Abstract

Multimode fibers (MMF) recently received considerable attention to realize space-division multiplexed (SDM) transmission systems, in order to increase the capacity of optical transmission systems. The nonlinear signal propagation in MMFs systems can be described by the nonlinear Schrödinger or the Manakov equation [1]. These are commonly solved numerically by applying split-step Fourier methods [2], which can be accelerated by GPUs. To simulate the nonlinear interaction between the channels in a mode-division multiplexed (MDM) transmission system, sufficiently long symbol sequences have to be simulated. Further, the number of samples per symbol has to be increased for the simulation of multiple wavelength channels [3]. However, the required memory for the simulation of MDM systems scales linearly with the number of modes. Thus, highly memory efficient split-step Fourier implementations are required. The iterative approach, described in [2], allows such a memory efficient implementation. However, the ability to simulate large scale MDM systems is still limited by the available GPU memory. Since most commonly all calculations are executed with double-precision, the question arises if single-precision is sufficient for parts of the calculations or, in the context of memory limitations, for saving intermediate results without introducing an unacceptable inaccuracy.

© 2017 IEEE