The Pulse
NVIDIA Makes cuDSS Generally Available for Simulation Software
NVIDIA says cuDSS is generally available, giving developers a GPU-accelerated library for sparse linear systems used in engineering and scientific simulations. The announcement names six companies integrating the library and reports perform
AI.info Team · Written with AI · fact-checked ·
NVIDIA made cuDSS generally available on October 21, putting its GPU-accelerated sparse solver into the hands of developers building engineering and scientific simulation software. The library targets a repeated calculation in many simulations: solving large systems of equations whose matrices contain mostly zeros.
NVIDIA says moving that work to its GPUs can shorten time to results and let teams run more simulations or examine more detailed models. The announcement names six companies already integrating cuDSS, spanning semiconductor design, multiphysics modeling and fluid-structure analysis.
cuDSS divides sparse solving into three stages
Many simulations repeatedly solve sparse linear systems as they model how a design or physical process behaves. NVIDIA’s documentation describes cuDSS as a library for solving those systems, with an execution flow divided into analysis, numerical factorization and solving. The analysis stage includes reordering and symbolic factorization; the library can also support refactorization for applicable workloads. In the default configuration, reordering runs on the host, while symbolic factorization, numerical factorization and solving run on the device. NVIDIA notes that hybrid host/device execution can change where some of these phases run.
That makes cuDSS a component for software developers to incorporate, rather than a standalone simulation application. NVIDIA says developers can use C, C++, Python and Fortran interfaces, and that the library can scale from individual workstations to large computing clusters. Its documentation lists single-GPU, multi-GPU and multi-GPU, multi-node configurations.
Six companies and their software are named
NVIDIA lists Applied Materials and its Ginestra platform, COMSOL and COMSOL Multiphysics, Cadence and Celsius, Dassault Systèmes and its Abaqus structural and CST electromagnetic solvers, Siemens and Simcenter STAR-CCM+, and Synopsys and Sentaurus Device. The use cases range from semiconductor materials and device reliability to heat, structural stress, warpage and fluid-structure interaction.
The company’s performance figures come from particular customer and partner workloads, not a single standardized comparison. NVIDIA says Applied Materials accelerated sparse solves in Ginestra by up to 10 times versus CPU-only approaches. It also says ZEISS SMT measured a fourfold speedup in an EUV lithography plasma simulation using COMSOL Multiphysics with NVIDIA accelerated computing; the full simulation took six days and 14 hours.
Dassault Systèmes reports speedups of eight to 10 times using cuDSS in Abaqus structural and CST electromagnetic solvers, according to NVIDIA. The announcement does not provide benchmark hardware or enough workload detail to compare those results directly with one another. The figures describe reported results for specific applications, rather than a general performance guarantee.
General availability puts integration in focus
For engineering software vendors, the release offers a GPU solver they can add to tools customers already use. Whether that shortens an entire simulation depends on how much time the workload spends solving sparse systems, as well as the model and computing setup; accelerating one stage does not automatically produce the same gain across every workflow.
NVIDIA’s announcement centers on adoption through existing simulation products, rather than a new end-user application. Its named integrations connect cuDSS to software used for chip design, materials analysis and industrial simulation, while the published performance examples remain tied to the workloads and comparisons each company reported.