NVIDIA Makes cuDSS Generally Available for GPU Simulation Workloads

NVIDIA made cuDSS generally available Wednesday, Oct. 21, bringing GPU acceleration to sparse linear solves used in engineering and scientific simulations. The software targets a repeated step in modeling work for chip design, aircraft development and industrial systems, where faster solves can shorten time to results.

cuDSS handles large sparse systems expressed as AX = B. Its documented capabilities include real and complex matrices, general, symmetric and positive-definite systems, multiple right-hand sides and batch processing. It also supports reordering, Schur complements, and hybrid CPU/GPU memory and execution.

The library runs on a single GPU, multiple GPUs or multiple computing nodes. C, C++, Python and Fortran interfaces connect it with engineering, scientific and optimization software. It can scale from workstations to larger computing clusters, and is aimed at workloads such as electronic design automation and computational fluid dynamics that use large sparse systems.

In Ginestra, Applied Materials’ sparse-solver workload ran up to 10 times faster with cuDSS than with CPU-only methods. For an EUV lithography plasma benchmark, ZEISS SMT completed the full COMSOL simulation in six days and 14 hours, with cuDSS delivering a measured fourfold speedup. Abaqus structural and electromagnetic solvers saw 8x to 10x speedups with cuDSS.

A June 23 update to version 0.8.0 added double-double arithmetic and expanded hybrid execution to analysis and multi-GPU or multi-node configurations. It also added support for mixed 64-bit row offsets and 32-bit column indices, host pointers for compressed sparse row input arrays, programmatic logging and NVTX profiling ranges.

Although cuDSS is generally available, its documentation was still labeled Preview as of Oct. 21 and noted that its APIs could change.