Methods of On-Line Computation Acceleration in Systems with Direct Connection between Units

A possibility to increase the effectiveness of the parallel systems with direct connections by means of data between computation units, which are connected one to another in accordance with data flow graph, is considered. Operations correspond to the graph vertices and data correspond to the graph arcs. In the computation units operations are performed in redundant symmetric number system. It enables to overlap digit-by-digit input of operands, their processing and digit-by-digit output starting with the high digits. A result digit we get does not require further correction and can be used as an operand digit within the next computation cycle in other computation units. It gives an opportunity to perform sequences of operations connected by means of data in the concurrency mode at the level of operand digit processing. An algorithm and a structure of a computation unit, which permits to form a result digit within one tact, are proposed. Dependence of the operation execution time on the number system base is considered. It is shown that due to digit-by-digit data representation the required number of internal connections between the units, as well as of the chip pinouts, is reduced.