Reconfigurable Intelligent Surfaces (RISs) are an emerging technology for future wireless communication systems, enabling improved coverage in an energy efficient manner. RISs are usually metasurfaces, constituting of two-dimensional arrangements of metamaterial elements, whose individual re-sponse is commonly modeled in the literature as an adjustable phase shifter. However, this model holds only for narrowband communications, and when wideband transmissions are utilized, one has to account for the frequency selectivity of metamaterials, whose response usually follows a Lorentzian-like profile. In this paper, we consider the uplink of a wideband RIS-empowered multi-user Multiple-Input Multiple-Output (MIMO) wireless sys-tem with Orthogonal Frequency Division Multiplexing (OFDM) signaling, while accounting for the frequency selectivity of RISs. In particular, we focus on designing the controllable parameters dictating the Lorentzian response of each RIS metamaterial element, in order to maximize the achievable sum rate. We devise a scheme combining block coordinate descent with penalty dual decomposition to tackle the resulting challenging optimization framework. Our simulation results reveal the achievable rates one can achieve using realistically frequency selective RISs in wide band settings, and quantify the performance loss that occurs when using state-of-the-art methods which assume that the RIS elements behave as frequency-flat phase shifters.