Reading a Crystal Datasheet: Three Tolerances, Two Ways to Add Them

ppm is a fraction, and that is the whole trick

One part per million is $10^{-6}$. So the absolute error scales with whatever you multiply it by:

At 32.768 kHz, ±20 ppm is ±0.66 Hz. At 25 MHz the same ±20 ppm is ±500 Hz. Synthesise a 2.4 GHz carrier from that reference and it becomes ±48 kHz — enough to fail a channel accuracy requirement.

The shortcut worth memorising: frequency in MHz equals Hz per ppm. A 25 MHz part moves 25 Hz for every ppm.

Three terms, not one

The number on the front of the datasheet is the initial or calibration tolerance — trim accuracy at 25 °C on the day it shipped. Two more terms sit behind it:

Temperature stability. For an AT-cut crystal, frequency versus temperature is a cubic with an inflection near 25 °C. It is quoted as a total excursion over the operating range rather than a coefficient, because the curve is not linear. Aging. Irreversible drift from stress relief in the mount and mass transfer at the electrodes. Quoted per year, largest in year one.

A typical 25 MHz part: ±20 ppm initial, ±15 ppm over temperature, ±3 ppm/year. Over three years the aging term accumulates to ±9 ppm.

Sum or root-sum-square?

This is the decision that matters, and both answers are correct for different questions.

Arithmetic sum: $20 + 15 + 9 = 44$ ppm. This is the guaranteed limit. No unit, in any condition, at any point in its life, exceeds it. It is what a protocol margin, a regulatory frequency-accuracy requirement, or a worst-case link budget has to survive. Root-sum-square: $\sqrt{20^2 + 15^2 + 9^2} = 26.57$ ppm. This assumes the three mechanisms are uncorrelated, which they largely are — a trim offset tells you nothing about where the turnover temperature landed. It describes the realistic production spread and is what a yield calculation should use.

The RSS figure is 40 % smaller. Using it where a guaranteed limit was required is how a design passes qualification and then fails intermittently in the field, on the units that happened to stack.

At 25 MHz, the ±44 ppm worst case is ±1100 Hz, putting the part somewhere between 24,998,900 and 25,001,100 Hz.

The timekeeping conversion

Because ppm is a fractional rate, it maps straight onto elapsed time with no extra physics:

• 1 ppm = 1 µs per second
• 1 ppm = 86.4 ms per day
• 1 ppm = 31.5 s per year

So ±44 ppm is up to 3.8 seconds of drift per day and 23 minutes per year.

This conversion is worth doing early, because seconds per day is a number product managers can judge and ppm is not. A ±20 ppm RTC drifts ±1.73 s/day — about ten and a half minutes a year. If the requirement is "within a minute a year unaided," you need roughly ±2 ppm, which means a TCXO, not a crystal.

The term that is not on the datasheet

Load capacitance mismatch.

A crystal is specified to hit its marked frequency against a particular load capacitance — commonly 12 pF, 18 pF, or 20 pF. Present it with a different load and it is pulled off frequency. The pulling is often tens of ppm, larger than the initial tolerance and temperature stability combined.

And it appears in none of the three datasheet terms, because it is a property of your circuit, not the part.

The load the crystal actually sees is the two loading capacitors in series, plus stray capacitance from the pads, traces, and oscillator pins — typically 3 to 5 pF that is easy to forget. Sizing the loading capacitors from the specified $C_L$ without subtracting stray is the single most common reason a board comes back running off frequency.

Aging over long service life

Accumulating a per-year aging figure linearly is deliberately pessimistic. Real crystal aging is roughly logarithmic — most of the shift happens in year one and the rate falls afterwards. Ten years is typically 2 to 3 times the first-year figure, not 10 times.

Linear accumulation is the conservative reading of the specification, and it is the right choice for a guaranteed limit. But if it is driving you toward a much more expensive part for a ten-year product, it is worth asking the vendor for their actual aging curve rather than designing to the pessimistic envelope.

Guidance

Size against the tightest consumer of the reference. One design can feed a UART that tolerates thousands of ppm and a radio synthesiser that tolerates ten. Two independent clocks means two budgets. When two devices must stay in step, the relative error is the sum of both — two ±30 ppm parts can differ by 60 ppm. If temperature dominates, buy a TCXO rather than tightening the trim. Initial tolerance is expensive to improve and temperature is usually the largest term in a wide-range design. Fix the load capacitance first. It is a layout and component choice, and getting it wrong swamps everything the datasheet promised.

The crystal PPM tolerance calculator stacks all three terms both ways, converts to Hz at your carrier, and reports drift in seconds per day and minutes per year.