Kog is going deeper to squeeze more inference out of GPUs