Modern cores can have more sram for branch prediction than they do for L1I$.