Do We Need to Normalize the Output?

I saw that you are normalizing the embedding output [here](https://github.com/deepinsight/insightface/blob/master/recognition/arcface_mxnet/verification.py#L321). Do we need to normalize the embedding? Since, if I removed the normalization part, the performance drop significantly. But, how does it run on the real-time implementation if we need to normalize the embedding? Thank you