New DRAM technologies such as SDRAMs, RDRAMs, EDRAMs, CDRAMs and others are vying to be the next standard in DRAMs and improve upon bandwidth limit of conventional DRAMs. With proliferation of power-aware systems, banked DRAM architecture has emerged as a promising candidate for reducing power. Prior work on optimizing applications in a banked memory environment has exclusively focused on uncompressed data. While this may be preferable from a performance viewpoint, it is not necessarily the best strategy as far as memory space utilization is considered. This is because compressing data in memory may reduce the number of memory banks it occupies and this, in turn, may enable a better use of low-power operating modes. In this paper, we explore the possibility of compressing infrequently used data for increasing effectiveness of low-power operating modes in banked DRAMs. Our experiments with five highly parallel array-based embedded applications indicate significant savings in memory energy over a technique that exploits low-power modes but does not use data compression/decompression.
展开▼