Voltage References: The Stable Yardstick

Why every supply needs a yardstick

In guide 2 you built a linear regulator as a thermostat: it watches a fraction of its own output and forever nudges the pass transistor to hold that rail steady. But a thermostat is useless without a setpoint — steady against what? The honest answer is that a regulator does not actually know what 5.00 V is. It only knows how to keep one voltage equal to another, and that other voltage — the one it trusts completely — is the voltage reference. The whole negative-feedback loop is a comparison, and the reference is the thing every measurement is compared to.

So the output is only ever as good as that yardstick. If the reference reads 2.500 V and the feedback divider scales it by 2, the rail lands at exactly 5.000 V; if the reference quietly drifts to 2.51 V on a hot day, your supposedly fixed rail drifts with it, to 5.02 V, and nothing downstream can tell. A reference made of rubber gives you a regulated supply made of rubber. Its one job, then, is to be a rock: a fixed voltage that ignores the input voltage, ignores the load current, and ignores temperature.

The Zener: breakdown turned into a benefit

The simplest reference reuses a part you already met in the diode rung. Push a Zener diode hard in reverse and, past its breakdown voltage, it conducts hard while holding the voltage across itself nearly constant — that steep vertical wall on the reverse side of its I-V curve. Feed it through a series resistor from a higher, messier rail, and the resistor swallows the difference while the Zener pins the node. You have turned a controlled, harmless breakdown into a fixed voltage.

      R = 1.4k
Vin o--/\/\/--+-------o  Vref ~ 5.1 V  (the reference node)
+12V         |
            _|_
            \ /   Zener 5.1 V, reverse-biased,
            ---       sitting in breakdown
             |
GND o--------+-------o  GND

Pick R so the Zener stays in breakdown at a sensible current:
  R   = (Vin - Vz) / Iz = (12 - 5.1) / 0.005  = 1.38k  -> use 1.4k
  P(R)     = (12 - 5.1) x 0.005 = 35 mW
  P(Zener) =  5.1       x 0.005 = 26 mW   (both run cool)

It works, and for a coarse job — a logic threshold, an over-voltage clamp — it is perfect. But be honest about its faults. The breakdown knee is soft, not a true vertical, so the Zener has a real dynamic resistance (often tens of ohms): draw more current and the voltage shifts a little, which is why any serious load must be buffered. Worse, the voltage drifts with temperature, and the direction depends on the part. Below about 5 V the true Zener mechanism gives a negative tempco; above about 6 V the avalanche mechanism gives a positive one. Right around 5.6 V the two effects roughly cancel, which is why 5.1 V and 5.6 V Zeners are the classic reference choices.

And there is a deeper limit. A Zener needs several volts of headroom just to break down, so you cannot build a 1.2 V or 2.5 V reference from one, and it is hopeless on a 3.3 V system. Its noise is also relatively high. The Zener is the rough draft of a reference — honest, cheap, and good enough for a clamp, but not the precision yardstick a 12-bit converter demands. For that we need a cleverer trick.

The bandgap: cancelling temperature on purpose

The masterpiece is the bandgap reference, and its idea is beautiful: build two voltages that drift with temperature in opposite directions, then add them so the drifts cancel. The first ingredient is the forward voltage of a transistor's base-emitter junction. That junction drop is not the fixed 0.7 V of the cartoon model — it slides down by about -2 mV for every degree Celsius of warming. Call that the falling ingredient.

The second ingredient is the rising one. Run two junctions at different current densities and take the difference of their two base-emitter drops; physics says that difference is proportional to absolute temperature — it climbs as things get hotter. Multiply that small rising voltage by just the right gain so its upward slope exactly matches the junction's downward slope, add the two, and the temperature dependence vanishes. The sum lands at about 1.205 V — the silicon band gap extrapolated to absolute zero, which is where the name comes from.

Reading a reference's datasheet

Once you stop treating a reference as just "a fixed voltage" and start reading its spec, a handful of numbers decide everything. Initial accuracy is how close it is at room temperature out of the box (say plus or minus 0.1%). Temperature coefficient (tempco), quoted in ppm per degree C, is how far it crawls as the temperature changes — the single most important number for a precision part. Then noise (how much it jitters, often quoted as microvolts over a band), line regulation (drift when the input rail moves) and load regulation (sag when you draw current), and finally long-term drift (how it ages over a thousand hours).

Tempco is worth a worked example because the units fool people. "ppm" is parts per million: 1 ppm of 2.5 V is 2.5 microvolts. Take a modest 25 ppm/degC reference at 2.5 V and warm it across a 50 degC range. The drift is 25 times 50 = 1250 ppm, which is 0.125%, or in volts 25 times 10^-6 times 50 times 2.5 = 3.1 mV. That sounds tiny — until you feed it to a converter, as you will see next. A premium part might be 2 ppm/degC; a buried-Zener reference (a Zener fabricated under the silicon surface, away from noisy contamination) can be lower-noise and more stable still, at a price.

1. Pick a topology. A shunt reference acts like an ideal Zener (two terminals, needs its own series resistor); a series reference is a three-terminal part like a small regulator (in, out, ground) and is usually easier to drop in.
2. Set your accuracy budget. Add up initial accuracy plus tempco-over-your-temperature-range plus drift, and check the worst-case total is smaller than the error you can tolerate.
3. Decouple it. Put the manufacturer-recommended capacitor right at the output pin — a clean reference is easy to dirty with noise, and the wrong cap can even make some references oscillate, so follow the datasheet.
4. Buffer it if it must drive a load. Follow it with a voltage follower so the reference itself stays lightly loaded and the world sees a low-impedance copy.

The reference as full-scale ruler

A reference's most demanding customer is not a regulator at all — it is a converter. An ADC does not measure volts; it measures fractions of its reference. A 12-bit ADC with a 2.5 V reference slices that range into 4096 steps, so one step (one LSB) is 2.5 / 4096 ≈ 0.61 mV. Now recall that 3.1 mV of drift from our 25 ppm part: it is about five LSBs of error, quietly corrupting the bottom bits of every reading. The expensive 12-bit converter is no more accurate than the cheap reference feeding it — the yardstick limits the ruler.

There is one gorgeous escape, and it turns the reference's flaw into a non-issue. Power your sensor and the ADC's reference from the very same voltage. Now if that voltage drifts, the sensor's output and the converter's full scale drift together by the same fraction, and the ratio — which is all the ADC reports — does not change at all. This is ratiometric measurement, and it is exactly how a resistive bridge or potentiometer is read: when you only care about a ratio, a shared reference makes its drift cancel out, and a humble reference becomes good enough.

That closes the reference. You now own the whole spectrum: the Zener rough draft, the bandgap masterpiece hiding inside every chip, the datasheet numbers that separate a good one from a great one, and the ratiometric trick that sidesteps drift entirely. Carry one fact forward: every regulator in this rung — including the efficient switching converter of guide 4 — still measures its feedback against exactly this kind of yardstick. Master the reference and you have mastered what "steady" really means.