The following talk is sponsored by the Department of Mathematics but
counts as a colloquium for Computer Science students.

Topic:  High Performance Computing in the Multi-Core Era -- Why Care About Performance

Time: 2:00pm, Friday, Oct. 23rd, 2009

Place: E&CS 1201, VMASC Auditorium


Until just a few years ago, the performance of single CPUs doubled
every 12-24 months, mainly owing to increased clock speeds and
manufacturing improvements. This trend stopped some years ago, mainly
owing to power concerns and cooling issues.  Thus, "free lunch" is
over. Moore's law which claims that the number of transistors (and not
the performance) is doubling every 12-24 months, however, is still
valid.  These days, the additional transistors are used for multiple
cores on a single die, large caches, etc. As a result, every new
desktop PC is a parallel computer. On the large scale, high-end
computing facilities are massively parallel with already up to O(105)
cores. To efficiently use multi-core systems on the desktop, clusters
at the faculty level, large federal system or emerging architectures
like GPGPUs, special care is required when implementing scientific
codes. The talk will give some insights on current hardware trends but
focus on implementation issues and give some best-practice guides.

Dr. Thomas Zeiser
Erlangen Regional Computing Center
University of Erlangen-Nuremberg, Germany
thomas.zeiser@rrze.uni-erlangen.de
http://www.rrze.uni-erlangen.de/hpc/