3D Bridge Super-resolution Crossbars

The work proposes a novel three-dimensional bridge super-resolution crossbar for improving the accuracy of analog-dot product computations for the Multiply and Accumulate operations in hardware neural network implementations. Compared to the series and parallel combinations of memristors used to create multi-level conductance states to map the neural weights, the 3D bridge configuration offers approximately tenfold higher resolution. The reduced relative current error by a more precise mapping process due to higher resolution makes the 3D-Bridge super-resolution crossbar tolerate the memristor non-idealities with significantly reduced area and power. The proposed approach is tested with artificial and convolutional neural networks with near-ideal inference accuracy, making the three-dimensional architecture flexible for high-precision, energy-efficient hardware neural network implementations.