Public data on precision timing

Why holdover is hard to measure honestly

Vendor datasheets routinely publish holdover performance figures — “1 µs / 24 h with OCXO,” “100 ns / 24 h with rubidium” — but rarely publish the methodology behind those numbers. The figures are typically derived under controlled laboratory conditions: stable temperature, recently disciplined oscillator, clean GNSS reference until the moment of test, no environmental perturbation. Real-world holdover rarely happens under those conditions.

The result is a published-figure problem: vendor datasheets are not lying, exactly, but they describe a best-case scenario that operators cannot replicate or rely on. This methodology paper exists to define a test procedure that operators can run in their own labs, against any grandmaster, with results that are directly comparable across vendors.

Test setup

The reference test setup uses a hardware-timestamping reference clock as the ground truth (we use a TimeBeat Open Time Appliance with rubidium holdover, externally GNSS-disciplined and continuously monitored). The grandmaster under test is locked to a separate, independently disciplined GNSS source until the moment of the test, then the GNSS antenna is physically disconnected and the local oscillator is left to free-run.

Time-of-day measurements are captured every 100 ms by comparing the grandmaster’s PTP output against the ground-truth reference using a cross-correlation timestamping technique that resolves to better than ±10 ns. The temperature of the grandmaster enclosure is held to ±2 °C for the duration of the test using a controlled environmental chamber. Tests run for 72 hours minimum, with a 24-hour stabilisation period before the holdover event.

Reference data from open literature

Pending publication of TimeBeat’s own measurements, the table below summarises holdover performance figures published in vendor datasheets and peer-reviewed literature for the four oscillator classes most commonly deployed in PTP grandmasters. These are typical-case figures under controlled conditions; real-world holdover typically performs worse, particularly during the first 30 minutes of free-run.

What we will measure (forthcoming)

TimeBeat’s own laboratory holdover programme is currently underway, running the methodology described above against the Open TimeCard, Open Time Appliance and Open Time Node WR product lines. We expect to publish the first set of measured data in mid-2026, including raw measurement files in Parquet format, the analysis scripts (Python notebooks) used to generate the published charts, and a detailed accuracy assessment of the ground-truth reference clock.

Until then, the methodology itself is publishable, citable and replicable. If you run this procedure on a grandmaster of your own and want to share results, the email address at the bottom of this page reaches the engineering team directly.