Towards an efficient hardware implementation of recurrent neural network based multiuser detection

Code division multiple access (CDMA) will be an important part of 3rd generation mobile communication systems worldwide. Besides its many advantages, CDMA has the disadvantage that CDMA systems are interference limited. However, this limitation can be overcome by proper multiuser detection (MUD). We propose a MUD based on recurrent neural networks (RNN). This algorithm provides a performance which is close to the optimum MUD, while keeping the computational complexity low. Furthermore, the regular structure of the algorithm facilitates a hardware (HW) implementation utilizing distributed computation. The HW platform we target at are field programmable gate arrays (FPGA). Compared to application specific integrated circuits, FPGAs allow a much faster and more flexible hardware implementation. We have implemented the algorithm with a fixed point arithmetic and investigated the minimum number of binary digits required so that the performance of the algorithm does not degrade substantially. Furthermore an efficient implementation of the tanh nonlinearity is presented.