cuSOLVERMp: A Distributed-Memory Multi-Node Dense Linear Algebra Library

Welcome to the cuSOLVERMp library documentation.

cuSOLVERMp is a distributed-memory multi-node and multi-GPU solution for solving systems of linear equations at scale, available through the HPC SDK.

Download: https://developer.nvidia.com/nvidia-hpc-sdk-downloads

Feedback: Math-Libs-Feedback@nvidia.com

Key Features

  • LU factorization, with and without partial, for general matrices distributed between GPU nodes as well as solving routines for single right hand side for computed factorization.
  • Cholesky factorization, for symmetric and Hermitian matrices, with corresponding solving routines for 1 right hand side.

Support

  • Supported SM Architectures : SM 7.0, SM 8.0
  • Supported OSs : RHEL 7/8, Ubuntu 20.04/18.04
  • Supported CPU Architectures : x86_64, ARM64, OpenPOWER

Prerequisites

  • Dependencies : cuda, cudart, cublas, cublasLt, cusolver, nvidia-ml, cal, cusolvermp.h headers