Citation Context

...adcast and unbounded fan-in, or have each input travel through an array of cells so that it is used at each cell. For modular expansibility of the resulting system, the second approach is preferable. =-=(2)-=- The design uses extensive concurrency. The processing power of a systolic architecture comes from concurrent use of many simple cells rather than sequential use of a few powerful processors as in man...

Citation Context

...ommunication. Criteria and advantages Having described a family of systolic convolution arrays, we can now be more precise in suggesting and evaluating criteria for the design of systolic structures. =-=(1)-=- The design makes multiple use of each input data item. Because of this property, systolic systems can achieve high throughputs with modest I/O bandwidths for outside communication. To meet this crite...

Citation Context

...ementations each cell could reasonably contain a high-performance arithmetic unit plus a few thousand words of memory. There is, of course, always a trade-off between cell simplicity and flexibility. =-=(4)-=- Data and controlflo ws are simple and regular. Pure systolic systems totally avoid long-distance or irregular wires for data communication. The only global communication (besides power and ground) is...

Citation Context

...support two-level pipelining.25 Since system cycle time is the time of a stage of a cell, rather than the whole cell cycle time, two-level pipelined systolic systems significantly improve throughput. =-=(3)-=- There are only afew types ofsimple cells. To achieve performance goals, a systolic system is likely to use a large number of cells. The cells must be simple and of only a few types to curtail design ...