Fig. 2: Setting the zero-reference point with differential read in Tiki-Taka version 2 (TTv2) and Chopped-TTv2 (c-TTv2).
From: Fast and robust analog in-memory deep neural network training
A Example ReRAM-like device response traces showing noise and variation in response to bi-directional pulses. Here we assume that the device gradually saturates with consecutive up or down pulses (see lower plot for pulse direction applied). Noise properties and update step sizes can be adjusted in the soft-bounds model Eq. (4) to e.g., reflect typical ReRAM (high noise), capacitor (medium noise, lower variation), or ECRAM (low noise) traces. B Due to the asymmetry, consecutive (pairwise) up-down pulses converge the conductance to a fixed point where up and down pulses are on average of the same size (symmetry point (SP), see Eq. (8)). Because of device-to-device variation each device has an individual SP value (dashed lines). C When the SP is estimated for each device of a crossbar array \(\breve{A}\), it can be programmed on a separate reference device \(\breve{R}\). Assuming that the circuitry allows for a matrix-vector multiplication with differential read, e.g., \({y}_{i}={\sum }_{j}\left({\breve{a}}_{ij}-{\breve{r}}_{ij}\right){x}_{j}\), then individual device responses are effectively set to zero when consecutive up-down pair pulses are applied.
