NVIDIA adds cuPotential and more CUDA-X libraries for AI-defined fabs: Lam Research targets up to 30x faster etch modelling and KLA's new inspection kernel runs 50x faster than earlier CPU versions

Start with NVIDIA's own definition. The release describes AI-defined fabs as semiconductor factories where AI helps engineers optimise processes, solve production issues and improve production efficiency. As demand for advanced chips rises, such fabs must expand production while managing increasingly complex manufacturing processes, and engineers need new ways to model materials, analyse inspection data and test process changes in digital twins before applying them in production. The October 14 release says semiconductor leaders are adopting NVIDIA CUDA-X libraries to advance AI-defined fabs, including the new cuPotential library, which developers can use to build AI agents that simulate how atoms and molecules interact. Tim Costa, vice president and general manager of industrial and computational engineering at NVIDIA, says the semiconductor industry made the AI revolution possible and AI is now transforming how chips are built. The release names Applied Materials, Cadence, KLA, Lam Research, Siemens and Synopsys among the companies using NVIDIA's accelerated computing and AI stack across chip design, manufacturing and testing.

Three libraries map to three concrete bottlenecks. cuPotential packages atomic-scale simulation into GPU-accelerated functions developers can add to existing software, so researchers can model electrical forces and other interactions across large numbers of atoms; NVIDIA says early-access testing showed the library accelerated already GPU-optimised workflows by an additional 2 to 3x, with Synopsys among the early adopters. NanoVDB, an open source library for working with 3D data, is described as accelerating chipmaking simulation for AI-defined fabs and helping Lam Research achieve 30x faster simulation; Lam Research intends to integrate NanoVDB into SEMulator3D, its advanced semiconductor process modelling platform, where the addition is expected to accelerate select wafer-etch modelling by up to 30x, with the capability targeted for release by the end of this year. NVIDIA NPP, a library for processing images on GPUs, is credited with accelerating inspection: KLA applied insights from its highly optimised production kernels to a new NPP NCC kernel that achieved performance levels 50x faster than earlier CPU versions, comparable with KLA's hand-tuned implementations.

Laid out partner by partner, the numbers land densely. Synopsys is using cuPotential, cuEST and cuDSS to accelerate semiconductor simulations, with early QuantumATK results showing a potential 30x speedup using cuEST and S-Device benchmarks showing a 12x speedup using cuDSS. Cadence is bringing agentic AI to the design and simulation of electromechanical systems, with its AuraStack AI Super Agent optimising advanced packages for heat and stress 8x faster using cuDSS with Cadence Celsius. COMSOL is accelerating semiconductor multiphysics simulations with cuDSS for up to 10x faster runs, and says its upcoming COMSOL MCP Server will use the Nemotron Ultra open model. Voltai's Voltaire platform pairs AI agents with GPU-accelerated solvers for 3D-IC design, compressing complex electro-thermal-mechanical simulation workflows from months to a day. Applied Materials is using Omniverse and CUDA-X libraries including cuDSS, cuEST and PhysicsNeMo across materials discovery, process development and fab optimisation, enabling up to 35x faster chamber simulations. On the inspection side, Teradyne uses NPP, RAPIDS cuDF and CuPy to deliver roughly 85x faster probe-maintenance inference, while Advantest has integrated Nemotron and Kumo-Anomaly into ACS RTDI and uses NeMo Switchyard agents to cut response times from hours to minutes. imec is denoising scanning electron microscope images and using cuLitho for computationally intensive process-window optical proximity correction, with the stated result a 30% improvement in the lithography process window.

Some boundaries are worth stating. First, every multiplier here comes from NVIDIA and its partners in a joint release; these are vendor performance claims, not independently reproduced third-party results, and cuPotential is in early access with no published general-availability date. Second, availability is specific and needs to be read separately: NanoVDB ships as open source within OpenVDB and NPP comes through the NVIDIA CUDA Toolkit, while Lam Research's NanoVDB-accelerated SEMulator3D capability is only targeted for release by the end of this year, which makes it roadmap rather than shipping capability. Third, no Intel or Samsung foundry-side activity appears in the partner list, and the release offers no capacity or yield figures, the metrics fabs live and die by. The more accurate framing is that this is progress at the software toolchain layer, not a production-capacity result.