GPUs for Smarties: From Gaming to General Purpose Computing

Speaker: Dr. Yong Cao, Virginia Tech
Location: Torgersen 2150
Date: Friday, April 24, 2009
Time: 11:15am-12:15pm

Due to the tremendous performance improvement of Graphics Processing
Units (GPUs) in recent years, the desktop PCs equipped with
state-of-art GPUs are labeled as "Personal Supercomputers". These
commoditized GPU systems can reach a computing capacity of
approximately 7 Tera FLOPs, and only cost around 4,000 dollars.
However, only a certain type of algorithms, data parallel algorithms,
can take full advantage of the GPU power, because of high memory
latency and the lack of synchronization mechanisms.

In this talk, I will discuss algorithm optimization strategies for
NVIDIA GPUs, focusing on how to maximize the level of parallelism and
how to hide memory latency. I will use several research projects as
examples. These projects include both data-parallel applications, such
as video processing and volume rendering, and non data-parallel
algorithms, such as a PDE (partial differential equation) solver and
temporal data mining.