HPL-GPU 1.0 Released For GPU Super Computing

Matthias Bach, a researcher and developer on high-performance computing with the Goethe Universitaet Frankfurt am Main, has written in to inform us of this German university's release of HPL-GPU 1.0. The HPL-GPU software package is a rewritten version of the LINPACK library that is re-engineered to run atop CALDGEMM, which is a DGEMM implementation developed at this university designed to run atop the latest AMD graphics processors.

The HPL-GPU 1.0 package is the code that's running atop LOEWE-CSC, which is a 832-node CPU/GPU cluster at Frankfurt University's Center for Scientific Computing. This LOEWE-CSC cluster HPL performance was measured at 285 TeraFLOPS this year making it one of the fastest super-computers in the world. LOEWE-CSC was ranked 22nd on this year's top 500 super-computer list and took the 8th spot on the green 500 list for the most energy efficient super-computers. This Frankfurt super-computer put outs out a massive 741 MFlops per Watt.

This open-source HPL-GPU code is designed to run on ATI/AMD Radeon graphics hardware under Linux when using the proprietary Catalyst driver (namely Catalyst 10.9 or newer) and using the latest ATI Stream SDK for its CAL support. The code is also engineered for the Radeon HD 5000 "Evergreen" series of graphics cards.

The release announcement (and download links/instructions) for this high-performance LINPACK library for the GPU can be found at uni-frankfurt.de.

Matthias Bach also mentioned to us in his email, "The really interesting thing however is, that this code can run AMD GPUs (Cypress Type, we use AMD 5870) at 100% load for hours, a load that except of our Linpack/DGEMM only Furmark can cause...The code also stomps NVIDIA, reaching 497 flops DGEMM on the HD5870 where NVIDIA only reaches around 300 on the much more expensive Tesla systems."