Introduction to Data-Stream-Based Processing on Graphics Processors
Robert Strzodka (Stanford University)
MSIDATE: 2005-12-05
TIME: 11:00:00 - 12:00:00
LOCATION: Bernhard Neumann seminar room, G35, John Dedman Bldg
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
The performance gap between processing elements and memory has become the main obstacle in fast processing of large data sets. The von Neumann computing paradigm reinforces this problem by focusing on instruction rather than data processing. Graphics Processors Units (GPUs) have traditionally been optimized for high data throughput. They subscribe to a data-stream-based computing paradigm which maximizes memory efficiency and exploits the parallelsim in situations when the same operation is applied to many data items. The presentation gives an introduction to the GPU programming model and explains how to exploit the enormous computing power of up to 192 GFLOPS in scientific computations. Important topics are parallel GPU programming, matrix vector products, accuracy and discretization schemes. The hardware acceleration is illustrated with solutions of different PDE and minimum problems in image processing and computer vision.
