Semiconductor Companies Adopt NVIDIA CUDA-X Libraries for AI-Defined Fabs

NVIDIA announced that semiconductor companies are adopting NVIDIA CUDA-X libraries to support what the industry is calling AI-defined fabs — manufacturing operations where artificial intelligence assists engineers across simulation, inspection, testing and automation.

"AI-defined fabs use AI to help engineers optimize processes, investigate production issues and improve manufacturing efficiency."

For manufacturing professionals, the significance lies less in the concept and more in the reported performance figures. Companies across the semiconductor workflow are posting acceleration numbers that, if sustained in production, change the economics of simulation-heavy engineering work.

Atomic-scale simulation: Synopsys and the new cuPotential library

The newest addition to the portfolio is the NVIDIA cuPotential library, which supports atomic-scale simulations of electrical forces and interactions across large groups of atoms.

"NVIDIA cuPotential provides GPU-accelerated functions that developers can add to existing software."

"The library can also be used by AI agents to run, evaluate and update molecular simulations."

"In early-access testing, cuPotential provided an additional 2–3x acceleration for workflows that were already optimized for GPUs."

That last point deserves emphasis: the 2–3x gain applies to code already tuned for GPUs, not to legacy CPU implementations.

Synopsys uses cuPotential, cuEST, and cuDSS for semiconductor simulations, and the company has shared early benchmark data:

"Early QuantumATK results showed a potential 30x speedup with cuEST, while S-Device benchmarks showed a 12x speedup with cuDSS."

Multiphysics and design: Cadence, COMSOL, Voltai, ChipAgents

Cadence uses NVIDIA accelerated computing and CUDA-X libraries for AI-powered design and simulation of electromechanical systems, including the AuraStack AI Super Agent and ChipStack AI Super Agent, with a reported 8x improvement in one advanced-packaging workflow.

COMSOL uses NVIDIA cuDSS to accelerate semiconductor multiphysics simulations, reporting gains of up to 10x. Voltai uses NVIDIA cuDSS and CUDA-X libraries in Voltaire, its agent-based multiphysics platform for 3D-IC design and simulation. ChipAgents, meanwhile, works with NVIDIA on Build-A-Chip Bench, an AI benchmark for evaluating agents across workflows.

Process modeling and package design

Siemens uses cuLitho to accelerate Calibre software for chip-pattern refinement and mask preparation. Lam Research plans to integrate NanoVDB into SEMulator3D to accelerate wafer-etch modeling tasks:

"Lam Research reports that NanoVDB can accelerate selected wafer-etch modeling tasks by up to 30x."

Luminary uses NVIDIA PhysicsNeMo for semiconductor package design to predict temperature, pressure drop, and thermal stress:

"Luminary reports that engineers can evaluate 100x more designs within the same schedule without running a separate simulation for every option."

Inspection and testing

Imec uses NVIDIA AI and accelerated computing to remove noise from scanning electron microscope images:

"Imec reports a 30% improvement in the lithography process window."

KLA applied production-kernel techniques to an NPP NCC kernel:

"KLA reported performance up to 50x faster than earlier CPU implementations."

Teradyne uses NVIDIA NPP and accelerated computing for image-sensor testing and RAPIDS cuDF and CuPy for GPU calculations:

"Teradyne reports approximately 85x faster probe-maintenance inference."

Advantest integrates NVIDIA Nemotron and NVIDIA Kumo-Anomaly into ACS RTDI for semiconductor production analytics. Applied Materials, which uses NVIDIA Omniverse alongside the CUDA-X libraries, reports its own figure:

"Applied Materials reports chamber simulation speeds up to 35x faster."

The takeaway

The pattern across these deployments is consistent: GPU-accelerated libraries are being dropped into existing engineering software rather than replacing it, and AI agents are being layered on top to run and evaluate simulations autonomously. Reported speedups range from 2–3x on already-optimized GPU code to 100x more design evaluations in package engineering. Whether those numbers hold at production scale is the question engineers will want answered — but the breadth of adoption, spanning simulation vendors, equipment makers and test houses, suggests this is more than a benchmarking exercise.