Memory errors are an important threat to computer system reliability. In production sites running large-scale systems, memory errors are one of the most common hardware problems that lead to machine crashes and are ranked near the top of component replacements. Memory errors are projected to increase in future systems not only because of the continuous increase in capacity, but also because of the integration of emerging non-volatile memory technologies within the memory hierarchy. As conventional memories scaling is in jeopardy, researchers are investigating the use of emerging technologies such as Phase-Change Memory (PCM) and Memristors as potential future scalable replacement for their traditional counter-parts. These memories promise several common advantages such as scaling, high density, and non-volatility. In addition, each technology has its unique properties such as analog-computations for Memristors and temperature-friendliness for PCM.

However, in order for these memories to be used as a successful replacement for conventional memories, we need to address their major challenges such as limited endurance, high defect rates and high write energy. In our work, we explore potential directions to improving the reliability of DRAM as well as new memory technologies. Our work is guided by current trends in designing reliability features and major milestones in emerging memory technologies. We build a modeling tool to guide the design of future reliability, availability and serviceability (RAS) features. Furthermore, we present two case- studies that exploit materials properties and unique characteristics to address their reliability concerns and/or improve their energy efficiency. The first case-study explores using Memristors to implement a highly accurate neural branch predictor.

This application relies on Memristors' high density and ability to do analog computations efficiently to implement a highly accurate predictor that lasts 5-years under a high frequency of updates and in presence of high defect rate. The second case-study explores using Phase-change memories to build heterogeneous 3D-memory stack. We exploit the temperature variation across layers within the 3D-stack to enable building heterogeneous stack. This improves the write energy by up to 3.5X and extends the lifetime of the memory system by 30%.