Fuse Devices for Pruning in Memristive Neural Network

In this study, we developed a fuse device for pruning implementation in a hardware neural network. A line-shaped fuse device was fabricated with aluminum metal and characterized in a one fuse and one resistive random-access memory (1F1R) structure. A blow time of 0.4 μs and read endurance of >10 7 were achieved, and the cut-off operation of the fuse was successfully verified in 1F1R. In addition, we developed a fuse design method for blow voltage and current via adjustment of the length, width, and thickness of the fuse. The adjustments were adopted to utilize the fuse in various synaptic devices. Finally, using simulations, we demonstrated a performance improvement due to the network pruning wherein the defective devices are disconnected by the fuse operations.