Training Method for Accurate Off-Chip Training of One-Selector-One-Resistor Crossbar Array with Nonlinearity and Wire Resistance

This work provides an off-chip training method for a one-selector-one-resistor (1S1R) crossbar array (CBA) device with wire resistance (rcc) and nonlinear conductance (g (i,j)) of 1S1R devices for hardware neural network (HNN) applications. An iterative method is introduced to calculate the node voltages of the 1S1R CBA, which arises from the variable voltage drop through the wires with rcc and g (i,j). Several mathematical approximations are introduced for fast and efficient calculation. The proposed method trains the HNN to have an inference accuracy of 85.9%, whereas the inference accuracy of HNN without the rcc consideration drops to 38.5%. The inference running time with the proposed method is 1% of the HSPICE-based simulation for the given HNN structure. As the rcc increases, the inference accuracy declines due to the decreased device voltage from the target values. The worst voltage model is adopted to identify the design factors that affected the accuracy. A CBA with a size almost three times larger can be used for the HNN if the rcc is appropriately addressed under the given device conditions.

Automated summary

This work presents an off-chip training method for a 1S1R CBA device with mathematical approximations for fast calculation, achieving an inference accuracy of 85.9% for the HNN. The proposed method also identifies design factors affecting accuracy and significantly reduces inference running time.