Reconfigurable Supercomputing:

 What are the Problems?
 What are the Solutions?

 
Reiner Hartenstein, IEEE life fellow
TU Kaiserslautern http://hartenstein.de/RC/


The dominant paradox of supercomputing is the contrast between decades of excellent 
technology development (e. g. see the Gordon Moore curve) versus the almost stalled 
progress in sustained performance in many application areas. The counterparts to this 
supercomputing paradox are the 3 paradoxes of Reconfigurable Computing: the low power 
paradox, the high performance paradox, as well as the education paradox. Despite to the fact, 
that are very power-hungry, software to configware migrations have been reported which 
reduce the tens or hundreds of thousand dollars electricity bill by an order of magnitude. 
Despite of the really awful technological parameters of FPGAs brilliant results with speedups 
by up to 4 orders of magnitude have been reported from software to FPGA migrations.

The presentation discusses the reasons of these paradoxes. A very important aspect is the way, 
how data and processors are brought together. From this point of view, supercomputing has 
used for decades the wrong road map, based on the wrong machine paradigm being extremely 
memory-cycle-hungry. The presentation illustrates, why the not instruction-stream-centered 
basic machine paradigm of Reconfigurable Computing provides the right road map to new 
horizons of supercomputing.

The Reconfigurable Computing education paradox shows, that its extremely high 
pervasiveness to applications in practically all disciplines of embedded system as well as 
scientific computing is possible, although computing-related curriculum recommendations 
completely ignore these subject areas. This leads to the conclusion, that these achievements 
have been implemented primarily from experts with backgrounds different from computing 
sciences. CS departments are the best possible institutions to overcome the methodology 
fragmentation between many FPGA application disciplines using their own domain-specific 
tool trick boxes, and, to develop models covering all aspects which the application disciplines 
have in common. Refusing to take this responsibility the undergraduate CS-related curricula 
also miss the most important job markets for their graduates.