As a community we have almost forgotten, what supercomputing was like
twenty years ago in 1985.  The state of the art system then was a 2
Gflop/s peak Cray-2, with at that time phenomenal 2 GBytes memory. It
was the era of  custom built vector mainframes, where anything beyond
100 Mflop/s sustained was considered excellent performance. The software
environment was Fortran with vectorizing compilers (at best), and a
proprietary operating system. There was hand tuning only, no tools, no
visualization, and dumb terminals with remote batch. If one was lucky
and had a account, remote access via 9600 baud was state-of-the-art.
Usually a single code developer developed and coded everything from
scratch.

What a long way have we come in the last twenty years! Teraflop/s level
performance on inexpensive, highly parallel commodity clusters, open
source software, community codes, grid access via 10 Gbit/s, powerful
visualization systems, and a productive development environment on a
desktop system that is more powerful than the Cray-2 from 20 years ago
-- these are the characteristics of high performance computing in 2005.

Of course a significant contribution to this progress is due to the
continued increase of computing power following Moore's law. But what I
want to argue here is, that progress was not just simply quantitative
alone. We did not just get more of the same at a cheaper price. There
were several powerful ideas and concepts that were shaped in the last 20
years, that made supercomputing the vibrant field that it is today.  As
an active researcher in the field for the last 25 years, I will offer my
subjective opinion, what were the real top breakthrough ideas that led
to qualitative change and significant progress in our field.

Retrospection leads to extrapolation: in the last part of the lecture, I
will envision, what supercomputing will be like 20 years from now in the
year 2025.  Can we expect similar performance increases?  How will
supercomputing change qualitatively? And what are the top challenges
that we will have to overcome to reach that vision of supercomputing in
2025?