Sustainable fault management and error correction for next-generation main memories

Abstract

As technology nodes continue to scale, main memories experience both increasing energy consumption as well as reliability challenges. In order to address rising failure rates due to problems with yield and runtime effects, such as crosstalk, due to process variation in small feature sizes, improved correction capabilities at the bit-level are increasingly essential. To address this challenge, we propose a sustainable approach to error correction in deeply scaled memories. In particular, we propose a novel area-efficient and sustainable fault map (SFaultMap) which targets holistic energy considerations to improve reliability while minimizing both operational and embodied energy. To demonstrate the effectiveness of SFaultMap we conduct a sustain-ability study, based on holistic energy consumption, to evaluate under which scenarios different solutions should be employed. In all cases and scenarios with moderate to high fault rates, SFaultMap has reduced energy over Error Correcting Pointers (ECP) for a five year lifetime. Moreover, as fault rate increases, the indifference time for ECP to recover upfront manufacturing energy increases from years to decades.