4. Parallel Computing 14. Introduction

Figure 299: SIMD multiprocessor organization

Notice that in the SIMD organization, each data processor is coupled with its own data

memory. However, in order to get data from one memory to another, it is necessary for

each processor to have access to the memories of its neighbors, at a minimum. Without

this, the machine would be reduced to a collection of almost-independent computers.

Also not clear in the diagram is how branching (jumps) dependent on data values take

place in control memory. There must be some way for the control processor to look at

selected data to make a branch decision. This can be accomplished by instructions that

form some sort of aggregate (such as the maximum) from data values in each data

processors' registers.

SIMD Multiprocessors are also sometimes called array processors, due to their obvious

application to problems in which an array can be distributed across the data memories.