I believe it's moreso for the (actively pursued) speed optimizations it provides. When inference is already computationally expensive any bit of performance is a big plus