As early as 1972, the advent of the ac Josephson junction promised to provide improved accuracy in its representation of the volt standard. When microwave energy at a precisely known frequency is injected into a stacked assembly of Jo — sephson junctions, held at 4 K by liquid helium, it is possible to generate a voltage that is accurate and stable to better than 0.1 JU. V/V, both theoretically and in practice (36).

Preliminary research confirmed that even the best Weston saturated standard cells had unexplained drifts and noises, of the order of 1 X 10—6, which the Josephson junctions did not. As the Josephson junction equipment became more reliable and easier to operate, it became obvious that they would soon make possible a new, more stable standard. After a consider­able amount of engineering and development, a new represen­tation of the volt was established. The Josephson constant KJ-90, adopted on January 1, 1990, was defined as 483,597.900 GHz/V, exactly.

The ac Josephson junction equipment for establishing ul — traprecise voltage references has typically a precisely known 72 GHz input frequency, and an output of 2.0678336 ^V/GHz. The output of each junction is about 149 ^V. To provide an output at a convenient level, an array of 2000 Josephson junc­tions is integrated, stacked in series, and enclosed in the cryo­genic (4 K) microwave-stimulated chamber, thus providing an output of perhaps 298 mV. This voltage is compared with the 1.018 V level using conventional potentiometric techniques, to calibrate the standard cells that act as secondary transfer references. Equipment to implement this stable reference tends to cost in the vicinity of $100,000, plus considerable la­bor and operational costs.

Thus on January 1, 1990, the magnitude of the US volt (as well the voltage standards in most other countries) was changed. The new 1990 US volt was established as +9.264 jU, V/V larger than the previous (1972) US standard. Since 1990, the international standard volt has still been defined as

1 W/1 A, but the standard representation of the volt is the output of the Josephson junction apparatus.


In theory, the volt is not an absolute standard. The volt has long been defined as the potential such that 1V X 1 A = 1 W. In turn the ampere is defined as an absolute standard, such that 1 A flowing through a pair of very long wires (of negligible diameter), separated by 1 m, will cause a force of

2 X 10—7 N per meter of length. In practice, the volt is a much more useful and usable standard. The ampere standard is not very portable. In fact, when a 1 A standard is required, it is normally constructed by establishing 1 V and 1 П, causing 1 A to flow.


In theory, the ohm is not an absolute standard, but the ratio 1 V/1 A, with the volt and ampere defined as above. As of 1990, the representation of the ohm was redefined using the quantum Hall effect (QHE), discovered by Klaus von Klitz — ing (37):

The QHE is characteristic of certain high-mobility semiconductor devices of standard Hall-bar geometry, placed in a large applied magnetic field and cooled to a temperature near one kelvin. For a fixed current I through a QHE device there are regions in the curve of Hall voltage vs. gate voltage, or of Hall voltage vs mag­netic field depending on the device, where the Hall voltage UH remains constant as the gate voltage or magnetic field is varied. These regions of constant Hall voltage are termed Hall plateaus. Under the proper experimental conditions, the quantized Hall re­sistance of the ith plateau RH(I), defined as the quotient of the ith plateau to the current I, is given by

RH(i) = UH (i)/I = RK/i

where i is an integer and RK is now termed the von Klitzing con­stant after the discoverer of the QHE….

Numerically, RK is about 25,813 ohms. The value agreed upon as an international constant was RK.90 = 25,812.807 ohms.

This was a considerable improvement, as the best older stan­dard resistors were shown to be drifting at about —0.1 ^П/П per year. With the quantum standard, such drifts are ban­ished.


As mentioned above, many kinds of ingenious internal archi­tectures have been used for bandgap references. The actual topology of the bandgap elements has been arranged in many ways—ways that may (or may not) be transparent to the user. However, there are also features that are useful to some users and of no value to other users. A list of typical features is provided here, along with a brief list of typical ICs that include those features. This list of ICs is by no means in­tended to be exhaustive, but merely indicative of what fea­tures one might look for.

• Low Power. Many users like the advantages of low power, but often a low-power bandgap reference is noisy, because transistors operated at small currents tend to have voltage noise inversely proportional to the square root of the emitter current. LM185-1.2 (10 ^A), ADR291 (15 ^A).

• Low Noise. A bandgap reference operated at higher cur­rent tends to be less noisy. LM336, ADR291.

• Shutdown. Sometimes it is important to turn the device off for minimum system power drain.

• Startup. Not all references start up very quickly.

• Shunt-Mode Operation. In some applications, a shunt­mode reference (which looks like a Zener) is easier to apply. LM4040, LM336, LT1009, AD1580.

• Series-Mode Operation. For applications where the load current may vary, series mode can be advantageous. AD581, LT1019, many others.

• Output Can Sink or Source Current. Sometimes it is con­venient if the output can drive load currents in both di­

rections. If you need this, be cautious, as load regulation when sinking is usually inferior to when sourcing. MAX6341, AD581, LT1019, many others.

• Output Is Trimmable—in a Narrow Range. Beware that if you use this feature to adjust Vout a significant amount, the tempco may be degraded.

• Output Is Adjustable—over a Wide Range. This is some­times a nice feature, but the accuracy, stability, and tempco of external resistors must be considered. LM385 (adjustable), LM4041-ADJ.

• Filter Pin. As band-gap references are fairly noisy (some­times comparable to 1 LSB of a 10- or 12-bit system) a provision for an external noise filter capacitor is some­times very useful. LM369, AD587.

• Temperature Sensor. Some units provide a temperature sensor output at —2.2 mV/°C. REF01, LT1019, many oth­ers.

• Heater. Some units provide a resistor on-chip that can be used to heat the unit to a constant warm temperature. LT1019.

• Low-Dropout Regulator. Many modern references need a supply voltage only 0.1 V or 0.2 V larger than Vout: a pop­ular feature. See also below.

• Requirement for Capacitive Load. Most low-dropout ref­erences require a capacitive load on their output to pre­vent oscillations: an unpopular feature, as sometimes the capacitor is bigger or more expensive than the IC.

• Tolerance of Cload. Some references will tolerate any ca­pacitive load, and may not require any load capacitor at all. LM385, LT1009.

• Low Tempco. A very desirable feature.

• After-Assembly Trim. This is a procedure for optimizing the low tempco of a reference. However, it uses pins for connection to in-circuit trims such as fuses or Zener zaps—so that these pins cannot be used for other fea­tures. LT1019, LM169.

• Small Packages. Many small systems require surface — mount devices. Packages such as SO-8 or SOT-23 are popular. However, tiny plastic packages tend to cause stresses which may degrade long-term stability.

• Long-Term Stability. A very desirable feature, but not trivial to find, and not easy or inexpensive to test for. In fact, stability is just about the most expensive specifica­tion that one can buy on a reference.

• Compromises. No reference can provide every advan­tage, so priorities and tradeoffs must be engineered.

• Price. Any combination of excellent features and/or spec­ifications is likely to command a high price. This leads to compromises; see above.


The concept of the bandgap reference was first published by David Hilbiber of Fairchild Semiconductor in 1964 (20). If a suitable circuit is used to add voltages with both positive and negative temperature coefficients, until the output voltage is approximately the same as the bandgap voltage of the mate­rial used (1.205 V for silicon, 0.803 V for germanium) a stable, low-tempco reference can be achieved.

The invention of the LM113 bandgap reference IC (21) by Robert J. Widlar in 1971 was rather like the birth of a baby— just a beginning. While this small IC was useful for instru­ment makers who needed a reference that would run on low voltages (such as 4.5 V or 3 V or even 1.5 V), it definitely did not have superior performance. The standard LM113 had an output voltage of 1.220 V with a tolerance of ±5%, a fairly broad spread of temperature coefficients, and mediocre long­term stability. Still, the principle and the feasibility of the bandgap reference had been proved, and the development of new bandgap reference circuits with needed improvements soon commenced. These efforts have continued for well over 25 years without much diminishment.

The bandgap reference was first used in the NSC LM113 reference circuit (1971) and the LM109 Voltage Regulator (1969) (22). The bandgap circuit employs transistors op­erating at different current densities, such as 10 p, A flowing through one transistor with one emitter, and the same amount of current flowing through another transistor 10 times as big. This generates a voltage (AVbe) of perhaps 60 mV, which is a voltage proportional to absolute temperature (VPTAT). This voltage is amplified and added to a voltage proportional to the transistor’s base-emitter voltage Vbe, which decreases fairly linearly with temperature. The addi­tion is scaled so that the total voltage is about 1.24 V dc. When the reference voltage is set or trimmed to this voltage, a low tempco is obtained.

The principle of the bandgap reference relies on a good un­derstanding of the Vbe of transistors. Widlar’s paper (23) on this subject clarified the mathematics and physics of Vbe and corrected various misconceptions.

We refer to the example of the LM113. In Fig. 2, the LM113 schematic diagram shows a basic bandgap circuit. When V+ is around 1.22 V dc, Q1 runs at a high current den­sity, about 230 nA/^m2. Q2 is operated at a low density, about 15 nA/^m2, and so its Vbe is much smaller, by about 70 mV. Now, let’s assume that the circuit is at balance and the output is near 1.22 V. Then the 70 mV across R5 is magnified by the ratio of R4 to R5, about 8.5: 1, up to a level of 600 mV. This


Figure 2. Schematic diagram, LM113 (simplified).

voltage is added to the Vbe of Q4 (about 620 mV at room tem­perature) to make a total of 1.22 V, as required. Q4 then am­plifies the error signal through Q7 and Q9, which provide enough gain to hold the V+ bus at 1.22 V. The beauty of the bandgap reference is the summation of the Vbe term, which decreases at the rate of about —2 mV/°C, and the AVbe term, which grows at about +2 mV/°C, to achieve an overall tempco that is substantially zero. All bandgap references employ this summation of a growing and a shrinking voltage to make a stable low-tempco voltage. Further, it has been shown (23) that when a circuit has been trimmed to the correct voltage, the correct tempco will follow, despite process variations in parameters such as Vbe, 3, sheet resistivity, etc. Consequently, bandgap circuits are often trimmed to their ideal voltage so as to provide a low tempco.

Since the LM109 and LM113 were introduced, many other circuits and configurations have been invented for bandgap references. The Brokaw cell is often used for output voltages larger than 1.25 V, as its Vout can be scaled by the ratio of two built-in resistors, R3 and R4. This unit, the AD580, was intro­duced in 1974 (24) (Fig. 3).

A similar circuit is used in the LM117 (25), whose output can be set by two external resistors R7 and R8 to any voltage in the range 1.25 V to 37 V (Fig. 4).

The above information was excerpted from an invited pa­per at the IEEE Bipolar Circuits and Technology Meeting, 1990 (26). The paper includes much information on how to



Figure 4. Schematic diagram, LM117 (simplified).

engineer a bandgap reference badly, and a little information on how to do it right. The paper is accessible on the World Wide Web at: http://www. national. com/rap/Application/


The bandgap reference was soon introduced into many kinds of voltage regulator ICs, instead of the old Zener refer­ences. Many of these ICs showed improved performance in one aspect or another. But most of these regulators had low accuracy, and will not be considered here. Our study here will concentrate on precision references with much better than 1% accuracy and tempcos much better than 50 X 10—6/°C.

Paul Brokaw at Analog Devices, Wilmington, MA, de­signed a curvature-correction circuit that canceled out the normal quadratic dependence of the bandgap’s output on tem­perature. The temperature drift over a 70°C span was re­duced below 5 X 10—6/°C. These were introduced in the AD581, about 1976. The related US patent (27) showed how to use different types of IC resistors, with different tempcos, to correct for the curvature.

Robert Widlar designed the NSC LM10 IC bandgap refer­ence (28). This circuit had a reference output of 0.200 V, which was easily scalable by external resistors up to 39 V. The basic reference was able to run on a 1.1 V power supply. It included curvature correction (29).

At NSC, Carl Nelson designed a quadratic curvature cor­rection circuit suitable for curvature correction of bandgap references and temperature sensors, as described in a US pa­tent (30). This was first introduced on the LM35 temperature sensor (31). While at LTC, Nelson later designed an improved logarithmic curvature-correction scheme (32). This circuit was introduced in the LT1019 (33).

Derek Bowers at Analog Devices designed a modified type of reference IC circuit that does not rely on the AVbe of transis­tors operated at different current densities. It depends on the offset of the threshold voltages of two junction field-effect transistors (JFETs) made with different implant doses. This IC was introduced in 1997 as the ADR291 (34). It has a typi­cal long-term stability of only 100 X 10—6 per 1000 h at + 150°C. The ADR291 as introduced does not have sufficiently low tempco or gain error to permit such excellent long-term stability to be fully appreciated. But since the feasibility of this principle has been demonstrated, it is foreseeable that ICs with fully optimized performance will be available soon.

Most of these IC references are not usable directly as stan­dards. However, the ones with the best specifications have adequate stability and are suitable for portable standards in small systems. They may need auxiliary circuits for trims, calibration, tempco correction, and so on—just as a Weston standard cell needs an oven to be useful.


For many years, makers of hybrid ICs were able to include chips of different technologies, wired together on a small ce­ramic substrate. For example, the old NSC LH0075 (intro­duced in 1975 but now out of production) included a quad opamp chip, an LM329 IC reference, and several laser — trimmed resistors. The trimming provided advantages of im­proved output accuracy, and a convenient 10.00 V output volt­age level rather than the 6.95 V of the reference itself.

Likewise, modern hybrid IC references such as Thaler Corp.’s Models VRE100/101/102 use chips of different tech­nologies (Zener diodes, precision trimmable resistors, opera­tional amplifiers, and thermistors) to provide the same kinds of advantages, such as output voltages trimmed to ±0.005%, and temperature coefficients better than ±0.6 X 10—6/°C or ±0.3 X 10—6/°C. Output voltages such as +10 V or —10 V dc (or both) to as low as +1.5 V or —1.5 V are available (35).


The LX5600 temperature-sensor IC, designed by Robert Dob — kin at National Semiconductor Corp. (NSC) and introduced in






Tempcos (10—6 °C—’)








Motorola, APD






Motorola, APD






Motorola, APD








” Tolerance ±5%.

1973 (13), had a hidden agenda: in addition to the tempera­ture sensor, this chip had an experimental IC Zener refer­ence. The Zener diode was connected in series with a transis­tor’s base-to-emitter voltage Vbe, and a buffer amplifier was provided. The reference’s actual performance was fairly medi — ocre—the tempco was typically +30 X 10—6/°C, and the long­term stability was stated as 1000 X 10—6 per 1000 h at +85°C. Still, this temperature-sensor IC went into production as a test bed, and the engineers were able to evaluate a large num­ber of the diodes. Best of all, the construction of a tempera­ture sensor on the same chip as the Zener reference made it easy to operate the temperature sensor as a temperature controller, and to make a little oven around the Zener, hold­ing it at a very stable temperature (such as +88°C). It was easy to evaluate a large number of these references, operating at a constant temperature.

This study soon led to the LM129 and LM199 IC refer­ences (14).

The LM129 was an improved, simplified, upgraded version of the LX5600’s reference, with a series resistance better than 1 П, and a tempco typically in the range 10 X 10—6/°C to 60 X 10—6/°C. These ICs could be tested (and graded in produc­tion test) for 50, 20, or 10 X 10—6/°C.

The LM199 was a new design. It used an integrated tem­perature controller on the die, to hold the die temperature at +88°C. It was housed in a four-lead TO-46 package (similar to a low-profile TO-18). A small plastic thermal shield over the package was used to minimize the power needed to hold the whole IC at that temperature. Under these conditions, the LM199’s reference could show a usable tempco better than 2 X 10—6/°C, 1 X 10—6/°C, or even j X 10—6/°C, selected and tested, over a temperature range from —55° to +85°C. Of course, this temperature-controlled IC did require a signifi­cant amount of power for the heater (typically 260 mW at 25°C, and even more at low ambient temperatures) to hold that +88°C temperature. But this was an acceptable require­ment in many systems.

The temperature sensitivity of any temperature-stabilized circuit depends at least as much on the layout of the heat — sensitive components, and on the gradients caused by the heater, as on the tempco of the circuit. Thus the LM199’s good tempco is related to good die layout.

Further disadvantages of the LM199 were its tolerance (±2%) and the fact that its nominal voltage (6.95 V) was not as convenient as 5.000 V or 10.000 V or even 6.20 V. And unless a charge pump was added, the LM129 or LM199 could not run on 5 V—it needed at least 8 V, at 1 mA for the refer­ence and at 50 mA for the heater. These disadvantages led to efforts to develop improved circuits that avoided some of these drawbacks.

The other significant advantage of the LM129 (and LM199) resided in its buried (subsurface) Zener diode. In most Zener references the breakdown occurred at the surface, where the concentration of impurities (doping) was maximum. Thus, sur­face contamination (even with high-quality planar processing) and electron charging of the oxide caused some degradation of noise and of long-term stability. The invention by Carl Nelson and Robert Dobkin of shallow diffusion layers with decreased concentrations at the surface caused the Zener breakdown to occur about a micron below the surface of the IC, where it is im­mune to surface conditions. This allowed superior consistency of low noise and better long-term stability.

Extensive testing of large numbers of LM199s showed that a large fraction of the units exhibited reference stability con­sistently better than 10 X 10—6 or 5 X 10—6 per 1000 h, when sampled once a week. (However, some units were consistently worse than 20 X 10—6 per 1000 h.) The units that tested better than 20 X 10—6 per 1000 h were designated LM199AH-20 and LM299AH-20, and were used in many precision systems as stable references. Also, they were popular in high-resolution DVMs. The LM199 is still the only successful temperature — stabilized IC in the industry.

Several good selected LM299AH references were evaluated by the National Bureau of Standards (NBS, now the NIST). They found that the long-term drift tendency was about —1 X 10—6 per 1000 h, a fairly consistent drift, presumably re­lated to operation at the die temperature of +88°C.

Other researchers found that if one group of LM299AHs were kept around room temperature, with their heaters off, and another group allowed to run at their normal tempera­ture of +88°C, the long-term drift trend of the units at room temperature was considerably lower than that of the warm units. The room-temperature units could be heated up to their normal +88°C on a specified schedule, perhaps one day per 3 months, and used to calibrate out the long-term drifts of the units kept at +88°C.

The use of buried Zener diodes has spread to other ICs. The LT1021 (15), designed by Carl Nelson at Linear Technol­ogy Corp. (LTC), used a buried Zener diode with temperature — compensating circuits and after-assembly trims to achieve 2 X 10—6/°C, without any heater. The LM169 (16) was engi­neered by Robert Pease at NSC to do likewise.

The LTZ1000 is a buried Zener designed by Robert Dobkin of LTC for laboratory standard use. All resistors are off chip, so that high-stability resistors can be utilized. Its on-chip heater can be activated for thermal stabilization. The die at­tach uses bubble material for high thermal impedance, as high as 400°C/W. Long-term stability approaching 1 X 10—6 is claimed (17).

The Analog Devices AD534 Multiplier IC was designed by Barrie Gilbert (18) using a buried Zener diode. The Analog Devices AD561 was a complete digital-to-analog converter (DAC) IC designed by Peter Holloway and Paul Brokaw, uti-

lizing a buried Zener diode and tempco trim circuits to achieve a gain tempco better than 10 X 10—6/°C (19).

Further research and development into circuits with bur­ied Zener diodes has waned, due to the improvements in bandgap references and to the concentration of research on low-voltage circuits and on CMOS technology, which pre­cludes the use of buried Zener diodes.

The design of a good reference on a large CMOS chip is not trivial. In many cases, a mediocre on-chip bandgap reference is adequate for a system on a chip. If a superior reference is needed, an external (off-chip) reference is often added. This can often provide cost, performance, and yield advantages.


In the era 1920 to 1960, there was a need for stable references in portable equipment. Gas discharge tubes such as OA2, OB2, and 85A2 were used as references for precision regula­tors. If these tubes and the related amplifiers and resistors were powered up and warmed up and aged properly, good short-term voltage stability in the range 0.01% to 0.1% per 24 h could be observed (9). This was considered adequate for most portable precision equipment. However, the advice in Ref. 9 stated that a stack of two 42 V mercury batteries such as Mallory type RM415 would provide improved stability over the 85A2. The degree of improvement was not stated. Gas discharge tubes are still used in high-voltage equipment, but not much in new designs.


Ever since the invention of the Zener diode in the 1950s, much hope has been engendered for using stable Zener diodes as stable references. Technically, diodes with breakdown be­low 5.4 V are true Zener diodes, which depend on a tunneling mechanism, whereas in diodes that break down above 5.4 V, the mechanism is avalanche breakdown. But the term ‘‘Zener diode’’ is applied to both types, unless there is some reason for distinguishing the mechanism.

Low-noise alloyed Zener diodes have existed since 1955. The Zener diode is made simply by doping a silicon pn junc­tion so heavily that it breaks down at some useful voltage— typically in the range of 3 V to 200 V. Some Zener diodes have displayed good stability, and others have not. Most Zener di­odes above 5.4 V have an inherent finite, positive, fairly con­stant tempco. To provide useful performance with low tempco, one or more forward diodes are then connected in series with the Zener. The forward diode chip is typically mounted right against the Zener chip, inside the conventional DO-35 glass diode package. The Vf of the forward diode (about —2 mV/°C) is used to cancel out the tempco of the Zener diode. The re­sulting reference voltage has, under favorable conditions, sta­bility rivaling that of inexpensive standard cells. These refer­ences have the advantage of portability. They can operate (with degraded accuracy and stability) over temperature ranges where standard cells are impractical. Much effort has been put into evaluating Zener diode references. Not all of it has been completely successful. Early manufacturers of refer­ence-type Zener diodes included Solitron, Hughes, PSI, and Motorola.

Alternatively, the Ref-Amp (10), invented by General Elec­tric Corp., utilized an npn transistor with its emitter con­nected to the cathode of a low-noise alloy Zener diode chip, acting both as a temperature compensator and as a gain stage. The TO-5 package provided low stress, low hysteresis, and good stability.

The author has evaluated a 6.2 V Zener reference from a Minuteman I nose-cone guidance system. It was hoped that this reference from about 1960 might exhibit superior stabil­ity. Actually, when it was set up in a precision bias circuit and aged, it had a tendency to drift not much better than 10 X 10—6 or 20 X 10—6 per week, somewhat inferior to modern Zener references.

When early integrated circuit references were built, they were evaluated and compared with the best reference Zener diodes. Soon a serious problem was noted: the glass-packaged Zener diodes had a typical thermal hysteresis of 200 X 10—6 to 600 X 10—6 when cycled over a 100°C loop. This is mani­fested when a Zener diode is cycled cold and then hot and back to room temperature; its voltage will be different than when it is cycled hot and then cold and back to room tempera­ture. Integrated circuits assembled in hermetic packages were observed to have typically 5 or 10 times better hystere­sis, due to the smaller stress (and smaller change of stress) on the die. This is a characteristic of references that is not usually documented or even mentioned. These reference Ze — ner diodes also showed some sensitivity to stress on their leads.

The best, most stable Zener references are often packaged with batteries and heaters in a portable temperature-con­trolled environment, just as saturated standard cells are. These instruments can be used as portable transfer stan­dards. However they have not taken over the task from stan­dard cells entirely, because their stability is not always as good. If the power is ever turned off, these Zener references sometimes exhibit a shift as large as 5 ppm—considerably bigger than that of standard cells—and not always reversible.

A study of low-tempco Zener references available in the

1972 era (11) showed at least 120 different JEDEC-registered part numbers, rated from —25° to +85°C, plus 100 A-grade versions rated for the military temperature range of —55° to +125°C. Many of them were for odd voltages (6.4 V, 8.4 V, 8.5 V, 9.3 V, 9.4 V, 11.7 V, 12.8 V, 19.2 V, 37.0 V, 37.2 V, etc.), at odd currents (choice of 0.5 mA, 1.0 mA, 2.0 mA, or 4.0 mA of bias current). However, as of this writing, almost all of these parts have been obsoleted or discontinued. A small number of popular, commercially viable reference-grade Zeners are still available, as listed in Table 1 (12).