Improvement Energy Efficiency for a Hybrid Multibank Memory in Energy Critical Applications

High performance, low power multiprocessor/multibank memory system requires a compiler that provides efficient data partitioning and mapping procedures. This paper introduced two compiler techniques for the data mapping to multibank memory, since data mapping is still an open problem and needs a better solution. The multibank memory can be consisted of volatile and non-volatile memory components to support ultra-low powered wearable devices. This hybrid memory system including volatile and non-volatile memory components yields higher complexity to map data onto it. To efficiently solve this mapping problem, we formulate it to a simple decision problem. Based on the problem definition, we proposed two efficient algorithms to determine the placement of data to the multibank memory. The proposed techniques consider the characteristic of the non-volatile memory that its write operation consumes more energy than the same operation of a volatile memory even though it provides ultra-low operation power and nearly zero leakage current. The proposed technique solves this negative effect of non-volatile memory by using efficient data placement technique and hybrid memory architecture. In experimental section, the result shows that the proposed techniques improve energy saving up to 59.5% for the hybrid multibank memory architecture.