true, for reductions you have algorithms like prefix scans, which produce a lot of output you're not necessarily interested in or algorithms which utilize fewer and fewer parts of the chips as the reduction progresses.