To meet emerging mobile traffic requirements, Centralized Radio Access Network (C-RAN) has been proposed to split the base station (BS) into two functional entities: the baseband units (BBU) and the remote radio heads (RRH). In C-RAN, by centralizing BBUs into BBU pools and leaving the RRHs in the cell sites, significant cost and energy savings and improved radio coordination can be achieved. However, C-RAN requires a costly high-capacity and low-latency access/aggregation network to support fronthaul traffic (i.e., digitized baseband signal). Hence, more recently, a new C-RAN architecture has been proposed (i.e., by 3GPP, IEEE 1914 WG), that defines three baseband function entities (or “splits”): central unit (CU), distributed unit (DU) and remote unit (RU). These three entities are expected to be interconnected by two external interfaces, called F1 and Fx. By transforming the RAN into a 3-layer (CU-DU-RU) architecture, more flexible deployment of the baseband functions can be achieved that better adapts to the heterogeneous characteristics of incoming 5G service requirements. It is also expected that, by properly placing CUs and DUs in the metro/aggregation network, higher benefits in terms of cost and power consumption can be achieved with respect to the previous 2-layer (BBU-RRH) architecture. In this paper, we investigate the optimal CU/DU placement problem in a 3-layer RAN architecture and formalize it by integer linear programming. We evaluate the benefits of the 3-layer architecture compared to the 2-layer architecture, showing that the consolidation degree of baseband processing depends heavily on fronthaul traffic latency, transport network capacity and processing capacity.