Figure 4.
Image processing demonstration using Ag2S FM array. (a) Convolutional kernel values are encoded into either an FTM or ITM Ag2S device array. The grayscale value of each pixel of the original image is mapped to the input voltage, which is fed to the top electrodes of different devices via a 3 × 3 input block. The postsynaptic current after MAC operation is recorded at the bottom electrode and back-transferred to grayscale values to generate the output image. (b) Software-based simulation result (i) and hardware outputs of FTM (ii) and ITM (iii) after sharpening operation. Software-based simulation result (iv) and hardware outputs of FTM (v) and ITM (vi) after softening operation. (c) Retention measurement of conductive states used in the image processing. (d) Measured (from the hardware) and arithmetic current (from the simulation) of FTM-1 under input reading voltages. The inset shows the current deviation distribution and its average value. Since the output current around ±0.4 mV is close to the background noise level, a significant deviation is generated when applying small reading voltages. This only contributes to the error for grayscale values below 4 in our demonstration. (e) Measured (from the hardware) and arithmetic current (from the simulation) of ITM-1 under input reading voltages. The inset shows the current deviation distribution and its average value. (f) Pulse power against elapsed time in the kernel encoding process for FTM-1. The energy consumption is calculated by integrating the pulse power against the device setting time. (g) Pulse power against elapsed time in kernel encoding process for ITM-1.