Main problem is not so much the speed of the individual processors, but the way to get data in the processor and the results out fast enough. This is a recurring mantra in HPC for as long as I can remember. Professor Willy Schonauer already mentioned that some 15 years ago. Today, even in the Earth Simulator, which is specially designed for this type of scientific applications, the processor only gets half of the necessary data speed. Other computers with commodity processsors do even worse.

High-speed network interconnects help a lot to build a fast machine, but at a cost. Building a cluster with gigabit ethernet costs about 100 euro per node for the network. Higher speed interconnects, like Quadrics or SCI cost in the order of at least 1000 euro per node.

Dongarra illustrated his analysis with data of all the Teraflop/s machines, 142, in the world. They more or less follow the complete TOP500 list. What is the efficiency of those machines. I.e. what is the percentage of the advertised peak performance that can be used? The efficiency of vector machines is 90% or more. They are specially designed, so one can expect that. For the other machines it is more in the order of 60% or lower, Dongarra said. Looking at the networks involved, Quadrics QSnet does well as does SCI.

Dongarra is a software person, so he ended with: "The real crisis with HPC is the software". On the horizon are machines with 100.000 processors. Programming paradigms are still Fortran and the like. How should we handle that?