Abstract :
Many of today´s parallel computer systems are built using standard off-the-shelf microprocessors and memories. This is not surprising since such devices, being produced in vast numbers, are significantly cheaper than custom devices would be. Even where special purpose devices are made, it is reasonable to expect that the long-term success of such designs would depend on them being made in some quantity. One of the first questions which must be answered by someone designing a parallel computer is, `All else being equal, how many processors?´. Given an unlimited number of nodes, connected in a perfect (zero latency, infinite bandwidth) network, how many nodes can a particular algorithm use? The author looks at this question
