4F²/Bit Memristive Multi-Bit Content Addressable Memory Enabled by Nonlinear Encoding for in-Memory Similarity Search

Search density and energy efficiency are critical metrics for similarity search, particularly in resource-limited edge AI applications. In this study, we present an innovative nonlinear encoding scheme for Content Addressable Memory (CAM) to enhance density and energy efficiency. Utilizing a 2 R memristor CAM, this work leverages multi-level programmable states and similarity-based matching within a single CAM cell by a Euclidean-like distance function. The design achieves a high density of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$4 ~\mathrm{F}^{2} /$</tex> bit and an energy consumption of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.88 \text{fJ} / \text{bit}$</tex>. Experimental results demonstrate software-comparable accuracy in few-shot learning tasks, offering 213x improvement in energy efficiency compared with GPU, highlighting its potential for scalable, low-power edge AI applications.