AI accelerates engineering work at semiconductor fabs
AI is also accelerating simulations, troubleshooting and the testing of manufacturing changes in semiconductor production. NVIDIA says these technologies together can create fabs where engineers make decisions faster while spending less time waiting.
- COMSOL's semiconductor multiphysics simulations can be up to ten times faster using cuDSS.
- Siemens' Calibre software can achieve up to 17x acceleration for some lithography tasks through cuLitho.
- PhysicsNeMo can predict plasma behavior in milliseconds, according to a Lam Research demonstration.
- Advantest's AI agents reduce the response time for model selection from hours to minutes.
- Applied Materials uses digital twins to test fab layouts and operating strategies.
Shorter simulation cycles
As demand for advanced chips grows, manufacturing processes are becoming more complex, while the margin for error is shrinking. According to an NVIDIA statement dated November 3, analyzing the impact of a process change, examining the interaction between materials and new manufacturing steps, and identifying the root cause of defects can take days. Physical experiments on the production line are slow and costly, so a poor decision about yield or process quality can have serious consequences.
Accelerated computing allows engineers to assess more alternatives before beginning physical trials. COMSOL uses the NVIDIA cuDSS library to accelerate multiphysics simulations for semiconductor manufacturing by up to ten times. In a collaboration between Applied Materials and NVIDIA, the Ginestra simulation platform also runs up to ten times faster than CPU-based solutions, while the Applied ACE+ multiphysics platform delivers topography simulations up to 35 times faster using NVIDIA PhysicsNeMo.
Siemens is integrating NVIDIA cuLitho technology into its Calibre lithography software. NVIDIA says this provides up to 17x acceleration for some simulation tasks, allowing more checks to be completed before the wafer is printed.
Physics-based AI and engineering agents
AI models trained on the laws of physics may be able to predict how materials and processes will behave without running a full simulation. NVIDIA cites plasma as an example. Plasma is used in semiconductor etching, and modeling its behavior requires equations that connect electrical, chemical and thermal processes, making repeated simulation costly.
According to a demonstration by Lam Research, NVIDIA PhysicsNeMo can predict plasma behavior in milliseconds, while a full simulation can take days. This allows engineers to examine different process conditions more quickly. Applied Materials uses digital twins built on NVIDIA Omniverse libraries to simulate fab-level layouts, throughput, costs and process control.
AI agents can also automate the engineering steps between decisions. Advantest uses NVIDIA NeMo Switchyard agents to automate the selection of which AI model to apply to a given task. At the company, this reduces response times from hours to minutes.
Digital twins before the production line
Digital twins enable equipment configurations, fab layouts and process changes to be tested in a virtual environment. NVIDIA says this can reduce the cost of experimentation, particularly in increasingly complex fabs where the risks of a poorly chosen change are also greater.
Applied Materials is creating physically accurate digital twins of entire semiconductor fabs based on NVIDIA Omniverse libraries. Teams can use them to simulate material movement, identify bottlenecks, and validate layouts and operating strategies before deployment.
NVIDIA calls these interconnected technologies an AI-defined fab. The company's vision is for production data to feed the models, the models to evaluate changes, and the results to return to engineers. Engineers will continue to make the decisions based on their own expertise, while learning faster from each manufacturing cycle.
