Finite Word-Length Effects in Implementation of a Block Floating-Point FFT Processor

This paper discusses issues of finite word-length effects in the block floating-point radix-8 FFT processor and presents a novel static model based on statistical analysis. When the input signal is white noise, the error caused by radix-8 unit and weight operation shows that the noise signal ratio (NSR) is proportional to r=log8N without the quantization error of input signal and coefficents, where N is the length of FFT. The paper also gives the error trend with the frequency point and transform stages. The simulation result from SPEED platform has a good consistent with the finite word-length effects analysis results. Consequently this scheme has better value in practice.