NVIDIA made its cuDSS CUDA-X library generally available, targeting sparse-system calculations used in engineering and scientific simulations.
NVIDIA made cuDSS generally available on October 21, presenting the CUDA-X library as a way to accelerate sparse-system solving in engineering and scientific simulation software. These calculations recur throughout simulation workloads and can limit how many designs teams test. The library uses GPUs for this computational step, and NVIDIA says it can help applications shorten run times and increase simulation throughput.
The release includes examples from software providers already integrating cuDSS. Applied Materials reports up to a 10x acceleration for sparse solves in its Ginestra simulation platform compared with CPU-only approaches. COMSOL and ZEISS SMT report a measured 4x speedup in an extreme-ultraviolet lithography plasma simulation benchmark; the full simulation took six days and 14 hours. Dassault Systèmes reports an 8–10x speedup within certain Abaqus electromagnetic solvers. These are company-reported results from specific applications, not a single standardized comparison across all workloads.
Other integrations span chip design and industrial simulation. Cadence uses cuDSS in its Celsius software for heat, structural-stress and warpage simulations, while Siemens incorporates it into Simcenter STAR-CCM+ for fluid-structure interaction workflows. Synopsys uses the library in its Sentaurus Device simulator for advanced logic, memory and power-chip work.
Developers can integrate cuDSS into existing engineering, scientific and optimization software, with interfaces for C, C++, Python and Fortran. NVIDIA says the library can scale from individual workstations to large computing clusters. The practical effect will depend on how well a particular application and workload can use GPU acceleration; the examples in the announcement illustrate results reported by the vendors using it.