Technical data

Vector Processing Concepts
1.10 PERFORMANCE
The performance of scalar computers has been measured for some time
using millions of instructions executed per second (MIPS). MIPS is not
a good measure of speed for vector processors, since one instruction
produces many results. Vector processor execution speed, instead, is
measured in millions of floating-point operations per second (MFLOPS).
Other abbreviations used are MegaFLOPS, MOPS (millions of operations
per second), and RPM (results per microsecond). Some of the largest
computers measure speed in gigaFLOPS or billions of floating-point
operations per second.
The peak MFLOPS value is a vector processor’s best theoretical
performance, in terms of the maximum number of floating-point
operations per second. For a vector processor having a processor
cycle time of 5 nanoseconds and 1 arithmetic unit per pipeline, its
peak MFLOPS performance (defined as 1 divided by the cycle time) is
determined as follows:
9 9
1.10.1 Amdahl’s Law
Amdahl’s Law indicates that the performance of an application on a
computer system with two or more processing rates is dominated by
the slowest process. Vector processing is faster than scalar processing
executing the same operation, yet the primary factor that determines a
computer’s speed is the speed of the scalar processor (see Figure 1–8).
Amdahl’s law, expressed as an equation, gives the time (T) to perform N
operations as:
T = N X (%scalar operations X time/scalar operation +
%vector operations X time/vector operation)
1–19