AIPS BOOM

cuDSS

3 min read

cuDSS

Type: Technology Tags: CUDA, NVIDIA, GPU, Sparse Solver, Direct Solver, Math, HPC, Autonomous Driving Related: cuSPARSE, cuSOLVER, cuBLAS, AmgX, nvmath-python, NCCL, CUDA-Graphs, NVIDIA-CUDA Sources: https://docs.nvidia.com/cuda/cudss/index.html, https://docs.nvidia.com/cuda/cudss/release_notes.html Last Updated: 2026-04-29

Summary

cuDSS is NVIDIA’s preview CUDA library for GPU-accelerated direct sparse solvers. It solves linear systems of the form AX = B where A is sparse and B/X may be vectors or matrices, with flexible matrix properties, solver configuration, CUDA stream execution, and single-GPU, multi-GPU, and multi-node modes.

Detail

Purpose

Direct sparse solvers are important when simulation, engineering, robotics, circuit, and control workloads require robust sparse linear solves rather than iterative convergence behavior. cuDSS brings this direct-solver workflow into the CUDA math stack with analysis, factorization, solve, refactorization, batching, and distributed execution paths.

Key capabilities

  • Sparse matrix types include real/complex, general, symmetric, positive-definite, and complex symmetric cases.
  • Analysis, numerical factorization, solve, optional refactorization, solve sub-phases, permutations, and iterative refinement.
  • Non-uniform batching for different systems and uniform batching for systems with the same sparsity pattern.
  • Single and multiple right-hand sides with single/double precision values and int/int64_t indices.
  • Multiple reordering and factorization algorithms, numerical pivoting controls, and memory estimates after analysis.
  • User-defined device memory handlers, memory pools, threading layers, and distributed communication layers.
  • Schur complement mode, hybrid host/device memory mode, and hybrid host/device execution mode.
  • Multi-GPU multi-node (MGMN), single-node multi-GPU (MG), and multi-threaded (MT) modes.
  • CUDA Graphs support and partially asynchronous APIs where host memory/execution modes are not enabled.
  • Optional deterministic computation for bit-wise reproducibility on the same hardware/software/input stack.

Current release signals

  • cuDSS v0.7.1 enabled user-provided elimination tree reuse in combination with user permutation data.
  • cuDSS v0.7.0 enabled single-node multi-GPU mode, int64_t sparse indices, solve sub-phases, Schur complement mode, deterministic mode, and CUDA 13 / newer Blackwell architecture support.
  • NVIDIA still marks cuDSS as preview, so API compatibility can change in later releases.

Use Cases

  • Autonomous driving simulation and sensor fusion
  • Industrial process simulation and control
  • Structural analysis and finite element methods
  • Circuit simulation
  • Real-time control systems requiring robust linear algebra

Hardware Requirements

  • NVIDIA GPU with supported SM architecture starting with Pascal, according to the matching CUDA Toolkit support matrix.
  • Linux and Windows support.
  • x86_64, Arm SBSA, and Arm aarch64/Jetson support for Orin and Thor devices.
  • MGMN communication backends include prebuilt OpenMPI 4.x and NCCL 2.x layers plus user-defined GPU-aware stream-aware backends.
  • MT threading backends include GNU OpenMP on Linux, VCOMP on Windows, and user-defined threading layers.

Language Bindings

  • C and C++ (primary API)
  • Python (via nvmath-python)
  • Sample code available via NVIDIA CUDALibrarySamples repository

Connections

  • cuSPARSE - sparse matrix context for GPU sparse linear algebra.
  • cuSOLVER - cuDSS docs include a comparison point to cuSolverSp and cuSolverRf.
  • cuBLAS - cuDSS has documented runtime dependency on cuBLAS in current release history.
  • AmgX - iterative algebraic multigrid alternative to cuDSS direct sparse solves.
  • nvmath-python - Python-facing sparse direct solver access.
  • NCCL - one documented MGMN communication backend for distributed cuDSS operation.
  • CUDA-Graphs - cuDSS documents CUDA Graphs support for solver workflows.
  • NVIDIA-CUDA - CUDA platform and toolkit context for cuDSS.

Source Excerpts

  • NVIDIA describes cuDSS as a preview GPU-accelerated direct sparse solver library for AX = B systems.
  • Current docs list single-GPU, single-node multi-GPU, and multi-GPU multi-node execution configurations.

Resources

Graph View

Backlinks