Custom instruction filter cache synthesis for low-power embedded systems

Filter cache has been shown to substantially reduce the power consumption in instruction memory hierarchy. Filter cache achieves energy savings due to the locality found in the frequent tiny loops, which are application dependent. In this paper we show that tuning filter cache to the needs of a particular application can save power and energy. Beside, a simple loop profiler directed methodology to deduce the optimal or near-optimal filter cache is proposed, without having to simulating all possible combinations of cache parameters from the specified space. Our experiments with MediaBench benchmark suite shows that the proposed methodology results in up to 49% energy reduction by tuning the filter cache. Moreover, the proposed filter cache tuning is done with the loop characteristics of the application, which in most cases are readily made available.