September 3, 2008
An Overview of Highly-Threaded Parallelism using Graphics Processing Units
Paul Gray, Associate Professor of Computer Science at UNI, will deliver a CS seminar on Wednesday, September 3, at 3:30 PM in 322 ITTC.
As supercomputers move from teraflops to petaflops, Graphics Processing Units (GPUs) become an important piece of the puzzle for general-purpose parallel computing. As GPUs evolve, driven by the demands of graphics software, they are also morphing into a high-performance computing environment suitable for general-purpose calculations. In recent years, GPU performance increases have greatly outpaced the marginal increases in CPU performance. For example, NVIDIA's G80 GPU has a peak floating-point performance of 330 gigaflops. The peak performance of ATI's Radeon HD 2900XT is about 475 gigaflops. With a little manipulation of the architecture -- as well as a bit of overclocking and a lot of cooling -- system boards today can support up to four of these cards in a single system for capability of up to a full teraflop in a single box.
This talk will give a general introduction to GPU programming using CUDA, a free CPU programming toolkit from NVidia for Windows and Linux platforms. It will highlight fundamental differences between GPU programming and traditional single-CPU programming and give simple coding examples in CUDA, along with illustrations of speedups on a system with multiple GPU units (NVidia 8800 GTS).
Paul Gray earned his B.S. degree in Applied Mathematics at the University of Iowa and his Ph.D. in Applied Mathematics at Michigan State University. After serving a postdoc in parallel and distributed computing environments under Dr. Vaidy Sunderam at Emory University, Dr. Gray joined the faculty at UNI, first in Math and later in Computer Science. His research across a number of topics in high-performance computing and HPC education, including such projects as BCCD, a "bootable cluster" on a CD, and LittleFe, a portable, low-cost cluster for parallel and distributed computing.
(the old East Gym)
Cedar Falls, Iowa
ph. (319) 273-2618
fax (319) 273-7123
With distributed computing, the devil is in the details. That's because a programmer is dealing with all sorts of different hardware and networks. Meshing all of these elements presents a significant challenge that Paul Gray, associate professor of computer science, is leading the charge to solve... [more]