next up previous

3.5.1 Vector Processors     continued...

The advantage of memory to memory machines is the ability to process very long vectors, whereas register to register machines must break long vectors into fixed length segments. Unfortunately, this flexibility is offset by a relatively large overhead known as the startup time, which is the time between the initialization of the instruction and the time the first result emerges from the pipeline. The long startup time on a memory to memory machine is a function of memory latency, which is longer than the time it takes to access a value in an internal register. Once the pipeline is full, however, a result is produced every cycle or perhaps every other cycle. Thus a performance model for a vector processor is of the form

where s is the startup time, N is the length of the vector and a is an instruction dependent constant, usually , 1 or 2.

Examples of this type of architecture include the Texas Instruments Inc. Advanced Scientific Computer and a family of machines built by Control Data Corp. known first as the Cyber 200 series and later the ETA-10 when Control Data Corp. founded a separate company known as ETA Systems Inc. These machines appeared in the mid 1970s after a long development cycle that left them with dated technology and disappeared in the mid 1980s. For a thorough discussion of their characteristics, see Hockney and Jesshope [13]. One of the major reasons for their demise was the large startup time, which was on the order of 100 processor cycles. This meant that short vector operations were very inefficient, and even for vectors of length 100 the machines were delivering only about half their maximum performance. In a later section we will see how this vector length that yields half of peak performance is used to characterize vector computers.