We present a graph matching algorithm that mitigates relaxation bias through a regularization term and achieves substantial acceleration via an innovative mixed-precision architecture.