2. Instrumented channel fundamentals

An instrumented channel is the physical and logical path that carries both timing information and control intent between a synchronization master and one or more high-speed cameras. Unlike a plain GPIO trigger line, an instrumented channel is designed so that every event can be attributed to a known time base, a known source, and a known propagation path. This chapter examines the four building blocks that determine whether a multi-camera acquisition is genuinely synchronized or merely "close enough": hardware timestamping, trigger distribution, clock domains, and synchronization signal encoding.

Hardware Timestamping: Latency You Can Measure

Software timestamps are taken when an interrupt or polling loop observes an event, so they inherit operating-system scheduling jitter — typically tens of microseconds to several milliseconds. Hardware timestamping moves the capture of the time value into the FPGA, PHY, or camera sensor interface, where the event is latched against a local counter with deterministic latency. The result is a timestamp whose uncertainty is bounded by the counter resolution and the fixed pipeline delay, not by software load.

Two quantities matter in practice. The first is resolution: the period of the timestamp counter. A 200 MHz counter gives 5 ns granularity. The second is accuracy: the offset between the timestamp value and the true instant of the physical event, including sensor exposure start, PHY receive latency, and pipeline delays. Resolution is easy to read from a datasheet; accuracy must be characterized. A useful technique is to feed a common electrical pulse into the trigger input of two cameras and compare their reported timestamps. The residual difference, after subtracting known cable delays, is the combined timestamping error of the pair.

Trigger Distribution: Fan-Out Without Skew

Trigger distribution is the problem of delivering one event to N cameras with minimal and, more importantly, matched propagation delay. Three topologies dominate:

• Daisy-chain — simple cabling, but each hop adds delay, so downstream cameras are systematically late. Acceptable only when hop delay is characterized and compensated.
• Star / fan-out — a single driver feeds all cameras through equal-length paths. Skew is dominated by driver output-to-output matching and cable-length tolerance.
• Encoded broadcast — the trigger is embedded in a continuous serial stream (for example, a sync-on-wire protocol or PTP-based event message), so all receivers decode the same event identifier rather than relying on edge arrival time.

For a star topology, skew budget is straightforward to estimate. If the fan-out buffer specifies 150 ps output-to-output skew and the cables are cut to ±2 cm, the cable contribution is roughly 2 cm ÷ 5 ns/m ≈ 100 ps. The electrical trigger skew is therefore on the order of 250 ps — negligible compared with a 10 µs exposure. The dominant error usually comes from the camera's internal trigger-to-exposure latency, which can vary by microseconds between models and must be compensated in software or by per-camera delay settings.

Clock Domains: Why "Same Frequency" Is Not Enough

Every camera, the trigger generator, and the host interface run in their own clock domain. Even nominally identical oscillators differ by tens of ppm, and each domain has its own phase. Two consequences follow. First, frequency offset causes timestamps to drift apart linearly: a 20 ppm offset accumulates 20 µs of error per second, or 1.2 ms per minute. Second, phase offset causes a constant bias that must be measured, not assumed.

The standard remedy is to discipline all local counters to a common reference. PTP (IEEE 1588) or a dedicated sync signal can steer each camera's counter, but the steering loop has a bandwidth and a settling time. During the first seconds after lock, residual offset may be large; acquisitions started too early will show skew that later disappears. A practical rule is to allow the synchronization loop to settle for at least ten time constants before trusting timestamps, and to log the estimated offset so that post-processing can flag marginal captures.

Synchronization Signal Encoding

The encoding of the sync signal determines what information each camera can recover. A simple level trigger conveys only "an event happened now," leaving the camera to infer identity and time from its own counter. A richer encoding embeds an event identifier, a sequence number, or an absolute time value, so that a missed or duplicated trigger can be detected rather than silently corrupting the dataset.

Common encodings include:

• Edge / level trigger — minimal latency, no event identity; suitable for free-running synchronized acquisition where every frame is expected.
• Pulse-width or pulse-position coding — a small alphabet of commands (start, stop, marker) distinguished by pulse width; robust but low bandwidth.
• Serial event messages — event ID and timestamp carried in a framed packet; higher latency but self-describing and loss-detectable.
• Continuous sync waveform — a periodic reference (for example, a 10 MHz or 1 PPS signal) used for clock disciplining rather than event delivery.

In a typical instrumented setup, the continuous sync waveform disciplines the clocks while a separate encoded trigger channel delivers events. Keeping these two functions on separate physical paths avoids the situation where a burst of trigger traffic perturbs the very reference used to time it.

Putting the Pieces Together: A Worked Example

Consider a four-camera setup capturing a ballistic event at 10,000 fps. The sync master distributes a 10 MHz reference and an encoded trigger over separate paths. Each camera timestamps the trigger-received event against its disciplined counter. Suppose the vendor specifies a trigger-to-exposure latency of 12 µs for cameras 1–3 and 18 µs for camera 4, which is a different model. Without compensation, camera 4's frames are 6 µs late relative to the others — at 10,000 fps, that is 6% of a frame interval, enough to matter for trajectory reconstruction.

The compensation procedure is concrete. First, measure the residual timestamp offset between cameras using a common electrical pulse injected into all trigger inputs; record the mean and standard deviation over at least 100 pulses. Second, subtract the known per-model latency from each camera's reported timestamp. Third, verify that the corrected offsets fall within the required skew budget — for this example, say ±2 µs. If a camera's offset drifts during the capture, the clock disciplining loop has not settled or the reference path is marginal; check the PTP offset log and the reference signal integrity before re-running.

Design Checklist for the Instrumented Channel

• Define the time reference point explicitly: exposure start, frame-valid, or trigger-received.
• Separate the continuous clock reference from the event trigger path.
• Characterize fan-out skew and cable-length tolerance; document the budget.
• Measure per-camera trigger-to-exposure latency and apply compensation.
• Allow the disciplining loop to settle; log residual offset per capture.
• Choose an encoding that makes missed or duplicated events detectable.

The instrumented channel is only as trustworthy as its weakest characterized element. Hardware timestamping provides the measurement, trigger distribution provides the event, clock domains determine the drift, and signal encoding determines whether errors are visible. Treat each as a separately budgeted and separately verified contributor, and the resulting synchronization will be defensible rather than assumed.