We present a 192-channel 1D convolutional neural network (1D CNN) based neural feature extractor for Brain-Machine Interfaces (BMI) that achieves state-of-the-art decoding stability at $\mathbf{1.8\ \mu W}$ and 12801 $\mathbf{\mu m^{2}}$ per channel in 65nm CMOS technology. Our device is a fully configurable, scalable, area and power efficient solution that supports models with 2-8 feature layers and a total kernel length of up to 256. This architecture reduces caching requirements by $\mathbf{5}{\boldsymbol\times}$ over conventional computation schemes. Channels and layers are individually power-switchable to further optimize power efficiency for a given neural application. We introduce an on-chip model, FENet-66, that achieves the highest cross-validated decoding performance compared to all previously reported feature sets. We show that this model maintains superior stability over time using recorded data from tetraplegic human participants with spinal cord injury. Our features have 18% higher overall average cross-validated R2 decoding performance compared to Spiking Band Power (SBP), with 28% better performance during the 4th year. Our proposed architecture can also extract mean wavelet power features at low power and latency. We show that custom 1D-CNN kernels achieve 10% better performance compared to wavelet features while compressing the neural data stream by $\mathbf{38}{\boldsymbol\times}$. The models and hardware were validated in real time with a human subject in online closed-loop center-out cursor control experiments with micro-electrode arrays that were implanted for 6 years. Decoders using features generated with this work substantially improve the viability of long-term neural implants compared to other feature extraction methods currently present in low power BMI hardware.