SDI Deep Dive: The Broadcast Wire, Word by Word

SMPTE 259M (1989) put studio digital video on the coax broadcast already owned: 75 Ω cable, BNC connectors, 800 mV peak-to-peak. ST 292 took the same idea to HD in 1998. The signal is scrambled NRZI: statistically enough transitions that the receiver recovers the clock from the data itself. That one property defines the culture: no handshake, no EDID, no HDCP, no retries. A source transmits; anything that locks, locks. That is why plants route SDI through matrices, patch bays, and distribution amps like water: every hop is passive-ish and deterministic, latency is fixed, and a signal is either present and correct or visibly broken. Compare the consumer world's two-way negotiation in the EDID module: SDI's answer to "what does the display support?" is "it will take what it's given."

Each line is bracketed by two four-word timing references: EAV (end of active video) and SAV (start). Both open with the only code values reserved on the whole interface, 3FF 000 000, followed by the XYZ word, which packs three flags: F (field 1/2), V (in vertical blanking), H (1 = EAV, 0 = SAV), plus four Hamming-protected parity bits so a single-bit hit can't fake a sync. During active picture lines a progressive signal's XYZ reads 274h at EAV and 200h at SAV; hover them in the diagram. After EAV come the line number (two words, LN0/LN1) and a two-word CRC computed over the previous line: a per-line, per-channel receipt. This is the wire-level reason codes 0–3 and 1020–1023 are forbidden to picture, the rule the Signal Range module teaches from the code-value side.

The blanking that analog TV spent on beam retrace, SDI spends on cargo. Ancillary packets open with their own flag (000 3FF 3FF, the mirror of the timing reference), then DID and SDID bytes naming the payload, a data count, the payload itself, and a checksum. Packets between EAV and SAV live in HANC; packets on vertical-blanking lines live in VANC. The residents are a who's-who of broadcast: embedded audio (ST 299, HANC), the payload ID (ST 352, VANC), timecode (ST 12 / RP 188), closed captions (ST 334 / CEA-708), AFD framing flags, SCTE 104 ad triggers, and HDR/WCG metadata (ST 2108). One 10-bit river carries all of it, in fixed positions, every frame: no side channel, no out-of-band anything.

ST 299 packs AES3 audio into HANC as audio data packets: four groups of four channels, sixteen 48 kHz, up-to-24-bit channels on one HD wire, organized as stereo pairs (group 1 = channels 1–4, group 2 = 5–8, and so on). Each group also sends an audio control packet carrying the exact clock-phase relationship between audio samples and the video line, which is how embedded audio stays sample-locked through a router the size of a wall. The catch is operational: every embed and de-embed is a format conversion. Groups can collide when two devices both try to write group 1; consoles that need audio must de-embed, process, and re-embed, adding a frame here and a frame there. The lip-sync drift that plagues plants is usually accumulated re-embedding, not "the cable".

A 1.485 Gb/s signal could be 1080i30, 1080p30, or 720p60: electrically identical. So SDI carries a passport of its own: the ST 352 payload identifier, four bytes in a VANC packet (DID 41h, SDID 01h). Byte 1 names the interface family; byte 2 carries scan flags and the picture-rate code; byte 3 the sampling structure; byte 4 the bit depth. Later revisions tucked colorimetry and transfer-characteristic flags into the spare bits, which is how a wire declares BT.2020 or HLG. It is SDI's entire answer to EDID, in the opposite direction: the source declares, once per frame, in-band. Receivers that honor it switch formats glitch-free; receivers that ignore it guess, and a wrong guess about 3G level A versus level B is the classic "my 1080p50 looks like two interlaced ghosts" service call.

3G came in two flavors that still bite: level A is 1080p50/60 mapped directly; level B carries two virtual 1.485 streams, the old dual-link arrangement folded onto one wire. Same connector, same bit rate, incompatible mappings; the payload ID (89h vs 8Ah) is the only honest tell. UHD repeated the story at the next octave: before 12G matured, 2160p60 went as quad-link 3G, either square division (each cable owns a quadrant; lose a cable, lose a corner) or two-sample interleave (each cable carries a full-picture subsample; lose a cable and the picture soft-degrades). 2SI won, and single-link 12G kept the 2SI arrangement internally: the four 1080 sub-images you see flagged in the diagram above. The ladder's genius is what never changed: word structure, timing references, ANC, audio. A 1989 design running at eight times the clock.

Coax loss grows with frequency, so every rung up the ladder shortens the leash: on premium RG-6 class cable, roughly 120 m at 1.5G shrinks toward 70–80 m at 3G and a few tens of meters at 12G. Exact figures depend on the cable datasheet, and serious plants read it. Receivers fight back with adaptive cable equalizers and reclockers that rebuild the signal completely, which is why SDI fails like a cliff, not a slope: perfect, perfect, perfect, gone. Test engineering matches the physics: pathological patterns (long runs of the scrambler's worst-case output) stress equalizer and clock recovery on purpose, and a waveform scope's eye pattern and jitter readouts show how much margin remains before the cliff. Past coax range, the same SDI rides fiber via SFP cages, unchanged at the protocol level.

Three checks solve most SDI mysteries. First, read the payload ID at the far end and compare it to what the source thinks it is sending: format-change leftovers and level-B surprises show up here, not on a monitor. Second, watch the CRC error counters: zero is the only acceptable number; occasional errors mean marginal cable, and marginal cable at 12G means a picture that dies on the warm afternoon. Third, audit audio group assignments before embedding anything new. On calibration jobs the same discipline applies before any probe touches the screen: if the wire is delivering 1080i when the chain believes 1080p, or level B into a level-A-only display, no amount of grayscale work will fix what the transport already broke. Book a calibration →