Zoom SFU, Simulcast & FEC Lab (Interactive)
Route packets, never mix pixels: price SFU versus MCU and push UDP loss at FEC parity. Run a gallery-view meeting through Zoom-style Multimedia Routers or a legacy MCU, tune packet loss against 45% FEC reconstruction capacity, and see simulcast layer selection shrink downlink bandwidth.
SFU Routing, Simulcast & FEC Lab
Route a 49-tile gallery through Zoom-style Multimedia Routers vs a legacy MCU, and push packet loss until the Reed-Solomon parity runs out.
How It Works Under the Hood
Human conversation breaks above roughly 150 milliseconds of round-trip latency, so conferencing cannot buffer and cannot use TCP head-of-line blocking. Zoom's proprietary UDP transport embeds Reed-Solomon parity packets that let receivers reconstruct lost audio mathematically with zero round trips, tolerating up to 45% loss. Meetings run on Multimedia Routers acting as selective forwarding units: no server-side decode or re-encode, just routing the publisher's simulcast layers — 1080p active speaker plus 180p thumbnails — according to each viewer's screen layout, which is how one platform went from ten million to three hundred million daily participants in months.
Core Architectural Principles
- SFU versus MCU economics: mixing and re-encoding caps a server near 15 participants; packet routing serves thousands at single-digit CPU.
- Simulcast downlink: one HD layer plus per-tile thumbnails versus naive all-HD, the source of the 80% bandwidth saving.
- FEC parity trades roughly 25% uplink overhead for loss recovery up to 45% with no retransmission latency.
In real-time rounds, choose forwarding over mixing with numbers: "An MCU re-encodes every stream; an SFU routes packets and holds thousands per box." Justify UDP by the 150 ms conversational threshold and explain FEC beats ARQ because retransmit costs a round trip. Mention simulcast plus active-speaker detection as the bandwidth story — that is exactly how Zoom survived the 2020 30x surge.
SFU routing with FEC minimizes server cost and interactive latency but shifts decode load to clients and wastes parity bandwidth on clean networks.