The polyhedra abstract domain is one of the most powerful and commonly used numerical abstract domains in the field of static program analysis based on abstract interpretation. In this paper, we present an implementation of the polyhedra domain using floating-point arithmetic without sacrificing soundness. Floating-point arithmetic allows a compact memory representation and an efficient implementation on current hardware, at the cost of some loss of precision due to rounding. Our domain is based on a constraint-only representation and employs sound floating-point variants of Fourier-Motzkin elimination and linear programming. The preliminary experimental results of our prototype are encouraging. To our knowledge, this is the first time that the polyhedra domain is adapted to floating-point arithmetic in a sound way.