Parasitic Resistance Effect Analysis in RRAM-based TCAM for Memory Augmented Neural Networks

Memory augmented neural networks (MANNs) enable trained neural networks to rapidly learn new classes from few examples. However, content-based addressing is inefficient in conventional computer system due to the von Neumann bottleneck. Ternary content-addressable memories (TCAMs) based on resistive random access memory (RRAM) provide a promising approach to accelerate the addressing according to the Hamming distances (HDs) between the search vector and stored vectors. Generally, the HD is sensed from the discharge rate of a match line, exhibiting a linear dependence on the number of mismatched bits. However, parasitic resistance effect causes that the location of mismatched bits also determines the HD. This work proposes a compact model to evaluate the discharge rate of a match line. The impact of parasitic resistance effect in RRAM- based TCAMs is analyzed. Parasitic resistance effect is also incorporated into MANNs during few-shot learning. Remarkable accuracy losses are observed as parasitic line resistance and columns of TCAM increase. Our analyses provide valuable design guidelines of RRAM-based TCAM for future MANN systems.