March 7, 2006
IEEE Delaware Bay Section Meeting

A New Era of High-End Parallel Computing
The Arrival of the Multi-Core Chip Technology.

Speaker: Guang R. Gao, PhD, MIT, Electrical and Computer Science. 1986 Endowed Distinguished Professor, Electrical and Computer Engineering, University of Delaware. Founder and leader of the Computer Architectures and Parallel Systems Laboratory (CAPSL)
Location: Trabant Student Center, University of Delaware


Dr. Gao made several references to the IBM Supercomputer program, information at web site http://www.research.ibm.com/bluegene/ 

In 2004, the processor speed had stopped in the 1 GHz area, due to limitations in CMOS technology. There are newer technologies coming along, but these will not add to processor speed in the near future. It is speculated that CMOS will be king until at least 2016. For some years now, the efforts to reach a processor speed of a Petaflop (1015 operations per second) have considered a large number of processors connected in parallel as being the best approach. One reason for this is that faster and more complex computer chips are:

  • Too complex to accurately test, and
  • Run too hot.
So, the alternative is to couple large numbers of simpler, slower computer chips. 

For many years, the US was dominant in the pursuit of the supercomputer. However, in 2002, Japan took the lead with the computer designed to be an Earth Simulator. This computer had the physical appearance of rows and rows of head high cabinets all filled with computer boards. 

The IBM/Lawrence Livermore project called Blue GeneL, (see web site, above) will have 131,072 CPU units, and will require 1.5 megawatts of power to run. 

University of Delaware will develop the software for this project. There is a significant change of emphasis in the way this software is developed. Previous attempts were to make the software "transparent" to the user. This turned out to be overly ambitious, resulting in designing code that not friendly or adequately responsive to the user. This time, selected users will join in with the software development team to define the type of response that is useful. Transparency is a useful concept, but will be reached in a much later stage of development.