Summary
NVIDIA describes how accelerated simulation, physics-trained AI, agents and digital twins are being used in semiconductor engineering. Its post cites partner-reported speedups of up to 35x for a simulation platform.
NVIDIA says semiconductor companies are using accelerated computing, physics-trained AI, AI agents and digital twins to speed up engineering work in chip factories. The examples in its November 3, 2026, post range from faster process simulations to virtual models of entire fabs. Engineers remain responsible for interpreting results and deciding what changes to make.
The approach, which NVIDIA calls an “AI-defined fab,” targets a practical constraint in chipmaking: testing process changes on manufacturing equipment can be slow and costly. Simulation and AI models can help engineers examine more possibilities before deciding which ones to try physically.
Where AI fits into fab engineering
Accelerated computing runs demanding calculations on hardware designed for parallel workloads. In semiconductor engineering, multiphysics simulations can model interacting effects such as heat transfer, electrical behaviour and mechanical stress. Faster calculations can let teams compare more equipment or process options within the same time.
NVIDIA reports that COMSOL is using its cuDSS library to accelerate multiphysics simulations for semiconductor manufacturing by up to 10x. Applied Materials is also integrating cuDSS into its Ginestra simulation platform, which NVIDIA says runs up to 10x faster than CPU-only approaches. Its ACE+ platform, using NVIDIA PhysicsNeMo, delivers topography simulations up to 35x faster, according to the post.
Some chip-manufacturing calculations are so demanding that they are difficult to run at production scale on CPUs. Computational lithography models how circuit patterns will print onto silicon. NVIDIA says Siemens is integrating its cuLitho software into Calibre lithography software, with up to 17x acceleration on key simulation workloads.
AI models trained on physics offer a different route: they predict process behaviour without rerunning a full simulation for every question. NVIDIA says Lam Research demonstrated that PhysicsNeMo can predict plasma behaviour in milliseconds, compared with days for a full simulation. Plasma is the electrically charged gas used in semiconductor etching, and modelling it involves interacting electrical, chemical and thermal processes.
From process models to virtual factories
AI agents can automate sequences of engineering tasks that would otherwise require manual steps. NVIDIA says Advantest is using its NeMo Switchyard agents to automate the selection of an AI model for a test-floor task, reducing response times from hours to minutes.
Digital twins extend simulation to equipment arrangements and factory operations. These virtual models can represent how materials move through a fab, helping teams examine layouts, identify bottlenecks and evaluate operating strategies before making physical changes. NVIDIA says Applied Materials uses Omniverse libraries to build digital twins of entire fabs.
Together, these tools could shorten the loop between observing a production issue, modelling a possible change and deciding whether to test it on the line. That is an engineering-workflow proposition, not a claim that AI itself makes production decisions: NVIDIA’s description assigns those decisions to engineers.
The speed figures are company-reported results for different tools and workloads. NVIDIA’s post does not provide shared benchmark conditions, so the figures should not be treated as directly comparable across those examples.