Separate thread for handling the CUDA parts seems a little better.