NOR Flash-based Multilevel In-Memory-Searching Architecture for Approximate Computing

Abstract

In the approximate in-memory-searching (IMS) system, the numeral-system-induced matching score counting error will cause computation error and degrade the system performance significantly especially at high resolution scenarios. This paper discusses the problem in detail and proposes a novel NOR-Flash-based multilevel IMS cell and the related operation scheme for approximate computing. The device characteristics and reliability performances are carefully evaluated on a high density NOR Flash array. Experiments and simulations suggest that the proposed approximate IMS cell not only can provide the matching score based on the difference between the input and the stored data, but also improve the system robustness toward data variation.