4. Parallel Computing 14. Introduction

Vector Processors

The name "vector processor" sounds similar to "array processor" discussed above.

However, vector processor connotes something different to those in the field: a processor

in which the data stored in registers are vectors. Typically there are both scalar registers

and vector registers, with the instruction code determining whether an addressed register

is one or the other. Vector processors differ from array processors in that not all elements

of the vector are operated on concurrently. Instead pipelining is used to reduce the cost of

a machine that might otherwise have one arithmetic unit for each vector element.

Typically, vector operations are floating-point. Floating point arithmetic can be broken

into four to eight separate stages. This means that a degree of concurrency equal to the

number of stages can be achieved without additional arithmetic units. If still greater

concurrency is desired, additional pipelined arithmetic units can be added, and the vectors

apportioned between them.

---

594

Parallel Computing