NVIDIA put a new CUDA-X library, cuPotential, into early access for simulating how atoms and molecules interact, and used the release to line up chip design and equipment companies behind what it calls AI-defined fabs: factories where AI helps engineers tune processes and chase down production problems.
NVIDIA says cuPotential gives developers GPU-accelerated functions they can add to existing software, including AI agents that run and iterate on molecular simulations. It sped up workflows that were already GPU-optimized by a further 2-3x in NVIDIA's early-access testing. Synopsys is named as an early adopter for materials simulation.
Other libraries carry the rest of the announcement. NanoVDB, an open source library for 3D data that lives inside OpenVDB, is slated to go into Lam Research's SEMulator3D process-modeling platform, where Lam expects select wafer-etch modeling to run up to 30x faster. That capability is targeted for release by the end of 2026. KLA, for its part, applied insights from its own production kernels to a new kernel in NVIDIA's NPP image-processing library, which NVIDIA reports ran 50x faster than earlier CPU versions with results comparable to KLA's hand-tuned implementations.
Synopsys reports early QuantumATK results with a potential 30x speedup using cuEST, plus a 12x gain on S-Device benchmarks using cuDSS. Applied Materials cites up to 35x faster chamber simulations from its mix of Omniverse and CUDA-X libraries. Cadence, Siemens, COMSOL, Teradyne and imec also appear in the announcement with their own work on NVIDIA libraries.
The pattern is that the gains come from the partners' own workloads, and several are labeled potential or expected rather than shipped. For the new library itself, the announcement offers only the 2-3x figure from NVIDIA's early-access testing. NVIDIA offers NPP through the CUDA Toolkit today, NanoVDB is already open, and cuPotential is the one piece still gated behind early access.
