Thank you for this great article! I am still having some trouble understanding the greater than linear scaling, is this because each GPU has less data to hold and thus can use more cores to build trees?