NVIDIA: AI Reshapes Chip Factories, Simulation Speeds Up by Up to 35x

On November 4, NVIDIA published an article on November 3 stating that artificial intelligence is driving semiconductor fabs into the "AI-defined factory" stage: through accelerated simulation, learned physics models, AI agents and digital twins, it is significantly speeding up design, testing and production decisions in chip manufacturing.

Accelerated simulation refers to using GPUs, parallel computing, AI agent models and other technologies to compress engineering simulations that traditionally rely on CPUs from hours or days to minutes or even milliseconds. In semiconductor manufacturing, it allows engineers to test process options and predict defect risks faster without first running expensive and time-consuming real production-line experiments.

Citing the blog post, as demand for advanced chips rises, manufacturing processes become increasingly complex, and the margin for error in yield and process quality keeps shrinking. Traditional production-line experiments are both slow and expensive, and engineers often need days to analyze process changes, material reactions or defect root causes. The value of AI lies in enabling engineers to compare options and identify risks faster in a virtual environment before committing to real experiments.

The goal of accelerated simulation is to obtain sufficiently reliable "virtual experiment" results faster before real manufacturing, thereby reducing trial and error, improving yield and lowering costs.

Learned physics models refer to models that use machine learning to learn physical laws from high-precision simulations, experiments or production-line data. They learn the mapping relationships among variables such as temperature, pressure, materials, electromagnetic fields and plasma, thereby quickly predicting complex physical processes. In its semiconductor manufacturing report, NVIDIA regards this as one of the key technologies for accelerating fab simulation and decision-making.

A digital twin refers to creating a dynamic digital replica of a physical object, system or process in a virtual environment. It is not just a static 3D model, but can receive real-time or historical data, simulate operating states, predict outcomes, and support a decision-making approach of "try first in the virtual world, then do it in the real world." In NVIDIA's report on semiconductor manufacturing, digital twins are used to build a virtual model of an entire fab, helping engineers identify bottlenecks and validate layout and operational strategies before production begins.

In specific cases, multiphysics simulation software COMSOL, with the help of NVIDIA cuDSS, increased the speed of semiconductor manufacturing-related simulations by up to 10 times; Siemens integrated cuLitho into its Calibre lithography software, accelerating key simulation workloads by up to 17 times.

Applied Materials' Ginestra platform, after integrating cuDSS, is up to 10 times faster than CPU-only solutions, and its ACE+ multiphysics platform, with the help of PhysicsNeMo, achieves up to 35 times acceleration in topography simulation.

AI can also reduce repeated simulations. For example, modeling the behavior of plasma used in semiconductor etching is complex and costly; Lam Research demonstrated that PhysicsNeMo can predict plasma behavior at millisecond speed, whereas a full simulation may take days.

In the past, engineers had to wait for long simulations, experiments or data analysis to determine whether a chip manufacturing plan was feasible; after introducing AI-accelerated simulation, learned physics models and AI agents, many analysis tasks that originally took "days" can be completed in "minutes."

In its report, NVIDIA gives an example: Advantest uses the NeMo Switchyard AI agent to automatically select the appropriate AI model, compressing tasks that originally took hours into minutes.

In advanced fabs, the questions engineers face are often not simple data queries, but usually require judgment:

Traditional workflows usually require high-precision physical simulation, result analysis and option comparison in sequence, and sometimes actual experiments must be arranged. A complete simulation may take days, making it difficult for engineers to quickly explore a large number of alternatives.

AI can first help engineers narrow the scope and identify high-risk options, while key process changes still require high-precision simulation, experimental validation and professional review by engineers. AI frees engineers from "waiting a long time for calculation results," allowing them to spend time on judgment, validation and optimization; the waiting time in the decision chain is shortened, but professional responsibility and final gatekeeping remain in the hands of engineers.

An intuitive analogy is that in the past a doctor had to wait for a full set of test reports before making a preliminary judgment on a condition; now AI is like quickly providing multiple screening results and possible causes first, so the doctor can immediately decide which key tests to prioritize. It is faster, but the final diagnosis and treatment decisions are still the doctor's responsibility.

More importantly, Applied Materials is building a digital twin of an entire fab based on NVIDIA Omniverse, which can simulate material flow, identify bottlenecks and validate layout and operational strategies in a virtual environment before deciding whether to implement them on a real production line.

Applied Materials uses digital twin technology built on NVIDIA Omniverse libraries to simulate factory-level optimization of layout, throughput, cost and process control.

NVIDIA says this closed loop will enable factories to learn from each production cycle and act earlier; engineers remain responsible for professional judgment and final decisions, but the time spent waiting for answers will be greatly reduced.

A realistic digital twin model of a fab built with NVIDIA Omniverse libraries, used to simulate material flow, identify bottlenecks and validate operational strategies before deployment. Image courtesy of Applied Materials.