ADC Reference Noise and Layout Design

Every conversion an ADC performs is a comparison against its reference, so the reference is as much a part of the measurement as the input itself. A stable input with a noisy reference produces a noisy result, and a reference that takes too long to settle produces a reading that belongs to the previous channel. Getting the reference right is often the difference between a converter that meets its specification and one that does not.

Why the Reference Limits Accuracy

The converter reports the ratio of the input to the reference, so an error in the reference appears directly in the result. If the reference is one percent high, every reading is one percent low, and if it drifts with temperature, the readings drift with it. An ADC reference that is specified to a fraction of a percent over temperature is a different component from a general purpose regulator output, even when the voltage is the same.

Noise works the same way. The input range is divided into steps, and any variation in the reference moves every step by the same amount in absolute terms but not in the code value. A reference with ten microvolts of noise in the bandwidth of the converter therefore limits the achievable resolution no matter how many bits the converter claims.

The reference also carries the load current of the converter input, which changes as the converter samples. That current is small but fast, and the reference has to supply it without moving. The combination of a quiet source and a local capacitor that can supply the sampling transient is what keeps the reference voltage constant during the conversion.

Types of Voltage Reference

A band gap reference relies on the difference between two base emitter voltages and is available with initial accuracy from a fraction of a percent down to a few parts per million. It is the usual choice for general instrumentation, and its temperature coefficient is the specification that matters most in a measurement channel.

A buried zener reference offers lower drift and lower long term instability at the cost of a higher supply voltage and more power. It is used where the measurement has to be stable over years rather than hours, such as in calibration equipment, and its low frequency noise is generally better than a band gap device of the same accuracy class.

A filtered regulator output can serve as a reference in a low resolution system, but it is a compromise. The output drifts with load, its temperature coefficient is that of the regulator rather than of a reference, and the noise is specified at a different bandwidth. Where a converter claims more than ten bits, a proper reference is usually cheaper than the engineering effort needed to make a regulator behave like one.

Reference Noise and the Noise Floor

noise floor of a measurement channel is the sum of the converter noise, the input amplifier noise and the reference noise, referred to the input. Reducing one of the three below the others produces no improvement, so the reference should be chosen to be comparable to the other contributions rather than as low as possible at any cost.

Low frequency noise is the component that matters in most instrumentation, because the bandwidth of the measurement is small and high frequency noise averages out. The 0.1 to 10 hertz peak to peak noise specification captures that region, and it is often the number that explains why a channel with excellent resolution still moves in the last few digits.

Filtering the reference reduces the noise that reaches the converter, but it also slows the response to a load change. A simple RC filter with a corner well below the conversion rate is usually enough, provided the capacitor is large compared with the sampling charge and is placed at the converter rather than at the reference. Where the reference has to change quickly, as in a ratiometric scheme, the filter has to be designed around that requirement instead.

Decoupling and Bypass Layout

decoupling the reference means more than one capacitor. The capacitor at the converter pin supplies the sampling transient, and its value is set by the charge that the converter draws and by the voltage change that can be tolerated. A capacitor that is too small allows the reference to dip on every sample, which appears as gain error that depends on the input voltage.

The capacitor type matters as much as the value. A ceramic capacitor with a stable dielectric holds its value over temperature and voltage, while a high dielectric constant part can lose half of its nominal capacitance at the working voltage. The physical size also affects the equivalent series inductance, which determines how quickly the capacitor can deliver the transient.

Place the capacitor directly between the reference pin and the converter ground pin, with the shortest possible connections, and route the reference trace away from any switching node. Where the reference serves more than one converter, each should have its own local capacitor and the trace should be routed as a bus with a defined order rather than as a star from a single point.

Reference Settling Time and Multiplexing

A reference that is shared between channels has to recover between conversions, and its settling time becomes part of the channel switching budget. The sampling capacitor of the converter presents a step load on every acquisition, and the reference has to return to its nominal value before the conversion is complete. If it does not, the error appears as crosstalk between adjacent channels.

The time constant that governs this behaviour is the product of the reference output impedance and the total capacitance on the node. A low impedance reference with a large local capacitor recovers quickly, while a higher impedance source with the minimum capacitor may need several microseconds. Comparing that figure with the acquisition time of the converter shows immediately whether the design has margin.

Where the converter multiplexes many channels at high speed, the reference is not the only node that has to settle. The input multiplexer and the driving amplifier have their own settling requirements, and the slowest of the three sets the achievable channel rate. Measuring the crosstalk between a channel connected to a full scale input and one connected to zero shows which node is limiting.

External multiplexing brings the same problem to the reference if the reference is switched. A switched reference needs a buffer with a fast settling time and a low output impedance, and the switch itself contributes charge injection that appears as an offset. Where possible, keep the reference fixed and switch the input, which removes the reference from the settling budget entirely.

Layout and Grounding of the Reference

The reference ground pin and the converter ground pin should be connected by a short, direct path, and that path should not carry any other current. The reference current returns to the source through that connection, and if a digital return current shares the copper, the resulting voltage appears as a reference error that changes with the data pattern.

Keep the reference away from switching regulators, clock lines and digital buses. The reference is a quiet node with a high impedance at some frequencies, so it couples easily. Where the board must route a clock past the reference, keep the distance generous and consider a ground trace between them, stitched to the plane at both ends.

Where the measurement is one part of a larger mixed signal board, the reference belongs in the analogue partition, close to the converter and away from the digital side. The organisation of those areas, including how the two grounds should meet, is described in mixed signal board design, and it applies to a single reference pin as much as to a whole front end.

Protecting the reference area from contamination is worth considering in humid environments, because surface leakage on a high impedance node produces a slow drift that is difficult to trace. The techniques in conformal coating keep the surface resistance high, and the general fabrication notes in PCB design and fabrication cover the process decisions that go with them.

Measuring Reference Performance

Measure the reference with the converter running, not with the board idle. The sampling current is part of the load, and a reference that looks perfect at no load may move by tens of microvolts once conversions begin. A scope with a low capacitance probe on the reference pin, triggered by the conversion start, shows the transient directly.

Measure the noise over the bandwidth of interest rather than with a wideband meter. A meter that reads to hundreds of kilohertz includes components that the converter averages out, and the resulting number is pessimistic. A low pass filter in front of the meter, or a Fast Fourier transform of a recorded time series, gives the figure that applies to the channel.

Finally, check the reference against a second unit at a stable temperature. Comparing two channels that share the same input but different references separates the converter contribution from the reference contribution, and it is the quickest way to establish which of the two is limiting a channel that does not meet its specification on paper.

FAQ

Can I use the supply as the reference for a ratiometric measurement? Yes, and it is often the better choice when the sensor is excited from the same rail, because the supply variation cancels. It is not suitable when the measurement has to be absolute.

How large should the reference capacitor be? Work from the sampling charge of the converter and the tolerable voltage change, then round up. Most converters specify a minimum value and a maximum equivalent series resistance in the datasheet.

Does a lower noise reference always improve the measurement? No. Once the reference noise is below the noise of the amplifier and the converter, reducing it further changes nothing. Balance the three contributions instead.