Formalized methodology for data reuse: exploration for low-powerhierarchical memory mappings

摘要

Efficient use of an optimized custom memory hierarchy to exploitntemporal locality in the data accesses can have a very large impact onnthe power consumption in data dominated applications. In the past,nexperiments have demonstrated that this task is crucial in a completenlow-power memory management methodology. But effective formalizedntechniques to deal with this specific task have not been addressed yet.nIn this paper, the surprisingly large design freedom available for thenbasic problem is explored in-depth and the outline of a systematicnsolution methodology is proposed. The efficiency of the methodology isnillustrated on a real-life motion estimation application. The resultsnobtained for this application show power reductions of about 85% for thenmemory subsystem compared to the case without a custom memory hierarchy.nThese large gains justify that data reuse and memory hierarchy decisionsnshould be taken early in the design flow