In-memory computing: characteristics, spintronics, and neural network applications insights

In today's digital computing landscape, In-Memory Computing (IMC) has emerged as a revolutionary approach to tackling critical energy efficiency and latency challenges, particularly in dealing with the widespread bottleneck between memory processors. While the core concept of IMC seems clear and promising, its practical application involves many complexities, encompassing a wide range of issues and solutions. This paper aims to act as a guide, navigating through the intricate terrain of IMC. This paper gives a concise yet thorough explanation of IMC, explaining its basics and essential features. It delves into the critical trade-offs between IMC and traditional memories, spotlighting the various aspects involved in bandwidth, energy, and latency dynamics. This paper discusses IMC applications in the Neural Networks (NN) domain and the integration of IMC with spintronics. Leveraging the distinctive capabilities of spintronic memory, the simultaneous activation of multiple word lines within an array has become achievable. Integrating IMC with non-volatile applications has demonstrated notable results, achieving a 98% maximum accuracy. In addition, a maximum throughput potential of around 660 GOPS and an energy efficiency of 66% has been achieved. Lastly, this paper delves into IMC's primary challenges and opportunities, providing a comprehensive understanding of this innovative computing paradigm.