NVIDIA and Semiconductor Leaders Use CUDA-X Libraries to Advance AI-Defined Fabs

On October 14, NVIDIA announced during SEMICON West 2026 that leading semiconductor companies are adopting NVIDIA CUDA-X libraries, including the newly released cuPotential. This library enables developers to build AI agents that simulate atomic and molecular interactions, advancing the development of "AI-defined fabs."

An AI-defined fab refers to a semiconductor facility where AI assists engineers in optimizing processes, solving production challenges, and improving efficiency. As demand for advanced chips continues to rise, these facilities must expand capacity while managing increasingly complex manufacturing workflows. Engineers require new methods to model materials, analyze inspection data, and test process changes via digital twins before committing solutions to mass production.

cuPotential is designed for modeling atomic and molecular interactions, leveraging dedicated GPU acceleration to speed up atom-level simulations. Developers can integrate it into existing software. It allows AI agents to run, evaluate, and iterate on molecular simulations. In early access testing, workflows already optimized for GPUs saw an additional 2x to 3x speedup.

In chip design, the CUDA-X suite also includes cuEST, which simulates electronic behavior, and cuDSS, which solves complex engineering simulations. Synopsys uses these three libraries to accelerate semiconductor simulation: cuEST delivered a 30x speedup in early QuantumATK tests, while cuDSS provided a 12x speedup in S-Device benchmark tests. Cadence has introduced agent-based AI into the design and simulation of electromechanical systems; its AuraStack AI Super Agent leverages cuDSS to accelerate thermal management and stress optimization in advanced packaging by 8x.

In the manufacturing phase, the cuLitho library accelerates lithography calculations for chip patterning, the open-source NanoVDB library enhances chip manufacturing simulations, and the PhysicsNeMo open framework aids in training AI models that predict physical systems. Siemens uses cuLitho to accelerate chip pattern optimization and mask preparation in Calibre software. Lam Research plans to integrate NanoVDB into its SEMulator3D process modeling platform, with some wafer etching modeling expected to see speedups of up to 30x. Luminary applies PhysicsNeMo to packaging design, allowing engineers to explore over 100 times more options within the same time frame.

In inspection and testing, NPP handles image processing on GPUs, while RAPIDS cuDF manages large-scale datasets on GPUs. imec used NVIDIA AI and cuLitho to improve the lithography process window by 30%. KLA’s new NPP NCC kernel is 50x faster than its earlier CPU version. Teradyne accelerated probe maintenance inference by approximately 85x. Advantest integrated Nemotron and Kumo-Anomaly into ACS RTDI for production analysis, reducing response times from hours to minutes.

For fab operations and automation, Applied Materials uses Omniverse and CUDA-X libraries to accelerate chamber simulations by 35x and build fab digital twins. Analog Devices, FANUC, Foxconn, and OMRON have connected sensor simulations, robot models, and vision agents to NVIDIA’s digital twin and open-world model frameworks.

Currently, the cuPotential early access version is live, NanoVDB has been released as an open-source component of OpenVDB, and NPP is available through the NVIDIA CUDA Toolkit. Lam Research’s SEMulator3D, accelerated by NanoVDB, is expected to launch later this year.