NCSA’s Private Sector Program supercharges Abaqus performance with GPUs | National Center for Supercomputing Applications at the University of Illinois
NCSA’s Private Sector Program supercharges Abaqus performance with GPUs
05.19.14 - Permalink
Seid Koric, a senior computational resources coordinator with NCSA’s Private Sector Program, says he experienced “an a-ha moment” recently when evaluating the performance of the Abaqus/Standard implicit finite element software on the GPU nodes of the center’s Blue Waters supercomputer. The results convinced him that “GPU computing has a future for commercial codes.”
Because many of NCSA’s industry partners are interested in exploring the potential of GPUs to accelerate their time to solution, Koric decided to benchmark the popular finite element software Abaqus—used in the auto, aerospace, and industrial products industries—to the Cray XK7 nodes of Blue Waters. Since a native Cray port of Abaqus does not exist, he used x86 binaries under cluster compatibility mode (CCM).
While many other attempts to leverage GPUs to run commercial finite element codes have been limited to single-precision floating-point operations and a single node, Koric used double-precision and multiple nodes. MPI was used for communication between nodes while the computationally intensive kernels were offloaded to an NVIDIA Kepler GPU attached to every XK7 node. This is one of the first hybrid programming models, successfully implemented by Simulia Dassault Systems, for a commercial code on GPUs.
A standard Abaqus benchmark (S4B) with 5 million degrees of freedom (DOFs) has scaled to six Cray XK7 nodes with performance between 50 and 80 percent faster than when using only CPUs on XE6 nodes. Koric plans to benchmark larger real-world problems from NCSA’s industrial partners in the future.
Blue Waters is supported by the National Science Foundation through awards ACI-0725070 and ACI-1238993.