Nvidia announced on 21 October the general availability of cuDSS, a GPU library for engineering and scientific simulations.
A sparse system is a large grid of numbers in which most positions are empty. Solving it is a repeated calculation that Nvidia says can limit how many designs a simulation team tests; cuDSS shifts that work to GPUs.
Testing a chip design before it is built could take less time where solving those equations sets the pace of a simulation, making it feasible to examine more designs.
Nvidia reports that Applied Materials sped up sparse solves in its Ginestra simulation platform by up to 10x against CPU-only approaches. Nvidia says ZEISS SMT measured a 4x speedup with cuDSS and COMSOL Multiphysics while modelling plasma for EUV lithography; the complete simulation took six days and 14 hours. In a 14 October release, Nvidia reported that COMSOL was delivering simulations up to 10x faster with cuDSS.
Nvidia says Cadence, Siemens and Synopsys also use cuDSS in simulation software. Developers can integrate the library through C, C++, Python or Fortran interfaces, on systems ranging from individual workstations to computing clusters.
