On the Accuracy of Analog Neural Network Inference Accelerators