Artificial intelligence sped up some semiconductor manufacturing simulations by up to 35 times, helping engineers evaluate changes before physical trials, Nvidia said in a blog published on Nov. 3.
The gains varied across applications, with engineers retaining responsibility for production decisions, the company said. Nvidia, whose technology powers the applications, published “AI Is Changing How Semiconductor Fabs Work — and Engineers Are Moving Faster Because of It.”
Nvidia described applications spanning materials development, manufacturing processes, chip testing and factory planning. The examples covered work by semiconductor equipment and software companies, without specifying a common testing period.
Applied Materials’ ACE+ platform delivered the largest reported speed increase for simulations of how manufacturing processes shape surfaces, using Nvidia’s PhysicsNeMo technology, Nvidia said. The company described the gain as an upper limit.
Applied Materials’ Ginestra platform also ran up to 10 times faster than approaches using only central processing units, Nvidia said. The companies are collaborating on materials discovery and process development at the atomic scale.
Software provider COMSOL used Nvidia’s cuDSS library to speed simulations by up to 10 times, Nvidia said. These calculations combine physical effects such as heat transfer, electrical behavior and mechanical stress.
Faster calculations allowed engineers to compare more equipment and process alternatives before choosing changes for physical testing, the company said. It said experiments on factory production lines are slow and expensive.
Siemens is integrating Nvidia’s cuLitho technology into its Calibre software, delivering speed gains of up to 17 times on selected simulation tasks, Nvidia said. The software calculates how circuit patterns are printed onto silicon.
Some of those calculations are too demanding for central processing units to complete at production scale, the company said. Faster processing expands the checks engineers can perform before printing a wafer.
Lam Research demonstrated predictions of plasma behavior in milliseconds using PhysicsNeMo, compared with days for a full simulation, Nvidia said. Plasma, an electrically charged gas, is used to etch semiconductor materials.
The approach lets engineers explore process conditions without repeating a full simulation each time, Nvidia said. Modeling plasma requires calculations linking electrical, chemical and thermal processes.
Advantest used Nvidia’s NeMo Switchyard AI agents to automate the choice of an AI model for a task, Nvidia said. Response times fell from hours to minutes for a process previously handled manually.
Applied Materials also uses Nvidia’s Omniverse libraries to build digital twins, virtual models of factories, Nvidia said. Teams can test material movement, find bottlenecks and compare layouts before physical trials.
Connecting production data with models could help engineers evaluate changes and apply findings to production decisions sooner, Nvidia said. Engineers would continue to direct the work and make decisions using their expertise.
