This paper analyzes different computational methods for real-time decoding of neural signals in primary motor cortex (M1). Specifically, we compare different classifiers as well as different Principal Component Analysis (PCA)-based preclassification strategies to identify how to proceed in terms of the necessary trade-off between computational complexity and accuracy. Our methods are applied to neural data in monkey, recorded while performing dexterous hand and finger movement tasks. We show that differences due to selection of a classifier using the same feature set are statistically significant for reduced sets of neurons, and specifically that neural networks are to be preferred to a linear classifier. Furthermore, we show that using PCA-based methods prior to neural network-based classification yields statistically equal real-time decoding accuracies using less than 20% of principal components. We therefore conclude that performing PCA prior to classification with a smaller feature space statistically provides the same or better decoding accuracies as those obtained using a larger feature space and a linear or non-linear classifier.