NVIDIA officially released the cuDSS (CUDA Distributed Sparse Solver) library on October 26, 2026. As a key component of the CUDA-X software stack, its core function is to offload the computationally intensive, repeatedly occurring step of solving large sparse linear systems in engineering and scientific simulations to GPU acceleration. Such simulations are widely used in advanced chip design, aircraft structural analysis, industrial equipment modeling, and other applications; traditional CPU-based solvers have become a bottleneck as model complexity continues to surge.
According to NVIDIA’s official blog, cuDSS has been integrated and validated on multiple mainstream industrial simulation platforms: Applied Materials achieved up to 10× speedup on the Ginestra platform; ZEISS SMT leveraged COMSOL Multiphysics + cuDSS to reduce the simulation cycle for extreme ultraviolet (EUV) lithography plasma simulations to 6 days and 14 hours (a 4× improvement); tools such as Cadence Celsius, Dassault Systèmes Abaqus, Siemens Simcenter STAR-CCM+, and Synopsys Sentaurus Device have also reported solution‑time accelerations ranging from 8× to 10×. All cases are based on real‑world industrial workflows and do not involve theoretical benchmarks or simplified models.
cuDSS supports multi-language interfaces for C/C++/Python/Fortran, can be seamlessly integrated into existing simulation software, and natively scales horizontally from standalone workstations to large-scale GPU clusters. Its GA status means developers can download and use it directly without needing to request preview access. Currently, no specific API details, open-source license type, minimum GPU compute requirements, or third-party independent verification reports have been disclosed, nor has support for mixed precision or heterogeneous memory architectures been specified.