NVIDIA has announced general availability of cuDSS, a CUDA-X library designed to accelerate the solving of large sparse systems on GPUs. These systems are a recurring computational step in engineering and scientific simulations, and NVIDIA says they often limit how many design variations teams can test before building physical prototypes.
The library is meant to drop into existing simulation software, with interfaces for C, C++, Python, and Fortran. NVIDIA reports that several vendors have already integrated cuDSS. Applied Materials saw up to 10x speedup in its Ginestra platform; Dassault Systèmes reports 8-10x speedup in Abaqus and electromagnetic solvers; and COMSOL, used by ZEISS SMT, achieved a 4x speedup in an EUV lithography plasma simulation benchmark.
Other integrations include Cadence's Celsius, Siemens' Simcenter STAR-CCM+, and Synopsys' Sentaurus Device. NVIDIA also says cuDSS scales from individual workstations to large computing clusters. Because this is a single vendor announcement, the speedup figures are as reported by NVIDIA and its partners, not independent benchmarks.