NVIDIA introduces cuPotential for atomic-scale chip modeling

NVIDIA introduced cuPotential in early access as a CUDA-X library for atomic- and molecular-scale simulations used in semiconductor engineering. The company said early users accelerated workloads already optimized for GPUs by an additional 2–3x, but it did not publish the hardware, datasets, baselines or test methods behind that figure.

cuPotential is part of a broader effort to apply NVIDIA GPUs and software across chip design, manufacturing, inspection and production testing. NVIDIA and its partners described uses ranging from materials modeling and package thermals to virtual wafer processing and test-cell anomaly detection. Their speedup, yield and capacity figures are vendor claims rather than independently audited fab results. In a separate announcement, NVIDIA said it contributed Vera Rubin rack and compute-tray designs through open MGX.

Atomic-scale simulation

NVIDIA describes cuPotential as GPU-accelerated functions that developers can add to existing engineering software. The functions model electrical forces and other interactions among large numbers of atoms, allowing researchers to iterate on atomic-scale simulations used to evaluate semiconductor materials and processes.

The company named Synopsys as an early adopter for materials simulation. Synopsys describes QuantumATK as atomic-scale modeling software supporting methods from density-functional theory to classical force fields. NVIDIA also said Synopsys is using cuEST and cuDSS in semiconductor simulations, reporting potential acceleration of 30x for early QuantumATK results with cuEST and 12x in S-Device benchmarks with cuDSS. Synopsys separately reported a potential 30x result for cuEST-optimized QuantumATK quantum-chemistry workloads against open-source CPU models, but neither company supplied a complete reproducibility package.

cuPotential extends CUDA-X, NVIDIA’s collection of GPU-accelerated libraries for AI, data science and mathematics. NVIDIA did not disclose pricing, licensing terms, a public API location, supported hardware or a general-availability date for cuPotential.

Design and manufacturing workflows

Cadence said its AuraStack platform uses NVIDIA Blackwell and CUDA-X for planning, implementation and multiphysics analysis in printed-circuit-board and advanced-package design. NVIDIA said an AuraStack workflow using cuDSS with Cadence Celsius optimized package heat and stress eight times faster; the release did not provide the full benchmark setup.

NVIDIA said Siemens is using cuLitho to accelerate chip-pattern refinement and mask preparation in Calibre. It also said Siemens is developing agent-driven Solido characterization workflows with Nemotron and NeMo. The research record did not locate an accessible Siemens release that independently documented those adoption claims.

In manufacturing, Lam Research describes SEMulator3D as a physics-driven virtual-fabrication platform for modeling complete process flows. NVIDIA said Lam intends to integrate the sparse-data library NanoVDB into SEMulator3D, with release targeted by the end of 2026, and expects up to 30x acceleration for selected wafer-etch modeling. The integration and timing remain future plans, and the performance figure is a company claim.

NVIDIA also said Applied Materials is using Omniverse and CUDA-X for materials discovery, process development, chamber simulation and fab optimization, with chamber simulations running up to 35x faster. The announcement did not provide an independent audit of that result or quantify fab-level yield gains attributable to cuPotential.

Inspection and testing

The announcement names imec work on electron-microscope image denoising and computational lithography, KLA image-matching kernels, and Teradyne analysis for image-sensor testing and probe maintenance. NVIDIA reported improvements ranging from 30% to 85x and from hours to minutes across the inspection and test examples it named. The research record could not independently corroborate the individual partner measurements.

Advantest separately says its AI-powered anomaly-detection product combines ACS RTDI with NVIDIA AI technologies to analyze live semiconductor test-cell data for real-time monitoring and diagnostics. NVIDIA’s release characterizes such tools as ways to improve yield and production capacity, but it reports no independently audited fab-level gain caused by cuPotential.

cuPotential remains available only through early access. NVIDIA gave no date for general availability, while Lam’s planned NanoVDB integration into SEMulator3D is targeted for the end of 2026.