What a timing error costs
Here is the arithmetic. An IMU turning at 100 degrees per second with a 1 ms offset against the lidar is 0.1 degree out, which puts a point 50 m away about 8.7 cm from where it should be. A vehicle moving at 20 m/s with a 10 ms offset between the camera and the position fix is 20 cm out. Those errors don't look like noise. They look like consistent motion, so the estimator believes them, and no amount of filtering removes them because the filter trusts the timestamps.
Where the stamp is applied
There are two places to stamp a measurement. A software stamp is taken when the driver hands the packet to your application, so it carries every delay on the way: USB or Ethernet buffering, interrupt latency and scheduling. A hardware stamp is taken when the sensor strobes an exposure or sync pulse, by a counter that runs from the same clock as everything else. One recent paper on robot sensor synchronisation calls timestamping "a key issue that rarely receives the attention it deserves". In its tests, a non-real-time Linux computer sat within single-digit microseconds of a hardware pulse-per-second reference, and a microcontroller within 100 nanoseconds.

The pattern that follows is to discipline one hardware clock from GNSS and hand that time to everything else. Sensors that take a pulse get PPS plus time of day over NMEA. Sensors on Ethernet get IEEE 1588 or 802.1AS. Sensors that only emit a strobe get captured by a timestamping input on the same clock.
The Open TimeCard is a PCIe card built for that job. It carries a u-blox LEA-F9T multi-band GNSS receiver, hardware-timestamped IEEE 1588v2, frequency outputs from 1 Hz to 100 MHz over LVDS, and eight timestamping inputs for sensor strobes. The receiver is specified at 5 ns absolute timing accuracy. If your mission computer has a PCIe slot, that puts the sensor strobes, the network time and the GNSS reference on one clock.
When GNSS goes
Poland's telecoms institute reports GNSS disruption on about 70% of days in May and June 2026. A platform that takes time and position from the same signal loses both. The sensors keep reporting, the clocks keep running, and the offsets grow without any alarm.
The oscillator on the timing card is what carries you through. Typical-case figures from our holdover guide:
| Oscillator | After 1 hour | After 24 hours |
|---|---|---|
| OCXO | about 100 ns | 1 to 10 µs |
| DOCXO | about 30 ns | 300 ns to 3 µs |
| Rubidium | 5 to 20 ns | 100 to 500 ns |
These are typical-case figures, not worst case. The method in the guide is to pick the longest GNSS outage you need to ride through and the largest phase error your estimator tolerates, then check the published holdover at twice the outage with a margin of two. For a 12 hour outage and a 1 µs tolerance, that means reading the 24 hour column and doubling it. Rubidium is the only class that stays inside 1 µs across its whole range, a DOCXO sits on the line at its best and well over it at its worst, and an OCXO is out. The Open TimeCard datasheet lists 1.5 µs over 24 hours for the DOCXO and about 500 ns for Rubidium Black.
When a source disagrees instead of disappearing
A failover chain switches when a source fails. A source that is being jammed or spoofed rarely fails cleanly. It starts to disagree, and the chain keeps trusting it until it stops.
Clock Ensemble runs inside the Timebeat Agent and treats every input as part of one weighted ensemble: GNSS, PTP, PPS, NMEA, PTP hardware clocks, White Rabbit and the local oscillator. Each cycle it computes a weight for every input from that input's running Allan deviation at a configurable tau, and normalises the weights to sum to one. In the example the Clock Ensemble article walks through, a GNSS, PTP and oscillator split of 0.7, 0.2 and 0.1 becomes 0.0, 0.85 and 0.15 when GNSS degrades. A transient outlier is rejected for that cycle and held at zero weight for the next several. A sustained shift keeps the input down-weighted until its distribution comes back, and it re-enters over minutes instead of in one step. Sync Insight shows the active sources and their current weights, so you can see why the output moved.
Try it on the bench before you fly it
If you want to run this against your own sensor stack before it goes near a vehicle, the Open Time Appliance works standalone on a bench. You can order it with a DOCXO or a Rubidium oscillator, which lets you test the holdover cases in the table above with the oscillator you would actually field. The datasheet lists 1.5 µs over 24 hours for the quartz OCXO and under 500 ns for Rubidium Black. When the timing source has to sit out at the edge with the platform, the Open Time Appliance Mini is the compact version built for edge deployment. It has 1 PPS and 10 MHz outputs, two further programmable frequency outputs, and an optional battery. It ships with an OCXO good for 1.5 µs over 8 hours, and a DOCXO upgrade gives 1.5 µs over 24 hours. If your test range spans several fixed sites that need to share time over single-mode fibre, White Rabbit is the sub-nanosecond option.
Three questions for your design review
- 01Which clock do your sensor stamps trace back to?
- 02How long can that clock hold if GNSS is jammed, and how much error can your estimator tolerate at the end?
- 03What does it do when a source starts to disagree instead of disappearing?
The answers differ between a ground vehicle and a small aircraft, because mass, power and temperature limits differ, which is why they belong in a design review early.
What to do next
- ●Size your holdover: read the holdover guide and run your own outage and tolerance through it.
- ●See the hardware: Open TimeCard for the vehicle computer, Open Time Appliance for the bench, Open Time Appliance Mini for the edge.
- ●Bring your block diagram to a call and we will walk through your three answers with you: book a call.
Related reading
Blog · Automotive
Why Timing Is Critical in Autonomous Vehicles
Autonomous vehicles depend on tight clock synchronisation across LiDAR, cameras, radar and the central compute platform. Why timing accuracy is non-negotiable in AV systems and where the failure modes live.
Blog · Hardware
Anti-Jam GNSS At The Hardware Layer: L1/L2/L5 and OSNMA On Every Shelf Unit
Why single-band L1 receivers are obsolete for critical timing. What the u-blox LEA-F9T multi-band receiver gets you that an incumbent single-band unit does not, and how OSNMA at the platform layer cross-validates the receiver's own authentication chain from outside the receiver.
Blog · Observability
Allan Deviation As Operational Observability: Detecting Oscillator Aging Before It Breaches
Oscillator aging is a slow, predictable, mostly-invisible failure mode. Allan deviation is the statistic that makes it visible weeks before it becomes a problem. Why trend-based alerting matters more than threshold-based alerting, and how to wire up a Grafana panel that surfaces the curve.

