Various interconnects for SCI-based systems

In this paper we wish to show some of the possible ways in which small SCI-rings can be connected together to form large systems of switches, processors and memories. Choosing a topology is very much dependent of the use of the network. A general multiprocessor requires an entirely different system than for example a data acquisition system. The topologies mentioned in this paper are relevant to general multiprocessors. The authors are currently implementing a simulation model for some of these topologies as part of their thesis. 1 Introduction The Scalable Coherent Interface (SCI) -- IEEE P1596 -- is a high performance interface standard for multiprocessors. The nodes communicate by sending packets on unidirectional point-to-point links. Each link consists of 18 bit-lines, 16 for data, 1 for flag and 1 for clock. The lines between nodes in the figures represent such a link. They are capable of an output of 1 GigaByte/second. 1 The nodes can consist of a processor with no memory, me..