Architecture Of A Neural Processor For Fourth Generation ´4g´ Mobile Terminals

Pipelined recurrent neural network (PRNN) has been used with lot of success in many applications. In recent works, we have also proven that the PRNN exhibits good performances when used for interference cancellation and channel parameters estimation for the different multiple access schemes proposed as physical layer of the fourth generation (4G) networks: wideband code division multiple access (WCDMA); orthogonal frequency division multiplexing (OFDM); and multi carrier CDMA (MC-CDMA). The use of a unique PRNN based module for the three multiple access techniques addresses a major challenge for the 4G mobile terminals: embedding many access techniques for reduced area and resources costs. In this paper, after a brief review of the performances of the PRNN based module used in the different multiple access receivers, we propose a mapping scheme for implementing the PRNN on a bi-directional systolic array. Finally, a neural network processing element architecture is derived