The precision required for neural network algorithm is an important question for hard-ware architects. This paper gives an excellent high speed BP algorithm for feed-forward muitilayer neural network using limited precision calculation. The experimental results show that the algorithm improves convergence speed with reduced precision bits.