Compute-in-memristor: A System and Technology Co-Optimization (STCO) Approach

Summary form only given, as follows. The complete presentation was not made available for publication as part of the conference proceedings. The reviving of artificial intelligence (AI) technology is changing our lives due to the superior performance in computer vision, nature language processing, and decision making and control tasks. Recently, computing in memristor (CIM) technology is considered as an unconventional approach to meet the urgent needs of high computing efficiency and throughput in continuously developing AI applications. In contrast to von Neumann architecture, computing in memristor technology avoids the data movement between computing units (such as CPU and GPU) and memory unit which would greatly reduce computing latency and power consumption. Memristor is one kind of novel non-volatile memory devices which could not only store information with multi-bit, but also conduct core computing operations of AI algorithm. Assembling memristors in a crossbar structure, it could carry out in-situ parallel multiply-accumulation (MAC) computation. To make the best use of the memristors in neuromorphic systems, a memristor-friendly architecture and the software-hardware collaborative design methods are essential, and the key challenge is how to utilize the memristor’s analog behavior. The talk will discuss recent development of analog computing in memory technology from device engineering, circuit design and architecture optimization perspectives. Further, the system and technology co-optimization (STCO) method is proposed to fully unleash the potential of CIM for practical applications.