The detectors of the CBM experiment, under construction at GSI, Germany, will generate an enormous data flow of 1TB/s, which is provided via approximately 6000 optical links to the DAQ system. For proper analysis, so called events have to be composed of data packets originating from different geometrical regions and with temporal coincidence, at a resolution of 1 nS. Routing and aggregating of packets is performed in a complex custom interconnect structure, where an active buffer is the interface to the commodity DAQ farm. We present the prototype of an active buffer board (ABB) in PCIe format. On each ABB, data from 2 optical links (each 2.5 or 5 Gbit/s) are received by an FPGA and, the packets being sorted according to spatial and temporal information and stored into a local buffer. To enable simultaneous readout towards the host, an efficient dual-port memory emulation was implemented, which is non trivial due to the small packet sizes and the high bandwidth requirements. Our advanced memory arbitration mechanism matches full incoming bandwidth and a concurrent DMA towards PCIe of 600 MB/s for a 4-lane PCIe interface, using an ordinary DDR2 memory module. In parallel to the data path the ABB implements a control flow interface, where requests can be sent from the host towards the detectors and in turn response messages are being received. The corresponding I/O mechanisms are implemented via DMA or PIO. At the software level, the DMA channels are controlled by a customized Linux device driver. Initial integration of the prototype ABB into the GSI software framework has been performed.