THE CONNECTION MACHINE

The Thinking Machines Corporation, Cambridge, Massachusetts, released the first CM-2 in April 1987. Since then, more than 40 units have been installed in the US and Europe. The CM-2 is one of only few parallel computers using the one-data--one-processor concept and concentrating on increasing the number of processors rather than using faster processors.
The most important feature of the CM's architecture is the concept of virtual processors (VPs). Each element of a parallel variable (usually an array element or an array dimension) is associated with one virtual processor. VPs are analogous to virtual memory: In either case the user can address a nearly infinite number of processors, or memory cells, independent of the actual hardware configuration.
If as usual the number of physical processors N_P is bigger than the number of virtual processors N_V, then every VP is related to one physical processor. If N_V >N_P, at least some physical processors have more than one VP associated with them. The ratio $\frac{N_{V}}{N_{P}}$ is called the VP-ratio. Because calculations can involve only the currently active set of VPs, i.e. the VPs currently present in physical processors, one has to switch from one VP-set to the next, if the VP-ratio is bigger than one. This is called VP-looping. Extensive VP-looping is costly, because it requires a physical shift from data between the external storage system and the processors' memories.
The high-level language-user can always think of the CM processors as representing a Cartesian grid, because Cartesian coordinates can be embeded in the 12-dimensional hypercube that forms the CM's hardware interconnectivity. The n-dimensional grid of VPs is described by the processor's geometry and can be defined by the user according to the problem being addressed. You will find an example in my first program sample.
The user accesses the CM via a front-end computer. All editing, compiling, and linking are done on the front end. Programs run on the front end, too, but parallelizable program parts are executed on the CM.
The Connection Machine itself consists of one, two, or four portions a 8192 or 16384 processors. Each of these parts is accessable to different users because it has its own sequencer that interprets the parallel instructions sent by the front end and activates its processor portion accordingly. I always used 8k processors, the smallest CM-portion available. Each CM processor has its own 8k memory and arithmetic logical unit.