In particular for this class of systems we cannot claim completeness of this overview. This has two reasons: First, at present this is the most dynamic area of development of new machines and it is quite probable that already new systems appear on the market while this report goes to print. This in no way implies that the systems not mentioned here should be in any way inferior to the ones that appear in this section. It is rather felt that many of these systems are in some sense equivalent and listing (almost) all of the systems would be counterproductive in the sense that the descriptions of the systems might lead to confusion.
For distributed-memory MIMD machines obviously the internode bandwidth and latency are very important system parameters. Unfortunately, it is very hard to come by reliable figures for these parameters. Therefore, we only can state the internode bandwidth point-to-point for the majority of systems, not for all. Where we do not have these figures we give the aggregate bandwidth which is less informative but better than nothing. We were not able to give latency figures for the systems for two reasons: manufacturers mostly state hardware latencies which, regrettably, does not say very much about the actual latency, except that the hardware latency is a guaranteed lower bound. The second reason is that the actual (software) latency, even if known at some point in time, decreases very fast, as better implementation of the communication software occurs continuously. Therefore, stating figures for this system parameter is next to useless at the moment even when very much desired.