Also in NLP domain a precision reduction may be applied: https://arxiv.org/abs/1706.06363