Why ST 2110 Determinism Is a Network Architecture Problem, Not a PTP-vs-TSN Choice

On paper the core was textbook: a leaf-spine fabric, 25 GbE to the endpoints and 100 GbE on the uplinks, average utilization held well below capacity, dedicated media VLANs, dual PTP grandmasters, and merchant-silicon switches whose datasheets all listed the TSN feature set.

Commissioning passed. Individual streams were clean. PTP locked to well under a microsecond. Every switch dashboard was green.

Then the facility went to full channel count. With every camera live and audio on the same fabric, receivers on two leaves began reporting intermittent artifacts, and one leaf logged egress drops on a handful of multicast groups. None of it reproduced on demand. The grandmasters never failed, and the links were nowhere near saturated on average.

The team had built a fast, well-provisioned IP fabric. What they had not designed was a deterministic media fabric — and the difference shows up only under the exact traffic a demo never produces.

Wrong Assumption

The assumption is reasonable on its face: enough bandwidth, a locked PTP clock, and TSN-capable switches add up to deterministic behavior in production. It skips the part of ST 2110 that actually causes the failures. PTP aligns time; it does not constrain queueing delay. Overprovisioning lowers the probability of sustained congestion but does not remove burst-induced jitter. And a “TSN-capable” switch is a set of datasheet features, not a configured schedule. Until the multicast design, the buffer behavior, the PTP domain, the ST 2110-21 pacing model, and — where used — the gate schedules are engineered against the traffic the facility will really carry, the fabric behaves as best-effort under the hardest load.

Why It Fails

PTP gives you time, not delivery. ST 2110 replaces SDI timing with IEEE 1588 Precision Time Protocol under the SMPTE ST 2059-2 profile, and it does that job well: sub-microsecond alignment, coherent RTP timestamps, lip-sync across essences, and clean switching points. But PTP says nothing about when a packet leaves a switch queue. If a leaf buffers and forwards under contention, arrival variation grows and latency climbs while PTP stays locked. A healthy timing plane can sit on top of an eroding delivery plane.

Statistical determinism is defeated by synchronized microbursts. Most ST 2110 cores rely on statistical determinism: oversized backplanes, substantial bandwidth headroom, media VLANs, and QoS. It works on average. But depending on the ST 2110-21 sender type and packet read schedule — narrow, narrow-linear, or wide — uncompressed video traffic can be line-related or otherwise bursty, which is exactly why the ST 2110-21 traffic-shaping model exists. When several sender patterns align at a leaf that then replicates to multiple egress ports, instantaneous egress demand can exceed a port’s buffer even while average utilization looks safe. Queue depth spikes and pacing is lost. Sender-side pacing and kernel bypass — the domain of low-latency IP transport — reduce host-generated packet variation, but they do not eliminate re-clumping caused later by switch buffering, replication, or egress contention.

Leaf-spine fabrics create multicast hot spots. In most facility-scale ST 2110 deployments, essence flows are commonly transported using multicast, although the standards also permit unicast delivery. Where multicast is used, behavior depends on IGMP snooping, querier placement, PIM tree construction, and where replication happens. ECMP distributes flows by hashing, and uneven hashing can concentrate replication load on some spine uplinks — a polarization effect that turns a specific routing state into a sporadic hot spot. Determinism at scale therefore depends as much on fabric and network switch design as on endpoint synchronization.

TSN is not a drop-in for broadcast. Time-Sensitive Networking can bound worst-case latency with the 802.1Qbv time-aware shaper and Gate Control Lists. Depending on the architecture, the TSN toolbox may also include 802.1AS time sync, 802.1Qci per-stream filtering and policing, 802.1Qbu/802.3br frame preemption, and 802.1CB frame replication and elimination — not every deployment uses all of them. Three realities complicate TSN in a broadcast core. First, 802.1AS (gPTP) is a different IEEE 1588 profile than SMPTE ST 2059-2, so a facility standardized on 2059-2 has to decide which domain the switches derive their scheduling time from, rather than assume the two interoperate. Second, much merchant silicon in broadcast cores implements Qbv only partially, or not at all, even when the datasheet lists TSN. Third, live production reroutes constantly: IEEE 802.1Qcc can support centralized stream configuration and resource management, but producing and safely updating coordinated Qbv schedules under dynamic live-production rerouting remains a system-level orchestration problem. Where TSN does fit, TSN development treats schedule generation and validation as design work, not a checkbox.

Convergence removes the isolation SDI gave you for free. Modern facilities converge ST 2110 video, AES67 or Dante audio, NMOS control, storage, and IT onto one fabric. PTP provides a common timebase to participating media endpoints, but it does not isolate media from control, storage, or general IT traffic sharing the same queues and links. Without enforced class boundaries, a burst of IT traffic or a control-plane storm can delay media packets on a shared egress queue.

In production these rarely arrive one at a time, and stacked together they produce a state that no single-stream test surfaces: every device reports healthy while the output is visibly impaired.

Hidden System Complexity

sender packetization (ST 2110-21 pacing) → PTP timestamp (ST 2059-2) → leaf ingress → QoS + buffer → multicast replication → ECMP spine → egress buffer → receiver buffer (Cmax/Vrx) → de-packetization → reconstruction against PTP → SDI/IP output

The network is the medium every other stage talks through, so a symptom at the output usually originates several hops upstream. A Qbv gate-control error, or a buffer sized for demo traffic, does not fail loudly: it lets media packets slip behind other bursts at the moment a leaf is replicating a dense multicast set, and the visible symptom is a sporadic artifact on one receiver that reproduces only when that precise traffic combination lines up.

PTP recovery across a maintenance event is just as quiet. If a boundary clock re-locks slowly after a spine reload, reconstruction can run briefly against stale timing — and because audio uses far finer sample timing than video, it can drift audibly first, so the operator reports an audio fault while the cause sits upstream in the timing plane.

Failure Patterns

Scenario 1 — Works until full count. The core validates for days at partial channel count. At full count, with concurrent camera and audio traffic, one or two leaves log intermittent multicast egress drops and receivers show sporadic artifacts, while average utilization never approaches capacity. The cause is a synchronized microburst exceeding per-port egress buffer during replication — a condition that exists only at production density and was never in the acceptance traffic.

Scenario 2 — Timing drifts after a maintenance event. Everything is clean until a grandmaster handover or a spine reload. Afterward, timing recovers slightly off — a boundary clock re-locks with an offset, or a path-delay asymmetry biases the offset during re-convergence. PTP reads as “locked” on a single snapshot; the problem shows only in the offset time-series.

Scenario 3 — TSN schedule that protects nothing. A facility enables 802.1Qbv on TSN-capable leaves and builds Gate Control Lists from a theoretical per-stream model. Under real bursts the protected windows are mis-sized, so either media frames miss their gate (a bounded-latency violation) or best-effort traffic starves in dead time. In the worst version, the switches take scheduling time from an 802.1AS domain while the media rides ST 2059-2, and the two profiles were never reconciled.

Deterministic Network Design for ST 2110 Facilities

Determinism in an ST 2110 core is engineered across the PTP domain, the multicast and buffer design, the ST 2110-21 pacing model, and, where justified, TSN scheduling. The failures that survive commissioning are microburst jitter, replication hot spots, PTP drift after failover, and TSN schedules built from theory. Promwad develops the hardware, FPGA, embedded software, and low-latency transport layers used in ST 2110 systems, and supports their integration into PTP-aware media networks — including PTP-aware endpoints, packet pacing, FPGA-based SDI-to-IP hardware, multicast and QoS integration, and realistic-load validation.

Explore ST 2110 / NMOS Integration →

Platforms and Silicon Ecosystems We Work With

Solution Approach

Step 1: Model the real multicast and burst traffic, not the spec. Instrument every class — video, audio, control, IT — on the target fabric under realistic load, and capture per-flow peak burst, replication fan-out, and the ST 2110-21 sender type (narrow, narrow-linear, or wide) for each source. Buffer sizing and any gate schedule are only ever as good as this model. This measured-pacing discipline is the foundation of low-latency IP transport.

Step 2: Engineer the PTP domain as a time-series, including failover. Confirm the ST 2059-2 profile and a consistent PTP domain across every device, place boundary clocks deliberately rather than by default, and measure offset and path-delay over time rather than as a single snapshot. Profile recovery after grandmaster handover and switch reloads, because that is where timing drift is born.

Step 3: Decide PTP-plus-headroom vs TSN on measured need, then engineer the fabric to match. Audit switch silicon for actual buffer architecture and real Qbv support, not datasheet claims. If overprovisioning suffices, fix buffers, multicast replication, and ECMP balance. If bounded latency is genuinely required, derive Gate Control Lists from the measured model, decide which timing domain the switches schedule against by reconciling 802.1AS and ST 2059-2, and validate through network switch design and TSN development. The control-plane, discovery, and rerouting side of the system — where NMOS state and re-routing live — is covered in the guide to NMOS IS-04/IS-05 for AV systems.

A gate schedule or buffer budget derived from a theoretical model and never checked against measured bursts will diverge from production traffic — and that gap is where the jitter and drops live.

Real Trade-Offs

PTP-plus-overprovisioning keeps configuration simple, uses a mature switch ecosystem, and stays flexible under dynamic reroutes, but it relies on statistical smoothing and stays exposed to microbursts, so it demands high-capacity silicon and disciplined multicast design.

PTP-plus-TSN scheduling bounds worst-case latency and isolates classes cleanly, but it adds hardware constraints, complex schedule orchestration, and reduced flexibility under live re-routing. In many broadcast facilities the operational simplicity of overprovisioning still outweighs TSN’s guarantees.

Deeper receiver buffers absorb jitter and tolerate imperfect pacing at the cost of end-to-end latency; tight buffers hit aggressive latency targets but leave no margin when a switch re-clumps packets.

More boundary clocks improve scalability and isolate PTP traffic, but each is another device that can recover slightly off and another link in the drift chain.

Single-vendor cores minimize interoperability seams and ship faster at the cost of lock-in; open multivendor ST 2110 and IPMX-based distribution buy freedom of choice and pay for it in integration and validation effort. Where broadcast and ProAV/IPMX environments diverge, and where each standard becomes fragile, is laid out in ST 2110 vs IPMX use cases.

At this point the work is network-architecture analysis, not more switch hardware: a measured multicast traffic model, corrected buffers and replication design, a PTP domain profiled across failover, and a deliberate PTP-vs-TSN decision validated under production composition. Where the same deterministic-Ethernet problem appears outside broadcast — industrial automation and automotive traffic on shared TSN fabrics — the underlying discipline is identical, and it is the subject of Promwad’s industrial network engineering. The control-plane and multivendor context that sits alongside the network layer is covered in the reciprocal ST 2110/NMOS field-failure analysis.

This class of problem shows up most in large ST 2110 cores and SDI-to-IP migrations where multicast burst modeling, buffer sizing, and PTP failover are treated as commissioning tasks instead of architecture-phase inputs, and in converged facilities where audio, control, and IT share the media fabric without enforced class boundaries.

Tell Us About Your ST 2110 Network