The technology of using Tilted Transverse Isotropic Reverse Time Migration (TTI RTM) for subsalt velocity model building has been playing an important role in the seismic industry. Since TTI RTM is a computationally expensive technique, improving the program efficiency to meet the project turnaround schedule becomes a critical topic. In this paper, we will present an algorithm called "GPU based Layer Stripping TTI RTM". In layer stripping RTM, the model will be decomposed into two or more regions, horizontally. The wavefield redatuming for the top region wavefield will be saved as an input to the bottom region for RTM. In this approach, we do not need to repeat the shallow wavefield extrapolation; the grid size of the deeper migration can be increased and TTI RTM can be replaced by Vertically Transverse Isotropic (VTI) RTM, due to the fact that the dip field is not as sensitive in the deep region. This is a key to improving the efficiency of the RTM. The super parallelism of the GPU RTM plays another important role in the efficiency of the algorithm. In practice the application of GPU based layer stripping TTI RTM reduces the computation time by two orders of magnitude. In this paper, we will discuss some practical issues such as how to manage balancing the computer resources and the 3D data explosion of the redatumed wavefields.