Cph CT Toolbox: A performance evaluation

TY - GEN

T1 - Cph CT Toolbox

T2 - A performance evaluation

AU - Bardino, Jonas

AU - Vinter, Brian

AU - Rehr, Martin

PY - 2015/9/2

Y1 - 2015/9/2

N2 - With the first version of the Cph CT Toolbox released and introduced, we turn to intensively evaluating the performance of the FDK and Katsevich reconstruction implementations in the second major release. The evaluation focuses on comparisons between different hardware platforms from the two major GPU compute vendors, AMD and NVIDIA, using our updated CUDA and new OpenCL implementations. Such a performance comparison is in itself interesting in a narrow CT scanning and reconstruction perspective, but it also sheds some light on the performance of those AMD and NVIDIA platforms and GPU technologies: something of general interest to anyone building or considering GPU solutions for their scientific calculations. Results from the best system reveals the chosen streaming strategy to scale linearly up to problem sizes one order of magnitude larger than the available GPU memory, and with only a minor scaling decrease when increasing the problem size further to the next order of magnitude.

AB - With the first version of the Cph CT Toolbox released and introduced, we turn to intensively evaluating the performance of the FDK and Katsevich reconstruction implementations in the second major release. The evaluation focuses on comparisons between different hardware platforms from the two major GPU compute vendors, AMD and NVIDIA, using our updated CUDA and new OpenCL implementations. Such a performance comparison is in itself interesting in a narrow CT scanning and reconstruction perspective, but it also sheds some light on the performance of those AMD and NVIDIA platforms and GPU technologies: something of general interest to anyone building or considering GPU solutions for their scientific calculations. Results from the best system reveals the chosen streaming strategy to scale linearly up to problem sizes one order of magnitude larger than the available GPU memory, and with only a minor scaling decrease when increasing the problem size further to the next order of magnitude.

U2 - 10.1109/HPCSim.2015.7237019

DO - 10.1109/HPCSim.2015.7237019

M3 - Conference article

SN - 1094-3420

VL - 2015

SP - 35

EP - 46

JO - International Journal of High Performance Computing Applications

JF - International Journal of High Performance Computing Applications

ER -