Spike sorting is often required for analyzing neural recordings to isolate the activity of single neurons. Wave shape analysis in the spike sorting procedure provides a means to detect spikes while minimizing the influence of false alarms. As neural recording techniques allow for recording hundreds of electrodes, power is too limited in neural implants for current spike sorting algorithms. Even with spike sorting at the back-end where more power is available, the bandwidth is too limited to transmit enough information for current spike sorting techniques. A low-power pulse-based feature extractor presented in the paper is a solution to the bandwidth bottleneck. It reduces the bandwidth of the neural signal by several orders of magnitude while preserving enough information for spike sorting.