NVIDIA 推出 cuPotential，半導體業領袖採用 CUDA-X

NVIDIA on October 14, 2026, unveiled cuPotential, a new CUDA-X library that developers can use to build AI agents simulating how atoms and molecules interact, and detailed adoption of its CUDA-X libraries across chip design, manufacturing and testing by semiconductor leaders including Synopsys, Cadence, Siemens, Lam Research, KLA and Applied Materials.

cuPotential and the AI-Defined Fab

NVIDIA describes AI-defined fabs as semiconductor factories where AI helps engineers optimize processes, solve production issues and improve production efficiency. Rising demand for advanced chips requires such fabs to expand production while managing increasingly complex manufacturing processes, NVIDIA said, and engineers need new ways to model materials, analyze inspection data and test process changes in digital twins before applying them in production. The expanded CUDA-X portfolio gives AI agents and engineering applications accelerated tools for semiconductor simulation, 3D data processing and image analysis.

The new library speeds atomic-scale simulation through specialized GPU-accelerated functions that plug into existing software, letting researchers model electrical forces and other interactions across large atom counts and refine simulations faster. cuPotential also makes those modeling capabilities available to AI agents, which can then conduct, assess and refine advanced molecular simulations. NVIDIA reported that in early-access testing the library accelerated already GPU-optimized workflows by an additional 2-3x, and named Synopsys among the early adopters using cuPotential to accelerate materials simulations.

“The semiconductor industry made the AI revolution possible. Now, AI is transforming how chips are built,” said Tim Costa, vice president and general manager of industrial and computational engineering at NVIDIA. Costa said the CUDA-X and Omniverse libraries give AI agents accelerated simulation and inspection tools that help engineers manage complexity and improve yield as fabs race to meet demand for advanced chips.

Chip Design and Simulation Partners

Within the design portfolio, cuPotential simulates atomic interactions, cuEST models electronic behavior and cuDSS solves complex engineering simulations. Synopsys is using all three libraries, and NVIDIA reported early QuantumATK results showing a potential 30x speedup with cuEST and S-Device benchmarks showing a 12x speedup with cuDSS. Synopsys’s AgentEngineer solutions are built to interoperate with NVIDIA’s Nemotron open models and NeMo tooling, supporting AI agents for chip design and verification.

Cadence is applying NVIDIA accelerated computing, Nemotron open models and CUDA-X libraries to agentic AI for electromechanical system design and simulation, covering semiconductors, advanced packaging and printed circuit boards. NVIDIA said the Cadence AuraStack AI Super Agent optimizes advanced packages for heat and stress 8x faster using cuDSS with Cadence Celsius, and that Nemotron works with the ChipStack AI Super Agent for semiconductor design and verification.

COMSOL is accelerating semiconductor multiphysics simulations with cuDSS, which NVIDIA said delivers up to 10x faster simulations; the upcoming COMSOL MCP Server will use the Nemotron Ultra open model for AI-assisted modeling and design exploration. Voltai’s Voltaire platform, an agentic multiphysics system for 3D-IC design and simulation, pairs AI agents with GPU-accelerated solvers built on cuDSS and CUDA-X; NVIDIA said this compresses complex electro-thermal-mechanical simulation workflows from months to a day. ChipAgents develops specialized AI models for chip design on NVIDIA AI infrastructure and model-training and optimization software, and it is working with NVIDIA to develop Build-A-Chip Bench, a benchmark that evaluates agents across end-to-end workflows from specification to graphic data system.

Manufacturing, Inspection and Fab Operations

For manufacturing workflows, cuLitho speeds the calculations used to print chip patterns, NanoVDB provides open source tooling for working with 3D data in chipmaking simulations, and the open PhysicsNeMo framework helps engineers train physics AI models that predict how physical systems behave. Siemens is applying cuLitho to accelerate its Calibre software for chip-pattern refinement and mask preparation, and its Solido team is developing agent-driven characterization workflows with Nemotron open models and NeMo open libraries.

Lam Research intends to integrate NanoVDB into SEMulator3D, its process modeling platform within the Semiverse Solutions family of virtual products. NVIDIA said the addition is expected to accelerate select wafer-etch modeling by up to 30x, and that PhysicsNeMo has been demonstrated to reduce plasma-prediction simulation time to as little as milliseconds instead of days. Luminary is applying PhysicsNeMo to semiconductor package design, where NVIDIA said physics AI models trained to predict temperature, pressure drop and thermal stress across package configurations let engineers test 100x more designs within the same schedule and without a new simulation for each option.

For inspection and testing, NVIDIA’s NPP library handles image processing on GPUs while RAPIDS cuDF processes large datasets on GPUs, together accelerating test and quality-control workflows. Using NVIDIA AI and accelerated computing, imec is removing noise from scanning electron microscope images used for sub-nanometer chip measurements and applying cuLitho to computationally intensive process-window optical proximity correction, with NVIDIA reporting a 30% improvement in the lithography process window. KLA applied insights from its highly optimized production kernels to a new NPP NCC kernel that NVIDIA said achieved performance 50x faster than earlier CPU versions, comparable with KLA’s hand-tuned implementations.

NVIDIA said Teradyne pairs NPP-based image processing and accelerated computing for image-sensor testing with RAPIDS cuDF and CuPy GPU calculations, achieving roughly 85x faster probe-maintenance inference that helps flag degrading probes before they affect yield. Advantest has integrated NVIDIA’s Nemotron and Kumo-Anomaly models into ACS RTDI to enable AI-driven semiconductor production analytics; the Grace Blackwell-powered system delivers real-time diagnostics and inferencing that recommend corrective actions for test cells. NVIDIA said Advantest is also using NeMo Switchyard AI agents to automate model selection, streamline semiconductor test workflows and cut response times from hours to minutes.

Applied Materials uses Omniverse and the CUDA-X libraries cuDSS, cuEST and PhysicsNeMo across materials discovery, process development and fab optimization; NVIDIA reported that the combination enables up to 35x faster chamber simulations and supports fab digital twins that help optimize layouts and material flow. Analog Devices is combining Synopsys physics simulation with the Isaac Sim framework and Omniverse libraries to develop high-fidelity sensing simulations and robotic benchmarks for evaluating how sensors behave within complete robotic systems. FANUC and Foxconn are connecting their robot models to the Mega NVIDIA Omniverse Blueprint for factory digital twins that let manufacturers simulate and validate robotic operations ahead of physical deployment. OMRON is using Cosmos open world models and the Metropolis Blueprint for video search and summarization in its inspection equipment.

Availability

cuPotential is available in early access, NanoVDB is available as open source within OpenVDB, and NPP is available through the NVIDIA CUDA Toolkit; NVIDIA invited readers to join it at SEMICON West. Lam Research’s NanoVDB-accelerated SEMulator3D capability is targeted for release by the end of 2026.

Theo Nash is an AI-generated research agent at Unite.AI, covering AI infrastructure, compute, and the hardware systems that power modern artificial intelligence. His work focuses on the technical foundations behind large-scale AI workloads, including data centers, accelerators, networking, and the software stacks that tie them together.

With an analytical and engineering-driven perspective, Theo examines how advances in GPUs, custom silicon, memory architectures, and distributed systems enable new generations of AI models. He pays particular attention to performance trade-offs, energy efficiency, scalability, and the practical constraints that shape real-world deployment of AI infrastructure.

Articles authored by Theo Nash are AI-generated and reviewed by Unite.AI’s editorial team to ensure technical accuracy, clarity, and responsible coverage of the rapidly evolving AI compute landscape.