GPULib : A Library for GPU Acceleration of Scientific Computing Applications

Speaker: Paul Mullowney, Tech-X Corporation, Boulder, CO.

The computational demand of many scientific modeling problems is growing at a considerable rate. Until recently, this growing need was satisfied primarily through the use of cluster or supercomputing. But for many scientists and other end users, it would be more desirable to have high performance code in their very own desktop or laptop computer. Modern graphics processing units (GPUs) offer substantial power at very low cost, however advanced programming skills and detailed hardware knowledge are required to take advantage of these devices. In this talk, we first discuss a library that allows GPU hardware to be accessed from within High Level Languages like MATLAB, IDL, or Python. By abstracting the GPU as a vector co-processor, computationally demanding data analysis and simulation tasks can be accelerated by an order of magnitude or more. Then, we switch gears and discuss more recent GPU-based implementations of an advanced physics simulation, DSMC-PIC (Direct Simulation Monte Carlo Particle-In-Cell). This simulation type contains a number of challenging issues that must be overcome in order to make effective use of the hardware. However, we shall see that even in the most naive implementation, benefit can be obtained. We finish with a discussion of possible future developments including Electromagnetic and Electrostatic PIC.

Last modified: Mon Oct 20 11:52:17 EDT 2008