Deep face recognition has made remarkable advances in the last few years, while the training scheme still remains challenging in the large-scale data situation where many hard cases occur. Especially in the range of low false accept rate (FAR), there are various hard cases in both positives ($\textit{i.e.}$ intra-class) and negatives ($\textit{i.e.}$ inter-class). In this paper, we study how to make better use of these hard samples for improving the training. The existing training methods deal with the challenge by adding margins in either the positive logit (such as SphereFace, CosFace, ArcFace) or the negative logit (such as MV-softmax, ArcNegFace, CurricularFace). However, the correlation between hard positive and hard negative is overlooked, as well as the relation between the margin in positive logit and the margin in negative logit. We find such correlation is significant, especially in the large-scale dataset, and one can take advantage from it to boost the training via relating the positive and negative margins for each training sample.To this end, we propose an explicit cooperation between positive and negative margins sample-wisely. Given a batch of hard samples, a novel Negative-Positive Cooperation loss, named NPCFace, is formulated, which emphasizes the training on both the negative and positive hard cases via a cooperative-margin mechanism in the softmax logits, and also brings better interpretation of negative-positive hardness correlation. Besides, the negative emphasis is implemented with an improved formulation to achieve stable convergence and flexible parameter setting.We validate the effectiveness of our approach on various benchmarks of large-scale face recognition and outperform the previous methods especially in the low FAR range.