HDMI Deep Dive: TMDS, FRL, and the Metadata Bus

HDMI (2002) is DVI's TMDS video with everything the living room needed bolted on: audio in the blanking, Y′CbCr carriage, CEC control, a smaller connector, and, decisively for the studios licensing content, HDCP encryption. Seven consumer giants founded it, and the licensing model shaped the ecosystem: HDMI is a product you certify, not a document you download, the consortium pattern the Color Authorities module maps. Its inheritance runs through this whole series: the connector carries DVI's DDC wires, which is why EDID works unchanged; and its range, colorimetry and 4:2:0 flags are the exact switches whose failure modes fill the Signal Range and Y′CbCr modules.

Transition-Minimized Differential Signaling runs four shielded pairs: channels 0/1/2 carry blue/green/red bytes (plus syncs and auxiliary data), and the fourth pair carries the pixel clock itself, at exactly the video's rate. Every 8-bit value becomes a 10-bit character chosen to minimize transitions: 20% overhead, which is why 18 Gb/s raw yields 14.4 Gb/s of payload. The transport is a slave to the raster: the wire literally speeds up and slows down with the format, and blanking crosses the cable as real time waiting to be filled. Elegant at 165 MHz, desperate at 600: skew between four pairs at 6 Gb/s per channel is what finally killed the architecture. SDI, for comparison, embedded its clock in the data in 1989; this series' recurring lesson that broadcast solved consumer problems decades early.

HDMI's cargo hold is the data island period: guard-band-fenced stretches of blanking where the three data channels switch to a denser code (TERC4) and carry 32-pixel packets of audio samples, InfoFrames and housekeeping. Audio has no clock wire of its own; the sink reconstructs it from N/CTS packets that express the audio rate as a ratio of the video clock, the regeneration scheme whose small errors surface as home theater's lip-sync and dropout gremlins. The parallel with SDI's HANC is exact and worth savoring: same idea, same location, opposite culture. SDI's ANC is inspectable on any scope, while HDMI's islands are encrypted along with the video the moment HDCP engages.

Fixed Rate Link (HDMI 2.1) is a philosophical surrender to the packet world: the clock pair becomes a fourth data lane, every lane runs at a fixed trained rate (3 to 12 Gb/s per lane) regardless of the video format, and pixels arrive as packets to be buffered and re-timed; DisplayPort's architecture, adopted a decade late (module M23 gives DisplayPort its due). Characters are coded 16b/18b (≈11% overhead against TMDS's 20%), Reed–Solomon forward error correction is mandatory, and the link is trained through SCDC status registers riding the old DDC wires. FRL is also DSC's landlord: Display Stream Compression runs only over FRL, unlocking 8K60 RGB and 4K120 12-bit beyond the raw 42.7 Gb/s payload ceiling.

The rungs tell the 4K story in three acts. 1.4 (2009) reached 4K, but only at 30 fps. 2.0 (2013) doubled TMDS to 18 Gb/s for 4K60; at 8-bit RGB that consumes 14.256 of 14.4 Gb/s, a 1% margin, which is why 4K60 HDR on 2.0 hardware always costs something: 4:2:2, 4:2:0, or 8-bit banding. 2.1 (2017) quadrupled the budget with FRL. Then the fine print: the licensing body folded 2.0 into 2.1, so a port may legally say "2.1" while supporting only TMDS speeds; every 2.1 feature is optional per port. The working rule on any install: ignore the version number; read the EDID's declared max FRL rate and DSC bit, then verify the trained state in SCDC. The Bandwidth Bucket module prices every rung; this module is why the prices differ.

Every frame, the source re-declares itself in CTA-861 InfoFrames. The AVI InfoFrame carries the switches this series keeps returning to: pixel format, colorimetry (which decode matrix and primaries the TV assumes), quantization range, and the video code. The DRM InfoFrame announces HDR: an EOTF byte plus ST 2086 mastering metadata and MaxCLL/MaxFALL for the tone mapper. Vendor-specific InfoFrames carry Dolby Vision. The failure mode is always the same shape: pixels encoded one way, flags claiming another; limited content marked full, BT.2020 decoded as 709, PQ pixels under an SDR EOTF byte. The picture survives the trip; its meaning doesn't. When a client says "HDR looks gray" or "everything is washed out," the first suspect is not the panel: it is a stale or wrong InfoFrame somewhere in the chain, usually an AVR or extender regenerating them badly.

Downstream audio travels in the data islands, negotiated against the EDID's audio block (LPCM channel counts, AC-3, DTS, TrueHD flags; module M19 decodes them). The return path grew in two steps. ARC (HDMI 1.4) reused a utility pin at S/PDIF-class bandwidth: stereo PCM and lossy surround, nothing more; Atmos only in its lossy Dolby Digital Plus wrapping. eARC (HDMI 2.1) repurposed the old HEC ethernet pins into a channel of roughly 37 Mb/s: uncompressed 7.1 PCM and lossless TrueHD/DTS-HD MA bitstreams, with mandatory lip-sync correction and its own discovery handshake, so it works even when video runs at TMDS speeds. The practical audit: an "eARC" badge on the TV means nothing unless every switch and AVR in the return path passes the eARC data channel too.

HDMI problems are almost never the cable's electronics failing outright: they are negotiations quietly concluding on the wrong terms. The audit order that works: read the EDID the source actually received (an AVR or extender likely rewrote it; the "why EDIDs lie" card in module M19); confirm the link mode and rate (TMDS vs FRL, DSC on or off, via the source's or display's link status); then compare the AVI and DRM InfoFrames against the content: range bits, colorimetry, EOTF.

Only when the declarations match the pixels is the display's own behavior on trial, and only then does a probe on the screen tell you anything about the panel rather than the plumbing. That is the order a calibration follows too. Book a calibration →