NVIDIA makes cuDSS generally available for simulation workloads
NVIDIA said on Wednesday it has made cuDSS generally available, positioning the CUDA-X library as a way to move a key sparse-solver step in engineering and scientific simulations onto GPUs. The company said the release is intended to reduce time to results and increase simulation throughput for workloads used in chip design, aircraft development and industrial systems.
Evidence and context
In the announcement published Oct. 21, 2026, NVIDIA described cuDSS as a library for accelerating large sparse systems, the math-heavy step that repeats throughout simulation workflows and can limit how many design iterations teams can test. The company said the software can help end users explore larger and more realistic models at lower cost by shortening simulation runtimes.
NVIDIA also said the library can scale from individual workstations to large computing clusters and can be integrated into existing engineering, scientific and optimization software through C, C++, Python and Fortran interfaces. In the same announcement, the company said developers can use cuDSS across single-GPU, multi-GPU and multi-node configurations.
The release follows earlier NVIDIA documentation that, as of Oct. 21, 2026, still described cuDSS as a direct sparse linear-solver library for systems of the form AX = B and noted that NVIDIA continued to label its documentation as Preview, with APIs subject to change. That documentation also listed support for real and complex sparse matrices, general, symmetric and positive-definite systems, multiple right-hand sides, batching, reordering, Schur complements, hybrid CPU/GPU memory and execution, and single-GPU, multi-GPU and multi-node operation.
NVIDIA cited several software partners as early users. The company said Applied Materials used cuDSS in Ginestra to accelerate sparse solves by up to 10x versus CPU-only approaches. It said COMSOL with NVIDIA cuDSS helped ZEISS SMT achieve a measured 4x speedup in an EUV lithography plasma simulation benchmark, finishing the full simulation in six days and 14 hours. NVIDIA also said Dassault Systèmes reported an 8-10x speedup using cuDSS within its Abaqus structural and electromagnetic solvers.
The background material also notes that version 0.8.0, released June 23, 2026, added double-double arithmetic, extended hybrid execution to analysis and multi-GPU or multi-node configurations, accepted host pointers for CSR input arrays across execution modes, supported mixed 64-bit row offsets and 32-bit column indices, and added programmatic logging and NVTX profiling ranges. The same material says cuDSS is relevant to electronic design automation, computational fluid dynamics, engineering simulation and optimization workloads involving large sparse systems.
NVIDIA did not provide pricing, market data or a next verified catalyst in the supplied material.