Non-Binary Low-Density-Parity-Check codes (NB-LDPC) have shown superior performance but its huge complexity and low throughput prevent it from practical applications. This paper presents a novel architecture to implement a kind of high-throughput low-complexity irregular quasicyclic NB-LDPC decoder over GF (16) based on Extended Min-Sum (EMS) Algorithm. Double clocks are adopted in this paper. The low frequency clock at 60 MHz serves as the system clock and the high frequency clock at 480 MHz works for check nodes and variable nodes thus they can be reused 8 times during one system clock period as a result the complexity can be largely reduced. Synthesis result shows that the throughput can achieve 68.57 Mbps at 5 iterations. FPGA testing result shows that the decoder has little performance degradation compared with its floating model and it can provide about 0.5 dB coding gain compared with its binary LDPC code of the same block length. The proposed architecture can be conveniently extended to higher Galois Filed such as GF (64) or GF (256) and it can be applied for different code length and rate by modifying a slight part of the decoder. Compared with previous works, the decoder proposed in this paper is more efficient for practical applications.