Engineering Strategies in HfOx RRAM-Based Analog Synapses Toward Linear Weight Update for Neuromorphic Hardware Accelerators

This study introduces two distinct engineering approaches for HfO
x
-based resistive memories (RRAMs) that can be implemented in various manners for neuromorphic hardware platforms. For highly dense cross-point synaptic arrays, we show that a gradual weight update can be achieved even in select-or-free RRAM by integrating a thin Al
2
O
3
nonlinear barrier. Meanwhile, in configurations that use conventional transistors as selectors, the weight update as quick as possible plays an important role in accelerating training process. We thus reveal that the abundant oxygen vacancies in sputtered HfO
x
enables nanosecond weight modulation driven by identical pulses.