NVIDIA cuDSS Launches With Up to 10x GPU Simulation Speedups
NVIDIA cuDSS is now generally available, accelerating GPU sparse solves for engineering simulations with measured and reported speedups of 4x to 10x.
Summary
NVIDIA made cuDSS generally available on October 21, 2026. The CUDA-X library moves repeated large sparse-system solves, a bottleneck in simulations for advanced chips, aircraft and industrial systems, from CPUs to GPUs. It scales from individual workstations to computing clusters and offers C, C++, Python and Fortran interfaces for existing engineering, scientific and optimization software. NVIDIA says higher throughput can support larger, more realistic and accurate models, broader design exploration, faster results and lower costs.
Applied Materials accelerated sparse solves in its Ginestra semiconductor materials and device-reliability platform by up to 10x versus CPU-only methods. ZEISS SMT measured a 4x speedup for a COMSOL Multiphysics EUV lithography plasma benchmark, completing the full simulation in six days and 14 hours. Dassault Systèmes reports 8x to 10x gains in Abaqus structural and Computer Simulation Technology electromagnetic solvers.
Cadence uses cuDSS in Celsius to model heat, structural stress and warpage in advanced and 3D chips, helping Nokia analyze complex designs. Siemens integrates it into Simcenter STAR-CCM+ for fluid-structure interaction, while Synopsys uses it in Sentaurus Device to accelerate advanced logic, memory and power-chip simulations for customers including Infineon. NVIDIA and Dassault Systèmes are demonstrating cuDSS at GTC Berlin.
Positives
- Applied Materials achieved up to 10x faster Ginestra sparse solves than CPU-only approaches for semiconductor materials and device-reliability exploration.
- ZEISS SMT measured a 4x COMSOL Multiphysics speedup for an EUV lithography plasma simulation.
- Dassault Systèmes reports 8x to 10x acceleration across Abaqus structural and Computer Simulation Technology electromagnetic solvers.
- cuDSS scales from individual workstations to large clusters and supports C, C++, Python and Fortran.
- Cadence, Siemens and Synopsys integrations extend GPU acceleration to chip design, fluid-structure interaction and semiconductor-device simulation.
Risks & concerns
- ZEISS SMT's accelerated EUV lithography plasma benchmark still required six days and 14 hours to complete.
- Cadence, Siemens and Synopsys provide no specific speedup figures for their cuDSS integrations.
- Dassault Systèmes' 8x to 10x gain is company-reported, while only the ZEISS SMT result is explicitly described as measured.
- cuDSS accelerates the recurring sparse-solve component rather than every stage of an engineering or scientific simulation.