It mostly shows up in decode but if you run DP=2 plus expert parallel on smaller models you can get a nice pre-fill boost too but it's smaller than the decode boost, about 1.3x.