In this paper, the authors discuss fine-grain power savings obtained through a technique which we call Effective Precision Matching Computation-PreMatch. It works by dynamically selecting arithmetic modules in a system designed with Functional Redundancy. Functional Redundant (FR) systems are those which contain different implementation instances of functional blocks or circuits on chip. The different instances are obtained by using different hardware algorithms and/or precision (bitwidth) of their operands. The selection of modules is completely dynamic and data-driven, i.e., it is done according to the contents of the parameters passed as input to those modules. Basically, according to the values of the operands different instances of adders, multipliers, and so on can be selected to “match" the required precision of the computation. Experiments showed that up to 28.4 energy savings can be achieved with PreMatch alone, and up 29.3 using both techniques (FR and PreMatch).
展开▼