Summary

  • RFC 9584 defines single-NAL packets, aggregation packets and fragmentation units for carrying EVC over RTP. These structures can establish ordered decoder input; they cannot establish that a picture was decoded or shown.
  • Sequence numbers, the marker bit, RTP timestamps, DON values, PLI and FIR each answer a bounded question. End-to-end playback needs separate receipts from packet ingress through renderer commit.

The network dashboard is green. Its trace shows an unbroken run of RTP sequence numbers, a marker bit at the end of the access unit and timestamps advancing at 90 kHz. Yet the display is black. There is no contradiction. The dashboard proved that one observer saw a well-shaped transport stream. It did not prove that the decoder had the right parameter sets, accepted the bitstream, produced a picture before its deadline or handed that picture to a visible surface.

RFC 9584 is useful precisely because it is narrower. Published on the IETF Standards Track in June 2024, it defines how Essential Video Coding NAL units travel in RTP. A packet may carry one NAL unit. An aggregation packet may carry two or more NAL units from one access unit. A fragmentation unit may carry part of one NAL unit when the whole object does not fit comfortably into a packet. The format inherits a familiar design lineage from H.264, HEVC and VVC payload formats. That lineage reduces novelty at the transport boundary; it does not eliminate failures beyond it.

The RTP header supplies several distinct observations. A sequence number helps reconstruct the sender’s packet order. It does not guarantee delivery, and a monitoring tap can miss the same packet as the receiver or see one the receiver drops later. The marker bit identifies the last packet of an access unit carried in the stream. It does not attest that all earlier packets arrived. The timestamp supplies the sampling and display-time reference, but it is not a receipt that the frame met its deadline. SSRC identifies a source within the RTP model; absent an appropriate security mechanism, it is not proof of a trusted producer.

Fragmentation makes the boundary sharper. The start and end bits identify the outer edges of a fragmented NAL unit. The receiver reconstructs the NAL payload by concatenating consecutive fragments from start to end. If a fragment is lost, RFC 9584 normally directs the receiver to discard the following fragments for that NAL unit unless the decoder is known to tolerate incomplete input. An end fragment therefore cannot stand in for the missing middle. “Reassembly ended” and “the complete object was reconstructed” are different claims.

Aggregation has a different risk. An aggregation packet contains multiple NAL units from the same access unit in decoding order; it cannot contain fragmentation units and cannot nest another aggregation packet. The packet can arrive intact while one contained NAL is malformed, refers to unavailable state or uses a tool the receiver cannot process. Transport integrity is necessary evidence for decoder submission. It is not a decoder verdict.

The decoding-order number is also easy to overread. DONL and DOND contribute to an absolute decoding-order value used to reorder NAL units before they are passed to the decoder. Gaps can be legitimate because a gateway may omit higher temporal layers or selected SEI units. Equal values may permit either order. AbsDon describes where input belongs in the codec pipeline; it says nothing about whether reference pictures exist, whether the decoder has enough memory, whether a security check passes or whether a frame leaves the decoded-picture buffer.

RFC 9584 separates the de-packetization buffer from practical jitter buffering. One reconstructs and reorders codec units; the other absorbs variable network delay. A receiver may have enough space for the advertised de-packetization requirement and still miss presentation deadlines because the jitter budget, decoder load or rendering path is insufficient. A single “buffer healthy” metric collapses two control surfaces that fail differently.

Recovery feedback is bounded too. A Picture Loss Indication says that an undefined amount of coded picture data was lost. A Full Intra Request requires the sender to produce an IDR picture and, unless already established out of band, the parameter sets. The feedback receipt proves a request entered the control loop. The IDR packet proves a new recovery object was sent. Neither proves the receiver decoded it or that the viewer saw recovery.

The specification states the larger limit plainly: the payload format in isolation is not a complete system. RTP itself does not reserve resources or guarantee quality of service. Confidentiality, integrity and source authenticity require choices at the application and RTP-security layers. The video/EVC registration proves code-point coordination, not implementation, negotiation or operating success.

An accountable evidence chain therefore keeps the joins visible: packets observed at the intended endpoint; source and integrity verified; fragments complete; NAL units reconstructed; decoding order resolved; parameter sets and profile accepted; picture decoded; presentation deadline met; renderer committed; display surface updated. Leadership can decide which stage authorizes which action. It cannot honestly rename the first stage as the last.

Sources