News

Version 1.1.0 Release

Added by Matthias Bach about 6 years ago

We are pleased to announce the release of HPL-GPU 1.1.0. This version includes two important new features:

  • Multi-GPU support — In combination with CALDGEMM 1.1.0 you can now use multiple GPUs per host
  • Heterogeneous system support — You can now utilize systems of different speed efficiently

These features are, however, still to be seen as experimental. Due to the limited of available multi-GPU systems and heterogeneous clusters with AMD GPUs they have not yet been tested extensively. If you experience problems in your specific combination feel free to contact the authors by joining ##caldgemm on irc.freenode.net or via the CALDGEMM mailing list.

You can download the tarball from the files section or get the source via tag v1.1.0 from the git repository. Installation instructions can be found in the wiki. If you are interested in some more in depth information you should have a look at the Technical Report that was published along with CALDGEMM 1.0.0.

Version 1.0.1 Release

Added by Matthias Bach about 6 years ago

Version 1.0.1 is a maintenance release that cleans up version 1.0.0 and fixes some issues found in certain system setups. Most changes are improvements to the build system and its documentation.

You can download the tarball from the files section or get the source via tag v1.0.1 from the git repository. Installation instruction can be found in the wiki. If you are interested in some more in depth information you should have a look at the Technical Report that was published along with CALDGEMM 1.0.0.

Version 1.0.0 Release

Added by Matthias Bach over 6 years ago

We are pleased to announce the release of HPL-GPU 1.0. This largely rewritten version of Linpack can sit on top of CALDGEMM, our high performance DGEMM library for AMD GPUs.

This is the code version that reached 285 Gflops on LOEWE-CSC, pushing the system to position 22 in this falls Top500 list.

Besides support for GPUs this version also features massive improvements to single process per node and initialization performance as well as lookahead, not to mention the many improvements in small details that help with things as debugging the code and the system.

You can download the tarball from the files section or get the source via tag v1.0.0 from the git repository. Installation instruction can be found in the wiki. If you are interested in some more in depth information you should have a look at the Technical Report that was published along with CALDGEMM 1.0.0.

Release Approaching

Added by Matthias Bach over 6 years ago

We are currently sorting out the last documentation and licensing issues. Once that is done, which will be before Christmas, the code will be published along with some technical documentation.

Also available in: Atom