Incomplete LU and Cholesky Preconditioned Iterative Methods
Type: Guide Tags: NVIDIA, CUDA, cuSPARSE, cuBLAS, iterative solvers, preconditioning, HPC Related: cuSPARSE, cuBLAS, cuSOLVER, AmgX, NVIDIA-CUDA, NVIDIA-Deep-Learning-Performance Sources: https://docs.nvidia.com/cuda/incomplete-lu-cholesky/index.html Last Updated: 2026-04-29
Summary
NVIDIA’s Incomplete-LU and Cholesky Preconditioned Iterative Methods whitepaper explains how to use cuSPARSE and cuBLAS for GPU-accelerated preconditioned iterative solvers. It focuses on sparse linear systems and Krylov methods such as BiCGStab and Conjugate Gradient.
Detail
The whitepaper addresses large sparse linear systems from computational mechanics, atmospheric modeling, geophysics, biology, circuit simulation, and other scientific computing domains. It describes using sparse matrix-vector multiplication and sparse triangular solve in cuSPARSE, with dense/vector operations in cuBLAS, to accelerate iterative methods.
This is a guide page rather than a library page: the canonical library pages remain cuSPARSE, cuBLAS, cuSOLVER, and AmgX.
Connections
- cuSPARSE - sparse matrix-vector multiplication and triangular solve primitives.
- cuBLAS - vector and dense operations used inside iterative methods.
- cuSOLVER - broader solver library for dense and sparse systems.
- AmgX - GPU-accelerated algebraic multigrid and Krylov solver library.
- NVIDIA-CUDA - CUDA parallel programming model underlying the implementation.
Source Excerpts
- NVIDIA’s whitepaper focuses on incomplete-LU and Cholesky preconditioning for iterative methods using cuSPARSE and cuBLAS.