previous | contents | next

Chapter 20 ½ The Iliac IV System 307

Before operations could be overlapped, control sequences between the components had to be decoupled. Certainly the CU could at least be fetching the next instruction while the ALU was executing the present one.

2 Replication: One of the four major components (or subcomponents within a major component) could be duplicated many times. (Ten black boxes can produce the result of one black box in one-tenth of the time if the conditions are right.) The replication of I/O devices, for example, was a step taken very early in the evolution of digital computers-large installations had more than one tape drive, more than one card reader, more than one printer.

Since the above two philosophies do not mutually exclude each other, a third approach exists which consists of both of them in a continuously variable range of proportions.

The overlapping philosophy was implemented largely through the buffer and pipeline mechanisms. The pipeline mechanism (see Fig. 2) breaks down an operation into suboperations, or stages, and decouples these stages from each other. After the stages are- decoupled they can be performed simultaneously or, equivalently, in parallel, The buffer mechanism allows an operation to be decoupled into parallel operation by providing a place to store information.

The replication philosophy is exemplified by the general multiprocessor which replicates three of the four major components (all but the I/O) many times. The cost of a general multiprocessor is, however, very high and further design options were considered which would decrease the cost without seriously degrading the power or efficiency of the system. The options consist merely of recentralizing one of the three major components which had been previously replicated in the general multiprocessor-the memory, the ALU, or the CU. Centralizing the CU gives rise to the basic organization of a vector or array processor such as Illiac IV. This particular option was chosen for two main reasons.

1 Cost: A very high percentage of the cost within a digital computer is associated with CU circuitry. Replication of this component is particularly expensive, and therefore centralizing the CU saves more money than can be saved by centralizing either of the other two components.

2 Structure: There is a large class of both scientific and business problems that can be solved by a computer with one CU (one instruction stream) and many ALUs. The same algorithm is performed repetitively on many sets of different data: the data are structured as a vector, and the vector processor of Illiac IV operates on the vector data. All of the components of data structured as a vector are processed simultaneously or in parallel.

The Illiac LV project was started in the Computer Science Department at the University of Illinois with the objective of developing a digital system employing the principle of parallel operation to achieve a computational rate of 109 instructions/s. In order to achieve this rate, the system was to employ 256

previous | contents | next