Analog Electronics

Bandgap Voltage Reference: Cancelling Temperature Drift at 1.25 V

Cool a silicon diode by 100 °C and its forward voltage climbs by about 0.2 V — roughly −2 mV for every degree. That drift would wreck any precision measurement, yet nearly every op-amp, ADC, DAC, and voltage regulator built since 1971 hides a tiny circuit that turns this drift against itself and holds a rock-steady voltage near 1.25 V. That circuit is the bandgap voltage reference.

A bandgap reference sums two voltages with opposite temperature slopes: a base-emitter voltage VBE that falls with temperature (CTAT), and a scaled thermal-voltage term that rises with temperature (PTAT). Tuned correctly, the slopes cancel to first order, and the sum lands almost exactly on the extrapolated silicon bandgap energy — about 1.205 V at absolute zero — hence the name.

  • TypeOn-chip precision voltage reference
  • Nominal output~1.20-1.25 V (silicon bandgap)
  • Key equationV_REF = V_BE + (kT/q)·ln(N)·(R2/R1)
  • InventedWidlar 1971 (LM113); Brokaw cell 1974
  • Typical drift10-50 ppm/°C (uncorrected), <5 ppm/°C corrected
  • Used inADCs, DACs, LDO regulators, op-amps, temp sensors

Interactive visualization

Press play, or step through manually. The visualization is yours to drive — try it before reading on.

Open visualization fullscreen ↗

Watch the 60-second explainer

A condensed visual walkthrough — narrated, captioned, under a minute.

What it is and where it hides

A bandgap voltage reference is an integrated circuit that produces a stable DC voltage — canonically near 1.25 V — that barely moves with temperature, supply voltage, or load. It is the silent workhorse of analog electronics: almost every data converter, regulator, and precision amplifier needs a fixed voltage to compare signals against, and the bandgap is the reference of choice because it can be built entirely from the bipolar transistors and resistors already available on any chip.

  • ADCs and DACs use it to define full-scale; a 12-bit converter's LSB is meaningless if its reference wanders.
  • LDO and switching regulators compare a divided-down output to the bandgap to set their regulated rail.
  • Op-amps, comparators, and PLLs use it to bias internal currents.
  • Silicon temperature sensors reuse the same PTAT term as a thermometer.

Unlike a Zener diode, which needs 6-7 V of headroom, a bandgap works from supplies as low as 1.5-2 V, which is why it became universal as chip voltages fell.

How it works: cancelling two opposing slopes

The trick is to add two voltages whose temperature slopes are equal and opposite. A forward-biased bipolar transistor's base-emitter voltage VBE is complementary to absolute temperature (CTAT): it falls at roughly −2 mV/°C because the collector current depends exponentially on temperature through the intrinsic carrier concentration.

The counter-term comes from the difference in VBE between two transistors run at different current densities. If transistor Q2 has N times the emitter area (or 1/N the current density) of Q1:

  • ΔVBE = (kT/q)·ln(N), where kT/q is the thermal voltage VT (≈25.9 mV at 300 K).

Because VT rises linearly with temperature (+0.087 mV/°C per unit), ΔVBE is proportional to absolute temperature (PTAT) with a positive slope. Amplify ΔVBE by a resistor ratio until its +slope exactly matches VBE's −slope, sum them, and the temperature dependence cancels to first order. The sum sits at the extrapolated silicon bandgap voltage Eg/q ≈ 1.205 V.

The governing equation and a worked example

The output of a classic bandgap is:

VREF = VBE + (R2/R1)·(kT/q)·ln(N)

Symbols: VBE = base-emitter voltage (~0.65 V, −2 mV/°C); k = Boltzmann's constant (1.381×10⁻²³ J/K); T = absolute temperature (K); q = electron charge (1.602×10⁻¹⁹ C); N = emitter-area (current-density) ratio; R2/R1 = gain resistor ratio.

Worked example. At T = 300 K, kT/q = 25.85 mV. Choose N = 8, so ln(8) = 2.079 and ΔVBE = 53.8 mV. The PTAT slope is +0.087·2.079 = +0.181 mV/°C per unit gain. To cancel VBE's −2 mV/°C we need a gain of 2/0.181 ≈ 11.0, i.e. R2/R1 ≈ 11. That adds 11 × 53.8 mV ≈ 0.59 V to the 0.65 V VBE, giving VREF ≈ 1.24 V — right at the bandgap. Residual drift from this design is typically 10-50 ppm/°C.

Designing and trimming one in practice

Real designs wrap the core in a feedback loop and then trim it. The dominant topologies are:

  • Widlar (1971): the original three-transistor cell in the LM113, using a Widlar current mirror to set the PTAT current.
  • Brokaw cell (1974): an op-amp forces equal currents through Q1 and Q2 (area ratio 8:1) across a shared emitter resistor; the reference is tapped at the transistor collectors. It is the most-copied topology in ICs.
  • Kuijk / CMOS bandgap: an op-amp servos two branches with a diode-connected substrate PNP, the standard in CMOS.

Key practical rules: (1) trim R2/R1 at the bandgap-magic voltage — nulling drift and setting output are the same adjustment, so trimming VREF to ~1.205 V minimizes slope. (2) Use a large N (8-24) laid out as a common-centroid array to fight mismatch. (3) Add a start-up circuit — the zero-current state is also a stable operating point. (4) Filter the PTAT resistor's noise; bandgaps are intrinsically noisy (~1-3 µVRMS in-band before filtering).

Bandgap vs. Zener and other references

The bandgap is not always the best reference — just the most integrable. Its main rivals:

  • Buried (subsurface) Zener: breakdown ~6.3 V, achieving 0.5-3 ppm/°C and far lower noise, because avalanche happens below the surface away from noisy interface traps. This is why the lowest-drift references (LM399 at ~1 ppm/°C, LTZ1000 at ~0.05 ppm/°C) are Zener-based — but they need >8 V and burn power heating themselves to a stable temperature.
  • Curvature-corrected bandgaps: the raw bandgap still has a residual parabolic curvature (the VBE(T) equation carries a T·ln(T) term). Adding a correction current squashes this to 1-5 ppm/°C, as in the LTC6655.
  • XFET and floating-gate references: lower power, competitive drift, no 1.2 V floor.

The bandgap wins whenever you need a reference on the same die from a low supply — which is essentially every mixed-signal SoC.

Failure modes, trade-offs, and significance

Bandgaps have well-known limitations:

  • Curvature error: first-order cancellation leaves a bowed VREF(T) that peaks mid-range and droops at the extremes — the reason uncorrected parts spec ~20 ppm/°C over 100 °C rather than zero.
  • Start-up latch: without a start-up cell the circuit can sit at 0 V forever; a classic silicon bug.
  • Resistor and VBE mismatch: a 1% area error in the N-ratio maps directly to drift; layout and dynamic element matching matter.
  • Package stress / drift: mechanical stress shifts resistor values, causing long-term drift (10-100 ppm/1000 h) and hysteresis after thermal cycling.
  • Noise: the PTAT gain amplifies transistor and resistor noise, so low-drift bandgaps trade bandwidth for filtering.

Significance: Widlar's insight — that the messy, temperature-sensitive VBE hides a fundamental constant (the bandgap energy) once you subtract its PTAT part — made precise analog possible on cheap silicon. More than 50 years later, essentially every ADC, DAC, and regulator still starts with that 1.25 V.

Bandgap reference vs. other voltage-reference technologies
Reference typeTypical outputTemp drift (ppm/°C)Notes
Bandgap (uncorrected)1.2-1.25 V10-50Integratable, low supply, ~1.2 V floor
Bandgap (curvature-corrected)1.2-2.5 V1-5Cancels V_BE curvature; e.g. LTC6655
Buried (subsurface) Zener6.2-7.1 V0.5-3Lowest noise/drift, needs >8 V supply
Surface Zener / temp-comp5.6-6.4 V5-20Simple but noisy, drifts with surface state
XFET / FGA2.5-5 V3-10Low power, good long-term stability

Frequently asked questions

Why is the output voltage always around 1.2 V?

Because the circuit is engineered so the temperature-independent sum equals the silicon energy bandgap divided by electron charge, E_g/q. Extrapolating a transistor's V_BE to absolute zero lands at ~1.205 V, and cancelling its temperature slope forces V_REF to that value. The exact number varies slightly (1.20-1.25 V) with process and curvature correction.

What do PTAT and CTAT mean?

PTAT is 'proportional to absolute temperature' — a voltage or current that rises linearly with T, like the thermal voltage kT/q. CTAT is 'complementary to absolute temperature' — it falls with T, like a diode's forward voltage (−2 mV/°C). A bandgap adds a scaled PTAT term to a CTAT V_BE so their opposite slopes cancel.

What is the difference between a Widlar and a Brokaw bandgap?

Widlar's 1971 original (LM113) used a Widlar current mirror and three transistors to generate the PTAT current directly. Brokaw's 1974 cell (AD580) uses an op-amp to force equal currents through two transistors of unequal emitter area (typically 8:1) sharing an emitter resistor, tapping the reference at their collectors. The Brokaw topology is more accurate and became the industry standard.

Can I get a reference higher than 1.25 V from a bandgap?

Yes. The core still produces the ~1.2 V bandgap voltage, but you buffer and gain it up with a precision resistor divider or amplifier to make common values like 2.5 V, 4.096 V, or 5 V. The temperature cancellation happens at the 1.2 V core; the output scaling just multiplies both the voltage and any residual drift.

Why does a bandgap need a start-up circuit?

The feedback loop has two stable operating points: the intended one, and a degenerate state where no current flows at all (zero in, zero out satisfies the loop equations). A small start-up circuit injects a kick-current at power-up to nudge the loop off the zero-current point; once running it disconnects so it doesn't disturb the reference.

How low can bandgap drift go, and how do you improve it?

A simple first-order bandgap achieves 10-50 ppm/°C, limited by the residual T·ln(T) curvature of V_BE. Curvature-correction circuits that add a nonlinear compensation current push this to 1-5 ppm/°C (e.g. LTC6655). For sub-1 ppm/°C you generally switch to a buried-Zener reference, at the cost of higher supply voltage and power.