I need to go through the gpu_utils.t file and make sure that we're actually using the most specific atomic and reduction instructions available. This is definitely not the case currently.