a Robot Data Flywheel with Isaac GR00T
NVIDIA and LG Electronics demonstrated NVIDIA Isaac GR00T-powered manipulation and manufacturing tasks using LG’s CLOiD robot in an actual manufacturing environment. The workflow connects real-world data collection, NVIDIA Cosmos-based data augmentation, GR00T policy training, onboard inference on Jetson AGX Thor, and simulation-based evaluation.
The important advance is not a single successful demonstration but a repeatable development loop. Real-world collection captures the current production state, synthetic data broadens coverage for long-tail situations, simulation provides regression testing and risk-scenario screening, GR00T maps multimodal observations to actions, and Jetson AGX Thor delivers real-time edge inference.
The practical bottleneck in robot learning is often data governance rather than the policy architecture alone. Different tasks, tools, and workstations need consistent action representations; synthetic visual data must preserve contact mechanics and object affordances; and simulation inevitably introduces dynamics mismatch. LG has not yet disclosed task-success rates, intervention frequency, cross-workstation transfer, or end-to-end training cost, which are necessary to judge production scalability.
The Next Layer: Engineering Agents Across Simulation and Design Space
Samsung is also evaluating NVIDIA Agent Toolkit to connect simulation tools and AI models, allowing future engineering agents to execute analyses and compare design options. The objective is to elevate agents from conversational assistants into controlled engineering executors, but every step requires traceable inputs, deterministic solver outputs, and explicit validation conditions.
A dependable system needs three layers. A tool layer restricts agents to certified solvers. A validation layer checks outputs against reference models, grid convergence, and physical conservation laws. An audit layer records parameters, model versions, and design decisions. Without these controls, an agent may generate plausible-looking simulation results quickly while allowing approximation errors to propagate into manufacturing choices.
Recommendations for Developers and Enterprises
Enterprises should distinguish three sources of value: acceleration of individual solves, fewer design iterations, and replacement of an explicit model. The first should be tested with identical grids, tolerances, and reference solutions while measuring end-to-end time and energy. The second requires counting design iterations and total engineering hours. The third requires comparing error distributions, not throughput alone.
A practical deployment path starts with bounded workloads such as sparse linear solving, quantum chemistry, or thermal analysis, then expands into robot policies and engineering agents. Teams should record GPU memory, CPU fallback time, data-transfer volume, and peak-load behavior so that apparent acceleration is not confused with idle resources or dataset differences.
The Korean projects show that physical AI competition is shifting from isolated model capability toward integration across GPU compute, numerical simulation, synthetic data, edge inference, and domain software. Most reported acceleration figures nevertheless come from controlled vendor examples, and their generalizability across processes, vehicle platforms, and production lines remains unproven.
— END —Source: NVIDIA AI News (2026-11-09)
Tags: #Physical AI #GPU-Accelerated Simulation #Robot Learning #Autonomous Driving #CUDA
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments