A Universal Framework for Reliable Computing-in-Memory based on Emerging Non-volatile Memories (CIMware)

点击翻译按钮获取中文摘要

Computing in Memory (CIM) is an emerging concept based on the tight integration of traditionally separated memory elements and processing cores, allowing minimizing time and energy needed to move data across digital architectures. Among emerging device technologies, non-volatile memory technologies such as resistive or magneto-resistive devices (STT, ReRAM and PCM) that are able to act as both storage and information processing units, favor increasing system complexity and performance at lower power consumption. Therefore, they provide the scientific community with opportunities for new computer architecture innovations being able to track todays limitation.While there are different flavors of implementing CIM operations using different memristive technologies, the architectural design of CIM computing paradigm and their integration within the memory hierarchy are open research problems. Additionally, despite the promising nature of the in-memory computing based architectures built with emerging devices, many issues related to the devices themselves and to their dual usage (storage and computing unit) have still to be solved. Due to immaturity of these emerging memory technologies – compared to more than 50 years of CMOS maturity – as well as inherent device stochasticity, it is difficult to meet today's highly demanding reliability requirements for the new technologies. In addition, performance and cost can be negatively affected by uncertainties in reliability. These issues introduce aggressive challenges on CIM reliability, which may mandate new models of CIM-based computing at higher abstraction levels. This project aims at providing a bridge between technological developments for emerging non-volatile memories for reliable and energy-efficient CIM realization from one side and architecture-level design exploration and computing paradigms from the other side. This is done by designing a library of reliable and energy-efficient primitive circuits based on various non-volatile technologies (memristive and spintronic devices), not only in order to enable computation-in-memory architectures, but also to provide accurate technology-aware architectural models to benchmark and explore the potential of each technology (e.g., MRAM, PCM, ReRAM) and CIM architecture designs to enable further development in software and compiler design and explorations for computing-in-memory paradigm.