AWARD: Approximation-aWAre Restore in Further Scaling DRAM

Abstract

DRAM further scaling becomes more and more challenging, making restore operation an serious issue in the near future. Fortunately, a wide range of modern applications are able to tolerate error or inexactness, providing a new dimension to mitigate the slow-restore issue. And thus, we can trade-off acceptable QoS loss in those applications to accelerate restore operations, and further to achieve performance and energy improvements. In this extended research abstract, we briefly explore DRAM restore-based approximate computing, and present a preliminary evaluation on impacts of quality-of-service (QoS) degradation and performance speedup. We show that restore-based approximate computing is a challenging work, and dedicated error correction/tolerance techniques are needed to balance QoS and performance.