SERVING RECURRENT NEURAL NETWORKS EFFICIENTLY WITH A SPATIAL ARCHITECTURE