Exploiting Partially-Forgetful Memories for Approximate Computing

While the memory subsystem is already a major contributor to energy consumption of embedded systems, the guard-banding required for masking the effects of ever increasing manufacturing variations in memories imposes even more energy overhead. In this letter, we explore how partially-forgetful memories can be used by exploiting the intrinsic tolerance of a vast class of applications to some level of error for relaxing this guard-banding in memories. We discuss the challenges to be addressed and introduce relaxed cache as an exemplar to address these challenges for partially-forgetful SRAM caches. Preliminary results show how adapting guard-bands to application characteristics can help the system save significant amount of cache leakage energy (up to 74%) while still generating acceptable quality results.