Brain-machine interface (BMI) is one of the most important tools in the neuroscience research and neuroprosthetics development. The investigation and development of BMI have achieved significant progress in the past decade. However, several bottlenecks from the electrical engineering perspective still have to be overcome. The next generation BMI system would feature bi-directional neural interface with on-chip neural feature extraction and machine learning. Moreover, the high integration, compact packing and wireless operation would allow the experiments in freely behaving animals. This paper reviews the state-of-the-art designs, summarizes the key design requirements and challenges in a BMI system, and provides insights in both circuit and system level design.