Samsung achieves 17-second chip simulation using Nvidia processors
The performance figures emerged at Nvidia AI Day Seoul 2026, held on November 9 and 10, where both South Korean chipmakers presented applications of accelerated computing and AI to semiconductor development. Samsung’s fastest configuration delivered an approximately 3,800-fold improvement over a conventional processor-based calculation, although the comparison involved different computational methods.
Samsung demonstrated a thermal and mechanical simulation designed to predict how semiconductor structures respond to heat and physical stress. The original calculation required approximately 18 hours on a 32-core central processing unit using a detailed computational grid. An AI-assisted calculation finished in 41 seconds on one Nvidia H100 graphics processor and 17 seconds using four processors.
The single-GPU configuration produced an approximately 1,600-fold speed improvement, the figure highlighted in Nvidia’s official announcement. The faster four-GPU result illustrates additional acceleration through parallel processing, rather than an identical test conducted with unchanged hardware and numerical methods.
Samsung’s approach combines Nvidia’s PhysicsNeMo artificial intelligence framework with a technique known as AI homogenisation. This replaces highly detailed representations of complicated semiconductor structures with simplified computational models that preserve important physical properties.
The companies demonstrated how a structure containing approximately one billion fine-scale computational elements could be represented using roughly one million larger cells. AI models estimate effective material properties, allowing engineers to calculate thermal behaviour and mechanical deformation without repeatedly processing every microscopic feature.
Samsung reported 99.17% accuracy in effective vertical thermal conductivity for the demonstrated workflow. That measurement applies to the specified physical property and test conditions, rather than establishing equivalent accuracy across every aspect of semiconductor simulation.
The development addresses a significant engineering challenge as advanced semiconductor packaging introduces increasingly complicated arrangements of materials and components. Temperature variations and mechanical stress can affect chip reliability, making accurate simulations important before designs enter manufacturing.
Samsung’s Computational Science and Engineering team worked with Nvidia to develop an integrated scientific machine-learning pipeline. The process generates synthetic semiconductor layouts, produces simulation data through GPU-accelerated calculations and trains physics-aware models to estimate material behaviour.
SK hynix separately outlined its TCAD 3.0 programme, which integrates AI with technology computer-aided design software used to model semiconductor manufacturing processes and device characteristics before physical production.
Kim Jung-han, head of the company’s TCAD Intelligence team, presented an AI model developed with Nvidia PhysicsNeMo to predict changes in material profiles during etching. The system produces results comparable to conventional simulations at speeds described as several dozen times faster.
Etching involves removing selected materials from semiconductor wafers through tightly controlled manufacturing steps. Individual processes can involve numerous variables, including temperature, pressure and gas flow, making comprehensive simulation computationally demanding.
SK hynix also demonstrated chip-scale thermal and mechanical calculations running more than 30 times faster than a conventional 32-thread CPU configuration. Its approach distributes calculations across multiple graphics processors while reusing information obtained from smaller computational regions.
The company said performance increased almost proportionately as additional GPUs were introduced, suggesting that larger computing installations could handle increasingly demanding engineering workloads.
The two manufacturers’ performance figures reflect different simulations, hardware configurations and measurement conditions. Neither demonstration establishes that complete semiconductor development cycles can achieve equivalent reductions in duration.
Samsung is extending its Nvidia collaboration into semiconductor process modelling, optical simulation and materials research. Its sMOL research platform has achieved up to fivefold acceleration in selected quantum chemical calculations using Nvidia’s cuEST software.
The company is also evaluating Nvidia’s cuISS and cuDSS libraries for computationally intensive mathematical operations involved in semiconductor device modelling. These tools address different engineering workloads from the thermal simulation demonstrated in Seoul.
SK hynix’s broader collaboration with Nvidia includes accelerating existing simulation software and developing AI-assisted manufacturing models. The companies have also agreed to cooperate on advanced memory technologies and digital representations of semiconductor fabrication facilities.
Nvidia’s PhysicsNeMo framework supports machine-learning models that incorporate physical principles and simulation data. Its Warp software provides GPU-accelerated computing capabilities used to generate training information for engineering applications.
Samsung’s demonstrated pipeline uses these technologies together, with the trained models supplying simplified material properties for subsequent large-scale calculations. The companies are exploring software agents capable of coordinating simulation tasks and evaluating alternative engineering designs.
Samsung has made an example of its physics-based modelling work available through Nvidia’s PhysicsNeMo resources, alongside research covering semiconductor simulation techniques presented at the 2026 Design Automation Conference.
http://www.wellerabangkok.com/
https://www.linkedin.com/company/nexaworkai
https://web.facebook.com/nexaworkai
