Home » It's HIP to Enhance Performance Portability of GPU Accelerated Software
It's HIP to Enhance Performance Portability of GPU Accelerated Software
Date and Time:
Monday April 8th 2019
Location:
CG Auditorium
Speaker:
Joe Schoonover
HIP is a new open-source API, similar in syntax to CUDA, that permits program execution on AMD and Nvidia accelerator hardware. Further, HIP comes with a "HIP-ify" script that can be used to convert existing CUDA code to HIP. This talk demonstrates GPU acceleration of a spectral element kernel for computing vector divergence in 3-D on AMD Nvidia platforms with a single HIP kernel, generated from an existing CUDA kernel with the HIP-ify utility. On Nvidia hardware, it is shown that performance is preserved relative to the original CUDA kernel. Further, we provide performance comparisons across CPU and Nvidia and AMD GPU hardware that highlight the performance portability of HIP kernels.
Speaker Description:
Dr. Joseph Schoonover holds degrees in Applied Mathematics, Physics, and Geophysical Fluid Dynamics from Florida State University. His graduate studies focused on Gulf Stream separation dynamics and high order methods for computational fluid dynamics. After graduating from FSU, he held a post-doc position at the Center for Non-Linear Studies at Los Alamos National Laboratory. This is where his interest in GPU acceleration was born through mentorship activities at the Parallel Computing Summer Research Internship. After leaving CNLS and LANL, Joe became an associate scientist at CU Boulder to work at NOAA's Space Weather Prediction Center to accelerate the operational WAM-IPE code for modeling ionosphere phenomena. Joe has since moved on to a startup company, Fluid Numerics, founded with Guy Thorsby and Elizabeth Simons, that aims to help domain scientists and software developers leverage the latest compute technologies and cloud computing platforms for high performance computing.
SEA2019_HIP_FluidNumerics.pdf
