Opportunities for Processing Near Non-Volatile Memory in Heterogeneous Memory Systems

Growth of data volumes in application domains such as highperformance computing, machine learning, and data analytics places increasing demands on main memory capacity of computing systems. Emerging byte-addressable, non-volatile memory (NVM) technologies can enable increased main memory capacities to meet these application demands more cost-effectively than is possible with DRAM alone due to their superior cell densities and lower cost-per-bit. However, NVM delivers lower memory bandwidth and incurs higher access latency than DRAM, especially for writes. As a result, system organizations that incorporate NVM typically use it as a large, “second-level” memory in addition to a smaller pool of DRAM. In this work, we have evaluated and compared the performance of different processing near memory systems using DRAM and NVM. Our simulation results show that, for several classes of applications with large datasets that cannot fit in DRAM, there can be significant amounts of data movement between DRAM and NVM. Therefore, depending on application characteristics, processing near NVM can improve energy efficiency and performance over processing near DRAM despite the limitations of NVM such as lower bandwidth and higher write energy. We show that processing near NVM can be up to a few orders of magnitude more energyefficient (based on energy-delay product) than processing near DRAM when the DRAM capacity can only accommodate a fraction of the application’s dataset.