In his talk Erich Strohmaier showed that the only thing constant in the high performance computing market is rapid change but the evolution of performance on a large scale is a steady process which seems to respect Moore's Law stipulating an increase in performance of two orders of magnitudes every decade over the last sixty years. In his presentation the speaker only covered the last thirty years analysing the procedures applied to draw the statistics and their usability, with market data provided by the Mannheim Supercomputer Seminar and the TOP500 project.

Thirty years ago modern supercomputing started out with vector systems, raw performance being the only argument to buy them. A decade later the focus was put on standard programming environments, operating systems and key applications. Scalable systems with distributed memory introduced the trend of massive parallel computing stimulated by various government programmes in order to overcome the existing scalability limitations. Later on the SMP systems were developed to focus on the lower and medium market segments. Towards the end of the nineties, industry preferred clusters of SMPs whereas the academic community went for clusters of workstations and PCs.

Since 2004, clusters built with components off the shelf are gaining momentum in the TOP500 list. This brought along a rapid rise in the application of Intel processors in HPC systems, according to the speaker. The introduction of the Earth Simulator, a highly efficient vector computer system, took the United States HPC community by surprise but since then they reacted with the low-power consuming IBM BlueGene/L system which has a record number of processors. This is high density computing with a limited main memory size. The next development will be a supercomputer delivering PetaFlop/s performance levels on real applications.

Cray was producing vector supercomputing systems since 1976. CDC had introduced the Cyber 205 in 1982. Later came the Japanese vector systems and the vector multi-processor at Cray. Between 1980 and 1985, the number of installed systems grew by 37 percent annually, as the speaker showed. The class of symmetric multi-vector processor systems dominated the supercomputing arena due to its commercial success in the eighties.

Erich Strohmaier explained that the share of United States installations decreased steadily in contrast to the increase of the Japanese share. By 1992, the worldwide dominance of Japan was growing, but in 1993, the first TOP500 showed the USA far ahead in supercomputers. This was mainly because a bigger number of relatively small entry-level systems of Japanese vendors have been installed in Japan.

Between 1985 and 1990, Cray dominated the martket with a constant share of 60 percent. In 1991 there was a CDC exit. The growth of the industry was declining. In the late eighties there were companies that started using supercomputiers with different architectures, namely parallel computers with distributed memory. A lot of new vendors pushed their products and wanted to be counted but entry level models were condidered too small to be counted, according to the speaker.

In the TOP500 list Erich Strohmaier is trying to balance statistics. As manufactueres are concerned, the Japanese market shares show 25 percent. IBM and HP are currently on the rise in the statistics. IBM has more than 50 percent than the rest.

Erich Strohmaier claimed that the TOP5000 proved itself by correcting the Mannheim Supercomputer Statistics. It is simplistic but or maybe because of this, it gets the trends right. It does not easlily allow to track market size. The list is inventory based: this smoothes seasonal fluctuation. The tunrover is very high so it still reflects the recent developments.

In the 25th TOP500 list, this is the Top 10:

• BlueGene/L
• BGW eServer Blue Gene - IBM, Thomas Watson
• Columbia SGI - NASA Ames Research
• Earth Simulator
• MareNostrum - Barcelona
• eServer Blue Gene - Astron, Groningen
• Thunder
• eServer Blue Gene - EPFL
• eServer Blue Gene - JAIST
• Red Storm - Sandia

Germany is ahead of the United Kingdom and ahead of France. France is now in the number two spot in Europe. The market in Japan has been exclusively for Japanese customers. The new economies of Korea and China are starting to use supercomputers, according to the latest developments.

Erich Strohmaier indicated that clusters have become very dominant in the TOP500 list. Contrary to the hopes of many small builders, IBM is now selling large quantities to the industry.

As far as interconnects are concerned, the industry just uses what comes, according to the speaker, such as Gigabit and Myrinet. What type of operating systems are used? Well, it seems that Unix and Linux have a strong market share. Most of the supercomputers in the TOP500 list use the Linux operating system.

If we look at the number of processors, ten years ago, an average system had 100 processors in the TOP500 list. Nowadays, this has evolved to 65.536 processors for the number 1, BlueGene/L.

The overall performance evolution is approaching the 1 TeraFlop barrier. In ten years from now, the last system on the list will be 1 TeraFlop. Erich Strohmaier predicted. At present, the total performance of all systems in the TOP500 list approaches 1 TeraFlop.

We face a lot of discussion about what a supercomputer is. This centers around using Linpack as the yardstick. Linpanck does not represent Scientific Computing very well and might need replacement at some point. But, so far the TOP500 statistiscs seem to be consistent, concluded Erich Strohmaier.

You do not have to believe Erich Strohmaier on his word. If you go to the http://top500.org site, you have access to the complete database of all the data of all the lists an you can do your own analysis.