Fig. 4: Monte Carlo simulation of our proposed synapse is illustrated to observe the final READ distribution at different READ voltage.
From: Leveraging stochasticity in memristive synapses for efficient and reliable neuromorphic systems
The noise distribution of our synapse is illustrated with a 3σ range. a Shows the READ current distribution at 0.55 V, where the current is distributed from ~14 μA to ~29 μA. About 100% and 93.48% of data ‘0’ and ‘1’ are overlapped with each other. b Exhibits about 100% of data ‘0’ and 46.68% of data ‘1’ are contained noise. c Shows much lower noisy data for both ‘1’ & 0’. Finally, (d) Shows the data noise is minimal at 0.7 V. e Exhibits the data noise for ‘0’ is increased to 100% from 18% compared to the last test case (d). Hence, (f) illustrates higher noisy data at 0.8 V. g Shows the noise pattern of data ‘0’ and ‘1’ are 100% and 98.22% respectively. A sensing line is utilized to differentiate the READ current between data ‘0’ to ‘1’. Finally, an optimal operating region is found at 0.68 V. h Shows an optimal operating region for READ operation, where data distribution illustrates a reliable READ operation for a neuromorphic system. Traditionally, highly noisy data is not useful, but data distributions like (a, b, e, f, g) can be utilized for applications where a biased dataset can enhance the performance. Moreover, this kind of stochasticity might be useful for neuromorphic applications. In addition, (c), (d), and (h) can be used to develop a reliable neuromorphic system.
