11. Crystals and oscillators

A crystal does not produce a signal. Two capacitors decide what frequency it runs at, and adding more of them is a defect.

A short chapter about one component. No other part in the course shows as clearly that a part’s value can matter as much as its presence, and here doing more of the obviously-good thing makes it worse.

Read chapter 3 first for capacitors, and chapter 7 for ratings.

This page teaches at EE3 and EE5, and the levels page says what each level means.

A crystal is not an oscillator (EE3)

The first surprise is that a quartz crystal does not produce anything. Put one on a bench supply and nothing happens. It is a mechanical resonator, a sliver of quartz that flexes when you apply a voltage and generates a voltage when it flexes. It does both very efficiently at one particular frequency, and how the quartz was cut sets that frequency.

The oscillator is the chip. Inside is an amplifier, and the crystal is wired across it so that the amplifier’s output feeds back to its input. Noise gets amplified, the crystal passes back only the component at its own frequency, that gets amplified again, and within a few milliseconds the whole thing is ringing steadily at the crystal’s frequency.

The two capacitors are part of that feedback path. Each terminal gets one, to ground.

The oscillator loop. An amplifier inside the chip drives XIN and XOUT, a crystal bridges the two, and each terminal carries a load capacitor to ground. The crystal produces nothing on its own; the amplifier is the oscillator and the crystal only decides which frequency survives the feedback. The two capacitors sit in series as far as the crystal is concerned, so a 22 pF pair presents about 11 pF, plus whatever the pins and the copper add. ground the chip holds the amplifier Y1, the crystal XIN XOUT 22 pF 22 pF the two capacitors sit in series as far as the crystal is concerned, so a 22 pF pair presents about 11 pF a resonator with its capacitors built in already presents that load, and fitting these as well doubles it

What the capacitors decide (EE3)

They are not decoupling. Nothing here is being smoothed.

A crystal’s exact frequency depends on the capacitance it sees across its terminals. Load it differently and it runs at a slightly different frequency, which is a property of the quartz rather than a flaw. So a crystal’s datasheet states a load capacitance, and that figure is a condition on the promise: run me with 12 pF and I will give you the frequency on the label.

The two capacitors are how a board supplies it. They sit in series as far as the crystal is concerned, so two 22 pF parts present about 11 pF, plus whatever the pins and the traces contribute.

That last term takes this chapter into EE7. Stray capacitance from the copper is part of the load, so the same schematic on two different layouts presents two different loads, and the usual practice is to keep those tracks short and away from everything else.

Presence is checkable, value is not (EE3) 

$ agni check --verdicts --rule crystal-load-caps cmd/agni/testdata/conformance/crystal.fires.edn
crystal-load-caps  1 fail, 1 pass
    fail  Y1.2  no capacitor sits on terminal net XTAL_OUT, and U1's datasheet was not read, so
                whether it integrates the load capacitance is unverified
    pass  Y1.1  capacitor C1 sits on terminal net XTAL_IN

2 verdicts across 1 rule(s), 1 pass, 1 fail

$ agni check --verdicts --rule crystal-load-caps cmd/agni/testdata/conformance/crystal.passes.edn
crystal-load-caps  2 pass
    pass  Y1.1  capacitor C1 sits on terminal net XTAL_IN
    pass  Y1.2  capacitor C2 sits on terminal net XTAL_OUT

2 verdicts across 1 rule(s), 2 pass

The first board is missing a load capacitor on one terminal. The subject is Y1.2, a terminal rather than the part, because each side needs its own capacitor and having one is not half a solution. The second board has both and passes twice.

The rule does not check the value, and by now the reason should be familiar. Presence is a fact about the netlist. The right value needs the crystal’s specified load from its datasheet, the chip’s own pin capacitance from another datasheet, and the stray capacitance of a layout that may not exist yet. Chapter 3 drew the same boundary around decoupling, and chapter 4 drew it around pull-ups. The rule handles it the same way, by making the check a netlist allows and stating the rest rather than pretending to it.

Getting the value wrong does not stop the clock. It shifts it, by a small percentage, which is enough to matter. A UART running a few percent off will still frame most characters and corrupt some. A CAN controller off frequency fails at the far end of the bus first. Bugs like these get blamed on cables.

More is worse (EE5)

Some parts are ceramic resonators with the load capacitors built in. Cheaper, less accurate, and physically one three-terminal part where a crystal plus two caps is three parts. They present their specified load internally, so they need nothing external.

The capacitors can also live at the other end, inside the chip that drives the crystal. Many real-time clocks and low-power MCUs build them into their oscillator pins, often selectable in a register, so a 32.768 kHz crystal sits straight across two pins with nothing else on the board. Only the chip’s datasheet says so. When that datasheet is in the run and binds the capacitance to that pin, crystal-load-caps passes such a terminal. Otherwise it reports the terminal and names the chip it could not read.

Fit external caps to one anyway and the load roughly doubles. The oscillator now sees far more capacitance than either part was specified for, so it starts slowly, starts off-frequency, or does not start at all over temperature and part spread.

The rule’s own Impact describes how that arrives: “The board usually oscillates on the bench and drifts or drops out in the field, and every timed peripheral clocked from it (UART baud, CAN bit timing) goes with it. It is a silent BOM/layout carry-over from a crystal design, invisible until the load is measured.”

The carry-over is the mechanism. Nobody adds redundant capacitors on purpose. Somebody copies an oscillator block from a previous design, swaps the crystal for a cheaper resonator, and leaves the two capacitors in place because they were already there and removing parts feels like the risky edit.

What you can now answer

  • Why a crystal produces nothing on its own, and where the oscillator actually is. (EE3)
  • What the two capacitors are for, given that nothing is being smoothed. (EE3)
  • Why the same schematic can run at two frequencies on two layouts. (EE3)
  • Why a rule checks that the capacitors are present and says nothing about their value. (EE3)
  • Why adding load capacitors to a part that already has them is a defect rather than margin. (EE5)

The rules this page explains

Rule Severity What it catches
crystal-load-caps warning a crystal terminal with no load capacitor to ground
resonator-redundant-load-caps warning a resonator with integrated caps carrying external ones too

The second has no fixture in this course, so it is described from its own catalog entry rather than run. The pair is worth reading together, because one fires when a capacitor is missing and the other when one is present, on parts that look nearly identical on a schematic.

The next chapter, when the copper matters, is the last one, and covers the level that everything so far has deferred to.