Statistical Study on Data Dependency of Memristor Crossbar Architectures for Neuromorphic Computing

In this work, we statistically present the data density dependency of the two memristor crossbar structures, namely the complementary crossbar architecture and the twin crossbar architecture, for neuromorphic image recognition. The two structures perform the Exclusive-NOR computing to find the similarity score between the input vector and the memristor arrays, and both employ the Winner-take-all circuit to determine the winning current. When considering the power consumption, the twin crossbar architecture has proven to consume less energy than complementary crossbar architecture. However, by analyzing the mathematical model of the circuit, we figured out the reduction of output currents as the input image has low density. The observation was verified by the simulation with the data densities of 0.5, 0.4, 0.3, and 0.2. When the data density decreases to low level of 0.2, the maximum current of columns yielding by the twin crossbar architecture decreases about 80% compared to the complementary structure. The Winner-take-all circuit could operate incorrectly under this significant reduction and it could degrade the performance of the twin crossbar architecture. In contrast, twin crossbar architecture can save more power consumption with low-density data. When the density of data is of low of 0.2, the power consumption of the twin crossbar architecture reduces by around 79% in comparison with the complementary structure. These results prove that although being affected by the low-density of data, the twin crossbar architecture has an advantage of low power consumption when working on low-density datasets compared to the complementary crossbar architecture.