NVIDIA Launches cuDSS Library to Accelerate Engineering and Scientific Simulations on GPUs

Image: NVIDIA · Source

NVIDIA cuDSS, now generally available, is a CUDA-X library designed to speed up large sparse system solves in simulations, enabling faster and more detailed engineering and scientific modeling by leveraging GPU acceleration.

NVIDIA has formally released cuDSS, a CUDA-X library that accelerates the mathematically intensive task of solving large sparse systems within engineering and scientific simulations by utilizing GPU computing power. These simulations are critical for predicting the performance of complex designs — from microchips to aircraft and industrial systems — before manufacturing or construction. As the complexity of these simulations grows, so do the computational demands, prolonging run times and limiting how extensively design variations can be tested.

cuDSS addresses this bottleneck by offloading the repeated sparse linear solves—often the most time-consuming part of these simulations—onto GPUs. This shift allows applications to significantly reduce time to results and increase throughput. Consequently, engineers and scientists can explore larger, more accurate models and evaluate a broader range of design possibilities at lower cost.

Several industry leaders have adopted cuDSS to enhance their simulation software workflow efficiency. Applied Materials achieved up to a 10x speedup in its Ginestra simulation platform for semiconductor materials and device reliability testing. COMSOL, combined with ZEISS SMT, realized a 4x acceleration in EUV lithography plasma simulations. Cadence integrated cuDSS in its Celsius thermal and structural simulation software, benefiting customers like Nokia by enabling faster complex chip design analysis. Dassault Systèmes reported an 8-10x speed improvement in electromagnetic solvers within its Abaqus toolkit. Siemens incorporated cuDSS to accelerate fluid-structure interaction simulations in Simcenter STAR-CCM+, and Synopsys used it in Sentaurus Device to hasten modeling of advanced logic, memory, and power chips for clients such as Infineon.

cuDSS is designed to scale seamlessly from individual workstations to large computing clusters. Developers can incorporate it into existing engineering, scientific, and optimization software using interfaces for C, C++, Python, and Fortran, thus bringing GPU acceleration to established tools used by engineers and researchers globally.

More information and demonstrations of NVIDIA cuDSS can be found at NVIDIA’s and Dassault Systèmes’ booths during GTC Berlin. The launch marks a significant advancement in enabling faster, more detailed simulation capabilities critical to multiple technology-driven industries.