The report Overview of recent supercomputers celebrates its tenth anniversary. It gives an overview of all the parallel and vector systems that are commercially available. Although there is somewhat arbitrariness about which systems to include and which not, in general all the important systems are there. Only the Siemens hpcLine is really missing.
Aad van der Steen notes that in recent years more systems have disappeared from his list than new ones have entered. However, currently this process seems to be stopped, with as many new machines entering as leaving.
The report gives a nice overview of the different architectural classes of supercomputers. It describes for instance SIMD, MIMD, and ccNuma concepts.
The core of the report consists of a description of each machine, such as operating systems, compilers, year of introduction, and provides this data in tabular form. It also includes system parameters like clock speed, performance, memory and interconnect speed. Remarks and notes on bench mark performance conclude each description.
For systems no longer available, a brief overview is given with no lengthy explanation.
In the section Systems under development, Aad van der Steen looks in his crystal ball and forecasts the following for the coming year. He expect that new machines will come out following the SMP cluster-based architectural model. Compaq for instance will come out with a 32 or perhaps 64 way cluster. The QsNet network technology that is now employed will soon allow for higher bandwidth as the underlying PCI technology speeds up.
For HP, Aad van der Steen sees a bottleneck coming up in the network interconnect. The PA-RISC processors deployed are becoming more and more powerful, but the interconnect has not changed for quite some time. One may expect HP will change this and come up with a composition of SMP clusters in a multi-level cross-bar. Whether the new systems will be PA-RISC based or use the IA-64 chip is still unclear.
The interconnect, a high-performance switch, that is at the hearth of the IBM SP supercomputers, will also become a bottle-neck for increasing overall machine performance in the coming years. Because of the Silicon-On-Insulator and copper technology, one can expect the processor speed to increase significantly, notes Aad van der Steen. Within the next few years, bandwidths of 1000 Mb/s should be feasible.
SGI will come with the new SN architecture, for which already the first orders have been signed. The machine will be equipped with a MIPS processor in the 2 nsec clock cycle range or perhaps an IA-64 processor. Aad van der Steen expects that, since the latter may not gain the acceptance predicted, the MIPS option will be there for some time to come.
With the Cray business moved from SGI to Tera, the future of the Cray T3E and the Tera MTA is somewhat unclear. Aad van der Steen anticipates that the SV line will gain enough support from the US government to be sustained. It is the only vendor of vector computers still alive in the USA. And although not dominant in supercomputing in general, it is still a very important niche architecture.
The report is available in printed form from the NCF . On the Web it can be accessed through: Overview of recent supercomputers . Primeur has produced an XML-version that is linked with the TOP500 data to show how augment data from one source with that from another.