The authors wrote this paper as an overview of the Connection Machine CM5 for a potential user or programmer. Although it proposes a strict distinction between architecture and implementation, the two are not always well separated, which makes some aspects difficult to understand for a reader who is knowledgeable about parallel processing but not familiar with these particular machines. A description of the processor node, data network, and system software gives a good insight into the design tradeoffs faced by the designer of general-purpose parallel machines, which are similar, though not identical, to the problems faced by the designer of a single-processor CPU. The support of multiple programming models is presented as a distinctive feature of this particular family of machines, and a description of SIMD versus MIMD styles illustrates that although the software view can be very different, the hardware architecture can be the same. Applications from a wide variety of fields are described and compared to traditional vector computer implementations.
As only one brand of MIMD machine is described, general characteristics of this class of machines are mixed with specific features of the CM-5, and the distinction is not always easy to make. Also, a lot of lessons from the past, such as those related to interconnection networks and applications on the CM2, may be particular to this brand of machine, and not always true for other, similar machines. The paper fulfills its purpose as an overview, though some nontrivial previous knowledge could be required (such as the architecture of the CM-2), which is not appropriate for an overview. The length of the paper is adequate. Its best feature is the remarks about the architectural requirements of general-purpose parallel processing. Its main weakness is that some sections could be difficult to read without specific prior knowledge in this field. This could have been avoided by trying to compare more than one general-purpose MIMD machine so that general features became clearer.
The intended audience is people with a general interest in the field of parallel computers. The list of references can act as a starting point for further reading.