Convolution neural network inference using frequency modulation in computational phase-change memory

In [1] we reported for the first time a frequency modulation method to control the conductance level in PCM cells. This increases the programming reliability of PCM, which is crucial for neuromorphic applications. We provided a physical picture and a physics-based analytical model to link the programming frequency to a target conductance. In this new report, we are the first to successfully demonstrate frequency modulation on 16kbit PCM array in a real case scenario. We first convert synaptic weights in target conductivities, next we translate them into modulation frequencies and transfer values to the PCM array. Eventually, we evaluate the accuracy of a test CNN (Convolutional Neural Network) based on pre-programmed PCM values for recognizing handwritten digits from the MNIST database. We evaluate different redundancies schemes and show up to 90% accuracy and high reliability. We complement our model after careful characterization of the distribution in the programming error (|G
target
-G
pcm
|) and show that we can fully predict the inference accuracy. PCM drift characterization shows good data stability over 24 hours at room temperature. Eventually we propose a block schematic of a FSM (Finite State Machine) to implement the frequency modulation on-chip, showing the benefit of such an approach as ease of design.