An SFU (Selective Forwarding Unit) is the media-server architecture that makes large WebRTC conferences practical. Each participant sends their audio and video once to the SFU, which then forwards those streams to the other participants without decoding or mixing them. Keeping streams separate means the server's CPU cost is low (it just routes packets), every client can render participants individually, and the media stays encrypted between endpoints better than a mixing server allows. For audio, an SFU reads the per-packet audio-level marking to forward only the active speakers rather than everyone, saving bandwidth. SFUs are the dominant design for group calls today, sitting between the cheap-but-unscalable mesh and the heavy, latency-adding MCU.
