Rafail Abramov : Graphic chips: a somewhat unexpected computational alternative

Posted by , part of the Graduate Student Colloquium.

At

April 4, 2008, 3 p.m.

In

SEO 636

Abstract

Modern computer video cards nowadays are used primarily for real-time 3D graphic applications, where fast 3D image rendering is a computationally intensive problem, naturally suitable for parallel processing. As a result, graphic processing units (GPU) eventually became powerful multiprocessors, tailored for massively parallel floating point computations with multiple (as much as 128) floating point arithmetic units on a single chip. Recently, one of the video card manufacturers, NVIDIA, started offering a common purpose software interface to program a GPU for general parallel floating point computations, allowing to obtain a speedup of 1-2 orders of magnitude in comparison with a modern CPU for essentially same price. I will try to give an overview of this new computational approach, as well as show a working program example and compare the GPU performance with a similar CPU program.