An Energy Efficient 3D-Heterogeneous Main Memory Architecture for Mobile Devices.

The demand for main memory capacity is ever increasing in mobile devices and embedded systems. Dynamic Random Access Memories (DRAMs) can not keep pace with the required main memory capacities because of the restrictions in improving the cell density due to the slowdown in scaling and the high leakage power consumption. Contrary, emerging Non-Volatile Memories (NVMs), primarily Resistive Random Access Memories (RRAMs), offer a high scaling potential and consume less leakage power than DRAMs. However, they are not suitable to completely replace DRAMs as the main memory, owing to their large read and write access latencies and limited endurance. In this paper, we present the architecture of a novel heterogeneous 3D-stacked on-chip main memory system composed of DRAMs and RRAMs that can fulfill the memory capacity demands of future mobile devices. We evaluate the energy savings of the new architecture for several applications, including some emerging machine learning tasks on mobile devices, by conducting system-level simulations in gem5 using ARM CPU models. We explore and analyze the impacts of different hybrid memory organizations and data allocation policies on reducing the energy and total number of RRAM writes. On average, the new 3D-hybrid architecture consumes 73% lesser energy and 61% lower average power than a 2D-Hybrid memory architecture for applications from the PARSEC benchmark. For a neural network training application, the 3D-hybrid memory saves up to 60% energy in comparison with a DDR4 DRAM-only main memory.