Reliability, scalability and clinical viability are of utmost importance in the design of wireless Brain Machine Interface systems (BMIs). This paper reports on the design and implementation of a neuroprocessor for conditioning raw extracellular neural signals recorded through microelectrode arrays chronically implanted in the brain of awake behaving rats. The neuroprocessor design exploits a sparse representation of the neural signals to combat the limited wireless telemetry bandwidth. We demonstrate a multimodal processing capability (monitoring, compression, and spike sorting) inherent in the neuroprocessor to support a wide range of scenarios in real experimental conditions. A wireless transmission link with rate-dependent compression strategy is shown to preserve information fidelity in the neural data. At 32 channels, the neuroprocessor has been fully implemented on a 5mm×5mm nano-FPGA, and the prototyping resulted in 5.19 mW power consumption, bringing its performance within the power-size constraints for clinical use. The optimal design for compression and sorting performance was evaluated for multiple sampling frequencies, wavelet basis choice and power consumption.