cuSOLVERMp: multi-node and multi-GPU linear solver

[ChrisRackauckas]

Dear Math Library Early Access members,

I'm excited to announce that we have released the initial delivery of cuSOLVERMp.

cuSOLVERMp v0.0.1 provides a distributed-memory multi-node and multi-GPU solution for solving systems of linear equations at scale! In the future, it will also solve eigenvalue and singular value problems.
Currently, supports LU solver, with and without pivoting (more to come).
The EA release of cuSOLVERMp targets P9 + IBM's Spectrum MPI.
The GA and future releases will also support x86_64 + OpenMPI.
The GA release will be hosted in HPC SDK.

For issues: https://forums.developer.nvidia.com/c/accelerated-computing/gpu-accelerated-libraries/12
For feedback/feature requests: [email protected]

Thanks,

Matthew Nicely, PhD | NVIDIA
Math Libraries Product Manager
(C) 205-492-5345 | [email protected]

[maleadt]

Only available for IBMP CPUs right now.

[kshyatt]

According to the docs this now supports x86 if someone wants to take a run at it.