NVIDIA cuDSS goes GA to speed engineering and scientific simulations
NVIDIA said on Oct. 21, 2026, that cuDSS is now generally available, positioning the CUDA-X library as a way to shift a key step in engineering and scientific simulations onto GPUs. The company says the software is designed to accelerate large sparse systems, which can be a limiting stage in workflows used to evaluate chips, aircraft and industrial systems before they are built.
Evidence and context
According to NVIDIA’s Oct. 21 blog post, cuDSS is intended to reduce time to results and increase simulation throughput by accelerating sparse solves across engineering, scientific and optimization software. NVIDIA said the library can be used through C, C++, Python and Fortran interfaces and can scale from a workstation to large computing clusters. The company also said developers can integrate cuDSS into existing software offerings rather than rebuilding workflows from scratch.
The release notes and background material supplied with the announcement describe cuDSS as a CUDA-X direct sparse solver library for systems of the form AX=B, with support for real and complex general, symmetric and positive-definite sparse matrices; single and multiple right-hand sides; uniform and nonuniform batching; single, double and double-double precision; and 32-bit and 64-bit indices. Execution is split into analysis, numerical factorization and solving, with refactorization and iterative refinement available. The material also lists numerical pivoting, reorderings, Schur complements, memory estimates and custom memory pools among the controls available to users.
NVIDIA’s general-availability announcement said several software vendors are already integrating cuDSS into their tools. Applied Materials said cuDSS accelerated sparse solves in Ginestra, its physics-based, defect-centric simulation platform, by up to 10x versus CPU-only approaches. COMSOL said ZEISS SMT achieved a measured 4x speedup in an EUV lithography plasma simulation benchmark using COMSOL Multiphysics with NVIDIA cuDSS on NVIDIA accelerated computing, completing the full simulation in six days and 14 hours. Cadence said it uses cuDSS in Celsius to accelerate heat, structural stress and warpage simulations in advanced chips and 3D chip designs. Dassault Systèmes said it reports an 8-10x speedup using cuDSS within Abaqus structural and electromagnetic solvers. Siemens said it uses cuDSS in Simcenter STAR-CCM+ to accelerate structures in fluid-structure interaction workflows, while Synopsys said it uses cuDSS in Sentaurus Device to speed simulations of advanced logic, memory and power chips.
The performance figures are vendor-reported and workload-specific, not general guarantees. The supplied background also notes a documentation discrepancy: NVIDIA announced general availability on Oct. 21, 2026, but documentation inspected the same day still carried a Preview label and listed version 0.8.0, which the archive identifies as a June 2026 release. The material does not include pricing, market data or a next dated catalyst.