Wild Dragon
Guide · Broadcast IP Engineering

SMPTE ST 2110 facility migration: from SDI to IP without going off air

How we design, phase, and commission SDI-to-IP migrations: fabric topology, PTP timing, ST 2022-7 redundancy, NMOS control, real bandwidth math, and the failure modes that show up in the field.

By Zachary Gaetano, Wild Dragon LLC · Updated November 17, 2026

When migration makes sense

SDI is simple and deterministic: a 3G-SDI link carries 2.97 Gb/s of multiplexed video, embedded audio, and ancillary data, and a router switches it on the vertical interval. The limits show up at scale. 12G-SDI for UHD needs premium coax with short runs, every format change means new cable, and pulling a single audio channel out of a multiplex needs de-embedders.

SMPTE ST 2110 splits each signal into separate essence streams over RTP. ST 2110-10 is the system layer underneath all of them: PTP-derived RTP timestamps, SDP signaling, and the standard UDP datagram size (1460 bytes or less; extended jumbo payloads are optional and must be supported end to end, so standard MTU is the safe default). The essence standards are ST 2110-20 uncompressed video, -30 PCM audio (AES67-compatible), -31 AES3 transparent audio, -40 ancillary data, and -22 compressed video. The Ethernet fabric becomes the router. That buys format-agnostic cabling, flexible audio, and room to grow, but it moves the engineering burden to timing, bandwidth, multicast, and control. If you only need a few more SDI inputs, a hybrid router may be the better answer; IP pays off when you are building new rooms, need UHD, or want to share a plant across rooms and sites.

Fabric topology: spine-leaf vs. monolithic core

Monolithic core. One large modular switch, every endpoint home-run to it. It is simple to reason about and has no inter-switch links to oversubscribe, but it demands long fiber runs, is expensive to expand, and concentrates failure in one chassis (which Red/Blue then has to duplicate).

Spine-leaf. Leaf switches sit near the endpoints (studio, control room, rack rooms) and uplink to two or more spines. It scales by adding leaves, shortens cable runs, and keeps failure domains small. The price is that uplinks must be sized deliberately: for uncompressed flows we design leaf-to-spine capacity to be non-blocking, not the 3:1 or 5:1 oversubscription typical of enterprise networks. As an example, thirty 1080p59.94 senders on one leaf is roughly 78 Gb/s of video, so 2×100G uplinks carry it comfortably, but lose one spine and the surviving 100G link is already tight once you add audio, ancillary data, and headroom. Size for the failure case, not the happy path, and favor switches with buffer depth suited to multicast media rather than enterprise defaults.

Whichever you choose, check the switch's multicast replication and buffer behavior, not just port speed. Packet bursts from many senders converging on one egress port are what exhaust shallow buffers.

Multicast, IGMP, and SDN orchestration

Every essence is a multicast group. A receiver "takes" a source by joining its group, using IGMPv3 on the edge (source-specific joins) and PIM-SSM across a routed core. Two design points matter:

  • Querier placement. Without an active IGMP querier on each VLAN, snooping state ages out and streams stop after a few minutes.
  • Flow placement. ECMP hashes flows onto uplinks without knowing their size. A handful of 2.6 Gb/s or 10 Gb/s flows can land on one link while others sit idle.

Pure IGMP/PIM works for smaller fabrics with conservative headroom. For denser plants we use SDN or broadcast-controller orchestration (vendor-specific products from switch and broadcast-controller vendors) so a route request is checked against available bandwidth before the join happens, and so operators see a router-style view rather than a multicast table.

ST 2022-7 and Red/Blue network design

SMPTE ST 2022-7 sends two identical RTP streams over separate paths; the receiver merges them packet by packet, so losing either path causes no visible hit. The design rules that make it real:

  • Two physically independent fabrics (Red and Blue): separate switches, power, and trunks, with no inter-switch link between them.
  • Separate subnets and multicast ranges per fabric so a misconfiguration on one cannot leak into the other.
  • Dual-homed endpoints with two NICs; single-NIC devices are fed through a gateway or accepted as a known single point of failure.
  • Receiver buffer sized for the worst-case Red/Blue path difference. Too small and the merge fails; too large and you add latency.

ST 2022-7 is not link aggregation. LACP aggregates links under one control plane and has to reconverge after a failure; ST 2022-7 has two independent networks and no reconvergence event for the receiver.

PTP timing: IEEE 1588 and SMPTE ST 2059

IP facilities distribute time over the same Ethernet using the Precision Time Protocol (IEEE 1588), profiled for media by SMPTE ST 2059-2. SMPTE ST 2059-1 defines how video frame and audio sample alignment is derived from the PTP timescale, whose epoch is 1970-01-01 00:00:00 TAI. Every sender timestamps its RTP packets against that clock, which is what lets receivers align sources from different cameras and rooms without genlock cabling.

  • Grandmasters. Two or more GNSS-locked grandmasters, ideally on separate antennas, with explicit priority settings so the best-master election is deterministic. Red and Blue each need their own path to a grandmaster, through physically separate GM ports or a GM per network. Never bridge the two networks through a shared switch or NIC to reach timing; that defeats ST 2022-7. Stay on the ST 2059-2 profile (default domain 127) consistently across every switch and endpoint.
  • Boundary or transparent clocks. We normally specify boundary clocks on every switch in the media path, so each hop terminates and regenerates PTP. A switch that does not support PTP correctly in the path is a defect, not a detail.
  • Symmetry. PTP assumes equal delay in both directions. Asymmetric paths, uncompensated firewalls, or media converters introduce a constant offset that looks like lip-sync drift.
  • Acceptance. ST 2059-2 targets sync within 1 µs across the facility; we verify with an analyzer rather than trusting switch dashboards.

Bandwidth math and sizing table

For ST 2110-20 the active payload is active pixels × bits per pixel × frame rate. 10-bit 4:2:2 averages 20 bits per pixel. For 1080p at 60/1.001 Hz: 1920 × 1080 × 20 × 59.94 ≈ 2.49 Gb/s. Packetization adds RTP, UDP, IP, and Ethernet overhead, which is why planning figures are usually quoted around 2.6 Gb/s per stream. For 2160p59.94: 3840 × 2160 × 20 × 59.94 ≈ 9.94 Gb/s payload, about 10.5 Gb/s on the wire.

Blanking is not carried, so a 2110-20 stream is smaller than the equivalent SDI wire rate. That is why a 3G-SDI signal (2.97 Gb/s) is smaller as 2110 video than its link rate suggests, and why 4K needs at least 25GbE at the edge. Wire-rate figures below are planning approximations (about 4–5% overhead) and should be confirmed against your endpoints' actual packet sizes.

Planning figures, 10-bit 4:2:2 (payload computed; wire rate approximate)
FormatST 2110-20 payloadApprox. wire rateLegacy SDI linkMin. edge portStreams per 100GbE at ~90% fill
720p59.941.11 Gb/s~1.15 Gb/sHD-SDI 1.485G10GbE~78
1080i59.941.24 Gb/s~1.3 Gb/sHD-SDI 1.485G10GbE~69
1080p59.942.49 Gb/s~2.6 Gb/s3G-SDI 2.97G10GbE (25GbE for headroom)~34
2160p59.949.94 Gb/s~10.5 Gb/s12G-SDI 11.88G25GbE~8
PCM audio, 8 ch, 24-bit, 48 kHz (ST 2110-30)9.2 Mb/s~10 Mb/sembedded in SDIanynegligible

Audio and ancillary streams are small, but they are numerous and timing-sensitive, so they drive multicast group counts and switch table sizes more than bandwidth.

ST 2110-22 and JPEG XS

ST 2110-22 carries constant-bitrate compressed video within the 2110 framework, and JPEG XS is the codec most commonly used with it. It is built for very low latency, a fraction of a frame, which is why it fits live production where H.264/HEVC does not. Typical uses are UHD over 10GbE links, and campus or WAN links between facilities. At a 5:1 ratio, UHD drops from about 9.9 Gb/s to roughly 2 Gb/s (arithmetic at the stated ratio; actual bitrates are set by the encoder and quality target). Both ends need compatible encoders and decoders, and you should confirm the generation-loss behavior your workflow needs with your vendors before putting several encode stages in series.

Control: NMOS IS-04, IS-05, and security

AMWA NMOS is what makes a multi-vendor 2110 plant operable. IS-04 handles discovery and registration: nodes announce their senders and receivers to a registry. IS-05 handles connection management on both ends: the controller stages transport parameters on the sender, passes the sender's SDP to the receiver's staged endpoint, and activates immediately or at a scheduled time. IS-08 maps audio channels inside endpoints, which avoids a multicast group per mono channel. Treat NMOS support as a procurement requirement and test each device's implementation, because "NMOS-compatible" varies between products. NMOS also defines security best practice (BCP-003: TLS for the APIs and authorization), and since these APIs patch your plant they should not sit on a flat, open network.

Migration phasing and gateways

Few facilities can shut down for a forklift upgrade, so we phase the work:

  1. Core and gateways. Stand up the IP fabric, timing, and NMOS/controller layer, and connect existing SDI through gateways so operators can route between baseband and IP.
  2. IP edge. Move multiviewers, switchers, replay, and processing to native IP interfaces, removing gateways from the primary path.
  3. Native studio floors. Convert cameras, audio, graphics, and intercom as they reach end-of-life, and extend to remote sites and cloud.

Gateways are where ancillary data gets lost, so test captions, timecode, and SCTE markers across them early. Our case studies show the range of this work. On the CVS / Aetna broadcast facility, we designed a Ross FR12 build in which one 12RU frame runs two independent control rooms, with a pathway to HDR and 4K. On the UBS Broadcast Center, the New York studio is headless, with only control surfaces, cameras, and mics on site and processing driven from Broadcast Management Group's Cloud Control Center. Not every IP build is ST 2110: the Washington Commanders facility is vMix-based with LiveU contribution and BetMGM runs a clustered vMix cloud environment. PTP and ST 2022-7 do not apply to those compressed, software-buffered workflows, but the redundancy and commissioning discipline below applies to all of them.

Commissioning checklist

  1. Optics and cabling. Inspect and clean every connector, and check optical power against each transceiver's datasheet.
  2. Timing. Soak-test PTP across the fabric with an analyzer, then force a grandmaster failover.
  3. Sender conformance. Check each video sender's ST 2110-21 sender type (N narrow, NL narrow linear, or W wide) against the receiver and switch buffers.
  4. Redundancy. Pull Red, confirm clean output, restore it, then pull Blue. Do this on live-quality program paths, not just test signals.
  5. Control. Confirm every node registers in the NMOS registry, then exercise IS-05 routing in volume from the production controller.
  6. Ancillary and captions. Push captions, timecode, and SCTE markers through gateways in both directions.
  7. Documentation. Deliver signal-flow drawings, IP plans, switch configs, and the test results, so the facility can operate and extend the plant without us.

Common failure modes

  • FEC mismatch. A 25G/100G link comes up but drops packets because the two ends use different FEC modes. Set it explicitly on both sides.
  • No IGMP querier. Streams work, then stop after a few minutes once snooping state ages out.
  • ECMP hash collisions. Drops while uplinks average well below capacity. Fix with flow-aware placement or orchestration.
  • Asymmetric PTP paths. Constant offsets and lip-sync complaints. Look for non-PTP-aware devices, firewalls, or media converters.
  • Wide-pacing senders on shallow buffers. Type W senders, common in software and GPU-based products, release packets in bursts that exceed what narrow receivers and shallow switch buffers can absorb, so egress queues drop packets despite free link bandwidth.
  • Audio packet time mismatches. ST 2110-30 conformance levels differ: Level A is 1 ms packets with up to 8 channels, while Levels B and C add 125 µs packet time (C up to 64 channels). A 125 µs-only console and a 1 ms-only device will not talk. Confirm packet time and channel count per stream at design time.
  • Red/Blue cross-connection. One shared trunk or management bridge between fabrics, and a loop on one fabric takes out both.

Frequently asked questions

Can uncompressed 2160p59.94 run over 10GbE?
No. The ST 2110-20 active payload for 2160p59.94 10-bit 4:2:2 is about 9.94 Gb/s before RTP/UDP/IP/Ethernet overhead, so it exceeds a 10GbE link. Use 25GbE or faster for UHD endpoints, or compress with ST 2110-22 (JPEG XS).
Do we need Red/Blue (ST 2022-7) redundancy?
For live and on-air paths, yes: two physically separate fabrics with no shared chassis or inter-switch links, and endpoints that send and receive both streams. Non-critical edit or monitoring paths can run single-network to save cost.
Boundary clock or transparent clock?
Both handle switch queuing: a transparent clock corrects for residence time in the correctionField. We default to boundary clocks in spine-leaf media fabrics because they scale: each switch serves Delay_Req traffic for its own ports instead of every endpoint talking to the grandmaster, and each hop isolates its downstream timing domain. Whatever you choose, validate it with a PTP analyzer on the finished network and follow your switch and endpoint vendors' tested profiles.
Multicast/IGMP or SDN orchestration?
Small fabrics can run IGMPv3 and PIM with careful design. Larger or tightly packed fabrics benefit from SDN or broadcast-controller bandwidth reservation, because multicast alone does not know whether a chosen path has headroom.
Do we have to rip out SDI on day one?
No. Most migrations are phased: IP core and gateways first, then IP edge devices, then native IP studio floors. SDI islands remain connected through gateways until their end-of-life.

Planning an SDI-to-IP migration?

We design, phase, and commission ST 2110 and hybrid facilities in the DC, Maryland, and Northern Virginia area and nationally. See broadcast systems integration and broadcast integration in the DMV.

Request a consultation