NVIDIA Details AI Tools for Semiconductor Fab Simulation Workflows

NVIDIA is presenting AI tools for semiconductor fabs as a way to evaluate manufacturing changes in software, citing faster results on specific simulation workloads as chipmaking processes grow more complex.

In a Nov. 3 post, the company described an “AI-defined fab” that combines accelerated computing, physics-based models, AI agents and digital twins. NVIDIA said engineers would steer the systems and retain authority over decisions. The examples span simulations, model selection and virtual models of factory operations.

For semiconductor manufacturing multiphysics simulations, COMSOL is using NVIDIA’s cuDSS library, which NVIDIA said can accelerate the work by up to 10 times. Applied Materials’ Ginestra platform also runs up to 10 times faster than CPU-only approaches when using cuDSS. Its ACE+ platform delivers topography simulations up to 35 times faster with NVIDIA PhysicsNeMo.

Siemens is integrating NVIDIA’s cuLitho into its Calibre lithography software, with acceleration of up to 17 times on key simulation workloads. The results cover different tasks and comparisons. NVIDIA also described work with Lam Research, where PhysicsNeMo predicts plasma behavior in milliseconds instead of waiting days for a complete simulation.

Advantest is using NVIDIA NeMo Switchyard agents to automate AI model selection. NVIDIA said the approach has shortened response times from hours to minutes. The examples show agents handling a specific selection workflow alongside software used to model physical processes.

Applied Materials uses digital twins built with NVIDIA Omniverse libraries to model fab layouts, throughput, cost and process control. Teams can use the virtual models to assess equipment setups and process changes before physical trials, which NVIDIA said can lower experimentation costs. The work extends NVIDIA’s earlier demonstrations of AI agents assembling Omniverse simulation apps.