Drowsy cache partitioning for reduced static and dynamic energy in the cache hierarchy

Power consumption in computing today has lead the industry towards energy efficient computing. As transistor technology shrinks, new techniques have to be developed to keep leakage current, the dominant portion of overall power consumption, to a minimum. Due to the large amount of transistors devoted to the cache hierarchy, the cache provides an excellent avenue to dramatically reduce power usage. The inherent danger with techniques that save power can negatively effect the primary reason for the inclusion of the cache, performance. This thesis work proposes a modification to the cache hierarchy that dramatically saves power with only a slight reduction in performance. By taking advantage of the overwhelming preference of memory accesses to the most recently used blocks, these blocks are placed into a small, fast access A partition. The rest of the cache is put into a drowsy mode, a state preserving technique that reduces leakage power within the remaining portion of the cache. This design was implemented within a private, second level cache that achieved an average of almost 20% dynamic energy savings and an average of nearly 45% leakage energy savings. These savings were attained while incurring an average performance penalty of only 2%