In his talk he presented the commodity processors with special discussion of Itanium and IA-32. Then he listed the high-speed interconnects, their features and costs. He concentrated on the Top131, the 1 TeraFlop/s Rmax machines in the last Top500. Interesting was the chart with the efficiency of these systems, which is in the range of 0.2 (i.e. Pentium) to 0.9 (Cray and NEC). The efficiency was analysed with regard to the interconnect too.

Jack Dongarra stated that high bandwidth systems have traditionally been vector computers designed for scientific problems and capability computing. On the contrary, commodity processors are designed for Web servers and the home PC market. The scientific computing needs are different, however. They require a better balance between data movement and floating point operations wich results in greater efficiency.

In the list of architectures relating to 100 TeraFlop/s, we can find SGI Altix, IBM with Power5, 5+, and Power 6, IBM BlueGene, Cray RS, X1e, X2. The real crisis within HPC is the software, as the speaker noted. The programming is stuck and arguably has not changed since the 70's. Jack Dongarra believes it is

time for a change.

The complexity is rising dramatically: today we are confronted with highly parallel and distributed systems, from 10 to 100 to 1000 to 10000 to 100000 of processors (!!) and multi-disciplinary applications. A supercomputer application and software are usually much more long-lived than the hardware. The hardware life typically amounts to five years at the most. Today, Fortran and C are the main programming models. The speaker stated that software is a major cost component of modern technologies and the tradition in HPC system procurement is to assume that the software is free.

Jack Dongarra concluded by citing some current unmet needs. He mentioned the problem of the performance and portability and the fault tolerance. He stated that we need better programming models for global shared address space and visible locality. He expected that the global address space (GAS) languages such as UPC, Co-Array Fortran, and Titanium, might be coming soon because they offer real benefits. They only have minor extensions to existing languages but they are more convenient than MPI and have performance transparency via explicit remote memory references. He ended by warning that the critical cycle of prototyping, assessment, and commercialization must be a long-term, sustaining investment instead of a one time, crash programme.