The crystal is not boring

The most important component in a TESSERA costs fifteen cents and sits in a metal can smaller than a pencil eraser. The note we keep on crystal selection runs twenty pages. This lab note is the short version, written because the question behind it keeps coming up in the community: why does a digital module's timing depend on a piece of rock?

TESSERA's buffer runs at 48 kHz from a crystal oscillator, and every timing decision in the module, from the grain window to the latency figure on the spec sheet, is a division of that clock. Get the crystal wrong and the module still works. Get it wrong in the subtle way and the module works today and drifts next year, which is worse, because a fault you can hear is a fault you can chase.

What we measure before a crystal gets a part number

The twenty-page note exists because a crystal's datasheet tells you what it does at 25 degrees Celsius in the first 48 hours, and the questions that matter are all about time: how the frequency ages over years, how it moves across the temperature range, and how it responds to the load capacitance the surrounding circuit actually presents, which is never exactly what the schematic says it is.

• Aging: we accept under 3 ppm in the first year and under 1 ppm per year after, measured on the 200-unit sample lot, not taken from the datasheet.
• Cut angle: the AT-cut family only, chosen for the temperature curve's shape, not its middle value. The inflection point matters more than the nominal frequency.
• Load capacitance: we measure the board's real capacitance with the rig, then specify the crystal to the measurement. The datasheet's assumption gets checked, never trusted.
• Drive level: overdriving a crystal ages it faster, so the oscillator circuit runs the crystal at about 60% of its rated drive, trading a little margin for a lot of years.

The 2025 shortage lesson

In 2025 the frequency-control market went tight, and the crystal specified for TESSERA's first revision went to a 26-week lead time. We did two things, and both are now policy. First, we qualified a second source in a different country with a different aging profile, and published both in the hardware documentation, so a TESSERA built after the change carries the note. Second, we bought enough of the original part to cover four batches, which is inventory we would rather not hold and insurance we were glad to have.

The community asked, correctly, whether the two sources sound different. The answer we measured is: not within any test we can devise, including the 9-engineer blind panel from the drift note. The answer we publish is the same, with the measurement attached, because "they are identical" without a measurement is a marketing sentence and this is a hardware company.