Low-Power Processors Require Effective Memory Partitioning

The ever increasing complexity of embedded systems demands for rising memory size, and larger memories increase the power drain. In this work, we exploit banked memories with independent lowleakage retention mode in event-driven applications. The resulting energy saving for a given number of banks is close to the maximum achievable value, since the memory banks access pattern of event-driven applications presents a high temporal locality, leading to a low saving loss due to wake-up transitions. Results show an energy reduction up to 77.4% for a memory of ten banks with a partition overhead of 1%.