Performance laws and metrics for parallel architectures
Introduction to parallelism and parallel architectures

This paragraph introduces various performance measures: peak performance, acceleration.

The simplest performance to calculate, but also the least useful for predicting the performance of a parallel architecture, is peak performance. This is simply obtained by multiplying the individual peak performance of all the processing units by their number in the architecture.

A widely used definition of peak performance is: the performance that the manufacturer guarantees you won't be able to exceed!

More relevant information is an application's computation time (or execution time). This can be used to derive the computation rate for floating-point or integer numbers. The following formula gives processing throughput as a function of calculation time: