Introducing AmgX - GPU-accelerated Algebraic Multigrid Accelerator

After 3 years of development, I’m very pleased to announce the beta availability of NVIDIA’s AmgX library:

AmgX is, as the name implies, a library of GPU-accelerated sparse linear solver, including 2 flavors of algebraic multigrid (AMG), Krylov solvers, preconditioners like ILU0, and a flexible composition system which lets you combine all of these building blocks to create custom solutions for your problems. AmgX works with a single GPU, multiples GPUs, and even interoperates with MPI for when you have a really big problem to solve across a cluster.

To get a beta copy of AmgX, you need to be a CUDA registered developer - apply here if you aren’t already:

https://developer.nvidia.com/registered-developer-programs

Then go to the main regdev website, download AmgX (link is off the regdev page), give it try, and let us know what you think. Please post any questions or issues on the forum and we’ll try to help you out as best as we can.

Happy solving!

Jon Cohen
Senior Manager, CUDA Libraries and Algorithms

When will AmgX be available for CUDA 6?

I wish there is more communication from the developers…

We installed amgX on a cluster where several nodes have K40m GPUs. We successfully ran several tests. We have several questions:

Our matrices fit into available K40m memory, they do not need partitioning, we rather need to be able to solve many matrices in parallel.

  1. Is it possible to use amgX with Intel MPI instead of OpenMPI?
  2. Is it possible to run a problem with MPI ranks solving their matrices independent of each other in parallel?
  3. If an MPI rank is running on a node without GPUs, will amgX be able to solve matrices locally on available CPUs?

We’ll appreciate any information or links to information that could help us to integrate amgX into our simulation software.

Is AmgX project still active?

Can you show me an example of a matrix file that uses AMGX extensions to matrix market format for inclusion of the right-hand-side?

Thanks in advance !!

When will AmgX be available for CUDA 7? Can AmgX based on the cuda 6.5 setup on a cuda 7.0 system?