Fast simulation of turbo codes on GPUs

Simulation of turbo codes with moderately long block lengths down to bit error rates of the order of 10^-9 requires long runtimes on conventional CPUs. The approach described in this paper is based on a CPU/GPU co-processing strategy which aims at effectively distributing the processing tasks between the two platforms. In this paper, the most computationally intensive parts of turbo codes simulation are analyzed and their implementation on the GPU parallel architecture is discussed. Results on a case study for a serial concatenated convolutional code are presented, showing a simulation speed-up in excess of 10×. These initial results show that the CPU/GPU approach is a powerful tool that allows the characterization of the high SNR behavior of turbo codes with a short simulation runtime.