NVIDIA has made cuDSS generally available, giving engineering-software developers a GPU library for one of their most demanding tasks: solving large, sparse systems of equations. The October 21 release targets calculations inside scientific and industrial simulations, where a faster numerical solve can shorten the wait between changing a design and seeing the result.[1]
NVIDIA’s examples span chipmaking and electromagnetic design. Applied Materials reports up to a tenfold improvement in sparse solves for its Ginestra simulation software. A ZEISS workflow using COMSOL for extreme-ultraviolet plasma modelling ran four times faster, shortening a six-day calculation to about a day and a half. The results come from NVIDIA and its partners and describe the particular workloads they tested.[1]
Why sparse equations consume so much time
Simulations divide a physical problem into many interacting parts: points in a mesh, circuit elements or regions of a material. Each part usually connects to only a small fraction of all the others. The resulting matrix contains many zeros—hence “sparse”—but finding the solution can still demand substantial memory and computation.
A direct solver works through a factorisation of that matrix and then uses it to recover the solution. Accelerating that stage is useful when it dominates the run. It also changes how an engineering team can work: shorter waits make it more practical to compare several geometries or operating conditions rather than commit early to one expensive simulation.
The workloads behind the reported speedups
| Software or partner | Workload | Reported result |
|---|---|---|
| Applied Materials / Ginestra | Sparse linear solves in semiconductor simulation | Up to 10× faster than the CPU comparison |
| ZEISS / COMSOL | EUV plasma simulation | About 4× faster; six days reduced to roughly 1.5 days |
| Dassault Systèmes / Abaqus | Electromagnetic simulation | 8–10× acceleration in the reported workload |
Integration determines the end-to-end gain
cuDSS is intended to fit into existing scientific software rather than require users to rebuild a complete simulation stack. NVIDIA lists access through C and C++, Python and Fortran, and describes work with engineering-software suppliers including Cadence, Siemens and Synopsys. Those integrations are the route by which the library reaches engineers who use applications rather than write numerical kernels.[1][2]
The whole application still has to create its mesh, prepare inputs, move data and process the solution. A tenfold reduction in one stage cannot produce a tenfold reduction overall unless that stage accounts for almost all the original runtime. Developers therefore need the profile of their actual application to decide where GPU acceleration will have the greatest effect.
For users, the practical payoff is more design iterations within the same schedule. The library’s release matters because it makes a specialised numerical acceleration available inside a wider range of established engineering workflows, where time saved on an equation solve can become time spent testing a better design.
Sources & further reading
- NVIDIA cuDSS Now Generally Available to Accelerate Engineering and Scientific Simulations NVIDIA Blog · 21 Oct 2026 · Accessed 25 Oct 2026
- Europe’s Industrial Leaders Reinvent Engineering and Manufacturing With NVIDIA NVIDIA Blog · 21 Oct 2026 · Accessed 25 Oct 2026