Nvidia Wires CUDA-X Into Fabs, Debuts cuPotential Library

Nvidia on October 14 named Synopsys, Cadence, KLA, Lam Research, Applied Materials, Siemens, Teradyne and Advantest as partners wiring its CUDA-X libraries into chip design, inspection and manufacturing tools, and introduced cuPotential, a new CUDA-X library for AI agents that simulate how atoms and molecules interact. In its announcement, the company says early-access testing showed cuPotential accelerated already GPU-optimized workflows by an additional 2-3x.

"The semiconductor industry made the AI revolution possible. Now, AI is transforming how chips are built," said Tim Costa, Nvidia's VP and GM of Industrial and Computational Engineering.

The partner benchmarks Nvidia published are specific and spread across the stack. Teradyne reported roughly 85x faster probe-maintenance inference. KLA's NPP NCC image-processing kernel ran 50x faster than earlier CPU versions. Lam Research is integrating NanoVDB into its SEMulator3D wafer-etch platform for a targeted 30x speedup, aimed at an end-of-year release. Synopsys clocked 30x on QuantumATK using cuEST and 12x on S-Device using cuDSS. Cadence's AuraStack AI Super Agent optimized advanced packages 8x faster through cuDSS inside Cadence Celsius. Applied Materials claimed up to 35x faster chamber simulations, imec said cuLitho delivered a 30% improvement in lithography process window, and Luminary said PhysicsNeMo lets it explore 100x more package configurations.

Every figure is vendor-reported and framed as early testing or a targeted release rather than a shipped production number; Nvidia does not publish the hardware or CPU baseline each was measured against. The story sits inside a steady drumbeat of coverage — our chips tracker has logged 313 stories over the last 90 days and Nvidia-specific coverage another 197.

cuPotential is in early access. NanoVDB is open source inside OpenVDB, and NPP ships through the Nvidia CUDA Toolkit.