Electronics Guide

The Role of a Voltage Reference

A voltage reference differs in purpose from a voltage regulator. A regulator exists to deliver power, holding a rail steady while supplying substantial current to a load; accuracy is secondary to current capability and efficiency. A reference exists to define a voltage, prioritizing accuracy and stability over current delivery; many references are intended to drive only high-impedance loads and supply at most a few milliamperes. The two are related, because a regulator typically contains a reference that sets its output, but their specifications and design emphases diverge sharply.

The accuracy demanded of a reference follows directly from the resolution of the system it serves. Consider a sixteen-bit analog-to-digital converter operating from a five-volt reference. One least significant bit corresponds to five volts divided by sixty-five thousand five hundred thirty-six, or about seventy-six microvolts. For the reference error to remain below half a least significant bit across the operating temperature range, the reference must hold to roughly thirty-eight microvolts, a fractional stability near eight parts per million. Achieving such stability against temperature, supply variation, loading, and years of aging is the central problem of precision reference design.

A reference is characterized by a small set of parameters, each of which contributes a term to the total error budget: initial accuracy at a standard temperature, temperature coefficient, output noise, long-term drift, line and load regulation, and output drive capability. No single parameter captures reference quality; the appropriate device depends on which error term dominates in a given application. A battery-powered sensor may be limited by quiescent current, a precision weigh scale by long-term drift, and a high-resolution converter by noise. The sections that follow treat each technology and parameter in turn.

Bandgap References

The bandgap reference, introduced by Robert Widlar at National Semiconductor in 1971 with the LM113, is the dominant technology for integrated references because it can be built in any standard semiconductor process and operates from low supply voltages. The technique combines two voltages with opposing temperature coefficients to synthesize a temperature-stable output whose value tracks the bandgap of silicon, near 1.2 electron volts. The figure that appears directly in the circuit is the base-emitter voltage extrapolated to absolute zero, which lands close to 1.205 volts; this extrapolated intercept, rather than the physical low-temperature gap of about 1.17 electron volts, is the value a first-order bandgap cell reproduces.

The principle rests on two complementary behaviors of bipolar junctions. The base-emitter voltage of a forward-biased transistor decreases with temperature at roughly two millivolts per degree Celsius, a quantity that is complementary to absolute temperature. The difference in base-emitter voltage between two transistors operating at different current densities increases with temperature in proportion to absolute temperature, a quantity termed proportional to absolute temperature, or PTAT. By scaling the PTAT term so that its positive slope exactly cancels the negative slope of a single base-emitter voltage, the sum becomes temperature-independent to first order and settles near 1.2 to 1.25 volts. Widlar's original LM113 produced 1.220 volts. The internal circuit techniques that realize these cells on silicon, including the PTAT generator and curvature correction, are developed in Bandgap References and Voltage Biasing; the focus here is on the finished reference as a system component.

A first-order bandgap cell cancels only the linear part of the base-emitter temperature characteristic. Because that characteristic is actually slightly curved, a bow-shaped residual error remains, peaking at the temperature extremes and limiting simple designs to temperature coefficients of twenty to fifty parts per million per degree Celsius. Curvature-corrected references add a higher-order compensating term, generated from the temperature dependence of transistor behavior or from resistors with deliberately chosen coefficients, that flattens this residual. Production curvature-corrected parts reach temperature coefficients below five parts per million per degree Celsius, and laboratory designs approach one.

The native bandgap output near 1.25 volts is inconveniently low for most systems, so integrated bandgap references include an on-chip amplifier that scales the cell to a standard output such as 2.5, 3.0, 4.096, or 5.0 volts. The value 4.096 volts is favored in data-converter work because it makes one millivolt correspond to one least significant bit in a twelve-bit converter, simplifying scaling. The gain stability, offset, and drift of this output amplifier add to the error budget and must be counted alongside the cell itself.

Zener and Buried-Zener References

The Zener diode is the oldest electronic voltage reference, exploiting the controlled reverse breakdown of a p-n junction. Two distinct mechanisms produce this breakdown, and their temperature behaviors differ. True Zener breakdown, dominant below about five volts, arises from quantum-mechanical tunneling and exhibits a negative temperature coefficient. Avalanche breakdown, dominant above about five volts, arises from impact ionization and exhibits a positive temperature coefficient. Near five to six volts the two mechanisms contribute comparably and their coefficients partially cancel, which is why reference diodes are often specified near 5.6 or 6.2 volts. A practical compensated reference diode places the breakdown junction in series with one or more forward-biased junctions whose negative coefficient trims the net temperature drift toward zero.

A conventional Zener junction forms at the silicon surface, where it is exposed to ionic contamination, mechanical stress, and surface trap states. These surface effects generate substantial low-frequency noise and cause the breakdown voltage to drift over time. The buried-Zener, or subsurface-Zener, structure addresses both problems by placing the breakdown junction beneath the surface, capped by a protective layer of silicon. Isolated from surface contamination and stress, the buried junction produces primarily well-behaved avalanche shot noise with a flat spectrum, and its breakdown voltage is far more stable over years of operation.

The performance advantage of the buried structure is large. Buried-Zener references exhibit low-frequency noise ten to one hundred times lower than surface Zeners and routinely achieve long-term drift below ten parts per million per year, with the best laboratory-grade parts specified near two parts per million per year or better. These properties make buried-Zener references the technology of choice for calibration equipment and the highest-resolution data converters, where noise and drift, rather than temperature coefficient alone, set the accuracy limit.

The highest-performance buried-Zener parts integrate a temperature-stabilizing system on the die. The Linear Technology LTZ1000, for example, combines a roughly seven-volt buried Zener with an on-chip temperature-sensing transistor and a heater resistor; with external bias circuitry, the heater holds the die at a constant elevated temperature so that ambient changes no longer reach the junction. The part is specified at a temperature coefficient near 0.05 part per million per degree Celsius and about 1.2 microvolts peak-to-peak of low-frequency noise. The later ADR1000 is a pin-compatible successor that carries the same on-die elements and still relies on external bias and oven-control circuitry, but a refined process lowers its noise by roughly a third and tightens its long-term drift, at a slightly reduced 6.62-volt output. The ADR1001 goes further and integrates the entire signal-conditioning chain, including the thermostat, heater control, and output buffer, into one package, so that a complete oven-stabilized reference no longer requires the discrete support network the LTZ1000 demands. Such parts represent the practical ceiling of solid-state reference performance, short of the Josephson quantum standards used to define the volt at national metrology institutes.

Series and Shunt Topologies

Reference integrated circuits are applied in one of two circuit topologies, and the choice affects supply current, supply-voltage range, and noise. A shunt reference behaves like a precision Zener diode. It is a two-terminal device placed in parallel with the load and biased through an external resistor connected to the supply; the resistor sets the current, and the reference sinks whatever current is not drawn by the load while holding its terminal voltage constant. Because a shunt reference contains no output amplifier, it tends to be quiet, tolerates a wide and varying supply because the series resistor absorbs the difference, and operates over a broad current range. Its drawback is that the bias resistor dissipates power continuously, and the designer must size it to guarantee adequate reference current at the lowest supply voltage and highest load, which wastes current at the opposite extreme.

A series reference is a three-terminal device with separate input, output, and ground connections that contains an internal pass element or buffer. It draws current from the input and delivers a regulated output without an external series resistor, much like a small low-dropout regulator dedicated to producing a reference voltage. A series reference is simpler to apply, dissipates only the quiescent current plus the load current rather than a fixed bias current, and is therefore preferred in power-conscious designs. Its limitations are a minimum input-to-output headroom below which it cannot regulate and, often, somewhat higher output noise than a comparable shunt part because of the active buffer.

The selection between the two is driven by the system. A shunt reference suits a fixed, well-defined bias point, an unusually wide or noisy supply, or a need for the lowest possible noise, and it is convenient for setting comparator thresholds and clamp levels. A series reference suits battery-powered and variable-load designs where the fixed dissipation of a shunt bias resistor is unacceptable, and it dominates general-purpose use because of its simplicity. Many adjustable shunt references, of which the TL431 with its internal 2.495-volt bandgap is the archetype, allow an external resistor divider to program the regulation voltage, which is why they are ubiquitous in switching-supply feedback paths.

Initial Accuracy and Trimming

Initial accuracy specifies how closely the output equals its nominal value at a standard temperature, conventionally twenty-five degrees Celsius. It is stated either as an absolute voltage tolerance or as a percentage of nominal. High-grade references guarantee initial accuracy of 0.02 percent or better, equivalent to two hundred parts per million, while general-purpose grades may specify 0.1, 0.5, or 1 percent. Initial accuracy reflects the manufacturer's ability to trim the reference at production test and the matching of the internal components.

Initial error is often the least troublesome reference parameter, because it is a fixed offset that system calibration can remove. A system that measures its reference once against a known standard, or that operates ratiometrically, can subtract the initial error in software and is then limited by the parameters that change after calibration: temperature coefficient, drift, and noise. For this reason a designer frequently accepts a looser initial-accuracy grade and budgets for a one-time calibration, reserving the cost of a tight-tolerance part for situations where no calibration step is available.

Manufacturers achieve their specified initial accuracy by trimming each part during test, historically by laser-trimming thin-film resistors and increasingly by storing digital correction in on-chip memory. The economics of trimming produce the grade structure visible in any reference data sheet, where the same die is sold at several initial-accuracy and temperature-coefficient grades sorted by test results. Selecting the grade is therefore a direct trade of cost against the portion of the error budget that calibration cannot reach.

Temperature Coefficient

The temperature coefficient quantifies how the output voltage changes with temperature and is usually the dominant error term in systems that operate over a wide temperature range. It is specified in parts per million per degree Celsius. A reference with a ten-part-per-million-per-degree coefficient changes by one thousand parts per million, or 0.1 percent, across a one-hundred-degree span. First-order bandgap parts fall in the twenty-to-fifty range, curvature-corrected bandgap parts reach below five, precision buried-Zener parts reach one or below, and oven-stabilized parts such as the LTZ1000 approach 0.05.

A subtlety of temperature-coefficient specification is the method by which it is computed, because a reference does not drift linearly with temperature. The box, or min-max, method divides the total peak-to-peak output variation across the operating range by the temperature span; this method is conservative and is preferred for precision parts because it bounds the worst-case error anywhere in the range. The slope, or butterfly, method fits a line through the endpoints and reports its slope, which can understate the true excursion when the characteristic is curved. A designer comparing two data sheets must confirm that both quote the coefficient by the same method before treating the numbers as comparable.

Temperature coefficient is, in principle, correctable if the system measures temperature accurately and applies a characterized correction, and some precision instruments do exactly this to reach effective coefficients below the raw device specification. In practice, the correction is limited by the accuracy of the temperature measurement, by hysteresis in the reference between heating and cooling cycles, and by self-heating that changes the die temperature with load and supply. For most designs the practical path to low temperature error is to select a part whose specified coefficient is small enough that no correction is needed.

Noise and Long-Term Drift

Noise sets the resolution floor that a reference imposes on a measurement, and it cannot be removed by calibration. Reference noise arises from several sources: shot noise from junction currents, thermal noise from internal resistances, and, most troublesome, flicker noise that rises at low frequency as the inverse of frequency. Two regions of the noise spectrum matter in different applications. The broadband noise density, quoted in nanovolts or microvolts per root hertz, governs high-speed sampling, while the low-frequency, or one-tenth-hertz-to-ten-hertz, peak-to-peak noise governs slow, high-resolution measurements such as those in a precision multimeter or weigh scale.

Buried-Zener references achieve the lowest noise because their subsurface junctions avoid surface-generated flicker noise; broadband densities below one hundred nanovolts per root hertz and low-frequency noise of a few microvolts peak-to-peak are characteristic, and the LTZ1000 reaches about 1.2 microvolts peak-to-peak. Bandgap references are noisier, typically by a factor of several, because the output amplifier and the bandgap cell both contribute; chopper-stabilized bandgap parts suppress flicker noise through auto-zeroing but can introduce a tone at the chopping frequency. Output noise can be reduced by filtering, but only at the cost of settling time, since a low-pass filter heavy enough to suppress flicker noise lengthens the time the reference needs to stabilize after power-up; the trade is acceptable in many systems because the reference voltage is essentially static once established.

Long-term drift describes the slow, irreversible change in output that accumulates over months and years and is specified in parts per million per one thousand hours or per year. It originates in aging mechanisms such as ionic migration in passivation, mechanical stress relaxation in the package and die attach, and gradual changes in defect populations; surface effects dominate aging in conventional Zeners, which again favors the buried structure. Drift is largest during initial operation and slows logarithmically thereafter, which is why precision parts are often specified after a burn-in period and benefit from extended operation before use in critical service. The best buried-Zener references hold below two parts per million per year, while bandgap references typically drift ten to fifty parts per million per year. For instruments expected to maintain accuracy between calibration intervals, long-term drift, together with noise, sets the recalibration period.

Output Drive and Application Practice

The output-drive capability of a reference describes how much current it can source or sink while holding its specified accuracy, and it is closely tied to load regulation, the change in output voltage per unit change in load current. Many precision references are intended to drive only high-impedance loads and tolerate at most a few milliamperes; loading them beyond that range introduces error through finite load regulation or, in the worst case, drops them out of regulation entirely. When a reference must drive a heavier or dynamic load, it is buffered by a precision unity-gain amplifier that supplies the current without changing the voltage. The buffer's own offset, bias current, noise, and drift then add to the reference error, so a low-offset, low-drift amplifier must be chosen, and the combined reference-and-buffer behavior should be calibrated together where the highest accuracy is required. Operational-amplifier building blocks suited to this role are treated in Operational Amplifiers and Linear Circuits.

Even an excellent reference is easily degraded by poor application. Supply noise reaches the output through the reference's finite line regulation and power-supply rejection, so a clean, well-bypassed input and decoupling with low-equivalent-series-resistance capacitors close to the supply pin are essential; a low-dropout regulator ahead of the reference often supplies the necessary quiet rail. Output capacitance must respect the reference's stability requirements, because some series references oscillate with the wrong capacitor value or equivalent series resistance, a constraint analogous to that of low-dropout regulators. Self-heating from internal dissipation shifts the die temperature away from ambient and should be minimized, and the reference should be kept physically away from heat-generating components to avoid thermal gradients across the die.

Printed-circuit layout determines whether the reference's rated performance survives in a real system. The output should be sensed with a Kelvin, or four-wire, connection so that current-carrying traces do not introduce series resistance into the sense path, and the ground reference should return to a single point to avoid ground-drop errors. Sensitive nodes can be surrounded by guard traces held at the same potential to intercept leakage currents, and the reference should be separated from high-current and high-frequency circuitry to prevent coupling. These practices, drawn from precision analog discipline generally, often make the difference between achieving the data-sheet specification and falling short of it.

Applications in Data Converters

Data converters are the application that most sharply defines what a precision reference must do, because the reference directly sets the converter's full-scale voltage and therefore the absolute meaning of every code. In an analog-to-digital converter the reference establishes the input voltage that corresponds to the maximum code; in a digital-to-analog converter it sets the output voltage that corresponds to full-scale digital input. Any error in the reference, whether offset, temperature drift, or noise, maps directly onto the converted result, scaled by the fraction of full scale at which the converter is operating. The companion topic Analog-to-Digital and Digital-to-Analog Conversion develops converter architectures in detail; the concern here is the reference they require.

The accuracy a converter can deliver is bounded by its reference. As the earlier sixteen-bit example showed, holding the reference error below half a least significant bit over temperature already demands stability of a few parts per million, and at eighteen and twenty bits the requirement tightens to the point that buried-Zener references and careful thermal design become mandatory. The reference noise matters just as much as its drift, because broadband reference noise adds directly to the converter's input-referred noise and degrades the effective number of bits; this is why high-resolution converters specify a maximum reference noise and why the reference output is often filtered to suppress flicker noise before it reaches the converter.

A consideration unique to converters is the dynamic load the reference presents. A successive-approximation converter switches its capacitive array on every bit decision, drawing brief charge transients from the reference at the sampling rate, and a delta-sigma converter draws a switched-capacitor load that depends on its clock. The reference and its decoupling network must settle from these transients within the available time, or the conversion incurs a settling error that appears as a code-dependent nonlinearity. The remedy is a reference with adequate dynamic output impedance, buffered where necessary, and a decoupling capacitor sized to supply the instantaneous charge, which is why converter data sheets specify a reference bypass capacitor and why a reference and a low-impedance buffer are frequently sold and applied as a matched pair.

A final consideration is whether the reference defines an absolute scale or only a ratio. In a ratiometric measurement, where the same reference excites a sensor bridge and sets the converter full scale, the reference value cancels from the result and its drift and initial error become almost irrelevant, leaving noise as the principal concern; this arrangement is common in strain-gauge and resistive-sensor systems and relaxes the reference requirement substantially. In an absolute measurement, where the converter result must correspond to a true voltage in volts, the full burden of initial accuracy, temperature coefficient, and long-term drift falls on the reference. Recognizing which case applies is the first step in specifying a converter's reference, because it determines whether a modest bandgap part suffices or a metrology-grade buried-Zener reference is required.

Summary

A precision voltage reference defines the absolute voltage scale of an analog system, and its quality sets the ceiling on the accuracy of every measurement and conversion that depends on it. Two technologies dominate: the bandgap reference, which sums complementary base-emitter and proportional-to-absolute-temperature voltages to produce a stable output near silicon's bandgap and integrates readily in any process, and the buried-Zener reference, whose subsurface breakdown junction delivers the lowest noise and best long-term stability for metrology-grade work. Each is applied through a shunt topology, quiet and supply-tolerant but continuously dissipating a bias current, or a series topology, simpler and more power-efficient but constrained by input headroom.

The parameters that distinguish references divide into those a system can calibrate away and those it cannot. Initial accuracy is a fixed offset that one-time calibration removes, while temperature coefficient, noise, and long-term drift change after calibration and therefore set the real accuracy limit; temperature coefficient is reported in parts per million per degree Celsius by box or slope methods that must be compared with care, noise sets a resolution floor in both broadband and low-frequency bands, and drift sets the recalibration interval. Output drive, line and load regulation, decoupling, thermal control, and Kelvin-sensed layout determine whether the rated performance survives in a real circuit.

Data converters express these requirements most exactly, because the reference sets full scale and its every error maps onto the result. The resolution of the converter dictates the stability the reference must hold, the converter's switched load dictates the reference's dynamic impedance and decoupling, and whether the measurement is ratiometric or absolute decides how much of the reference's accuracy actually matters. Specifying a reference well is, in the end, a matter of identifying which error term governs a given application and choosing the technology, topology, and grade that bring that term within budget.

Related Topics

• Voltage References - Broader survey of reference circuits, from simple Zeners to integrated reference ICs and current references
• Bandgap References and Voltage Biasing - On-chip design of bandgap cells, PTAT generators, and curvature correction
• Analog-to-Digital and Digital-to-Analog Conversion - Converter architectures whose full-scale accuracy a reference defines
• Linear Voltage Regulators - Regulators that deliver power and are built around an internal reference
• Operational Amplifiers and Linear Circuits - Buffer and amplifier building blocks used to drive loads from a reference
• Power Supply and Voltage Regulation - Parent category covering regulation, references, and supporting circuits