Implementation of ultra-low-power neural networks on quantized and pruned RRAM crossbar arrays

We demonstrate ultra-low-power spiking neural network (SNN) inference on an RRAM crossbar array by applying network lightweight techniques, and predict average power consumption using a highly accurate array-level model. A 24 × 24 crossbar array was fabricated using non-filamentary HT-RRAM, and quantized and pruned weights were transferred to the array. The compact model of HT-RRAM was used as a synaptic device to simulate a crossbar array model of the same scales the fabricated array, and the same network lightweight techniques were applied in the simulation. Both the crossbar array and the array model successfully transferred over 94% of the weights within an error margin of 2 nS, and the SNN inference results over 25 time steps showed highly consistent output currents. With this reliable array model, power consumption during MNIST inference was estimated for arrays with lightweight techniques applied. Based on our experimental results, the power consumption of image inference operations is predicted to be 243 nW with weight quantization only and 222 nW with weight quantization and pruning, across 10 classes. These findings suggest that ultra-low-power operation can be achieved in the RRAM array through the application of lightweight network techniques.