RESOLUTION
Bits count codes, not truth.
An ideal N-bit ADC divides its input range into 2^N nominal code regions. For a unipolar converter spanning reference voltage Vref, the ideal code width is approximately Vref / 2^N.
That is resolution. Real measurements also include offset error, gain error, integral and differential nonlinearity, reference uncertainty, input noise, quantization, source error, temperature drift, and digital processing effects. A 12-bit output field is therefore not a certificate that the result is accurate to one part in 4096.
REFERENCE
The ADC measures a ratio against its reference.
For many converter architectures, the result depends directly on Vref. Noise or drift on that reference becomes measurement noise or gain error. Using a noisy supply rail as the reference can be acceptable for ratiometric sensors that move with the same rail, but disastrous when measuring an independent absolute voltage.
Reference selection, decoupling, source impedance, thermal drift, load behavior, and routing therefore belong in the measurement model. If the reference moves one percent, the conversion scale can move with it.
SAMPLE-AND-HOLD
The converter input often has to charge an internal capacitor.
Successive-approximation ADCs commonly connect the external signal to an internal sample capacitor during an acquisition interval. The source must settle that node close enough to its final value before conversion begins.
A high source resistance, large external RC filter, multiplexer switch resistance, or short acquisition time can prevent full settling. The resulting error may depend on the previous channel voltage, which is why multiplexed ADC systems can appear haunted when adjacent channels influence one another.
Longer acquisition time, a lower source impedance, or an input buffer can solve the mechanism. Increasing digital averaging cannot correct a deterministic settling bias.
GROUND
The converter measures between pins, not between concepts.
An analog source and ADC only share “ground” to the extent that the physical reference path between them remains quiet. Load current flowing through common impedance creates reference error. Fast digital return currents can couple into the analog path. Cable shields and external grounds can import common-mode voltage.
Follow current through the real board. Place reference and decoupling components according to device guidance. Keep high-current switching loops away from sensitive nodes. Measure the analog input and reference at the converter pins when the digital load is active.
INPUT PROTECTION
Protection components become part of the signal path.
Series resistors, clamp diodes, RC filters, TVS devices, multiplexers, and isolation components can add leakage, capacitance, resistance, distortion, and temperature dependence. Those effects may be negligible or dominant depending on sensor impedance and required accuracy.
Design protection against the fault environment, then include its normal-operation behavior in the error budget. A resistor selected only for fault current may make the sample-and-hold settling problem worse.
NOISE + ENOB
Effective number of bits measures useful dynamic performance, not register width.
Noise causes repeated conversions of a constant input to spread across codes. Dynamic converter specifications often summarize the combined effect of noise and distortion through signal-to-noise-and-distortion ratio and effective number of bits, or ENOB.
ENOB is useful because it acknowledges that a nominal 16-bit ADC can have substantially fewer useful bits at a given input frequency and sampling condition. For slow sensor measurements, histogram noise, standard deviation, offset, and gain calibration may be more directly useful than one headline ENOB number.
OVERSAMPLING
Averaging can reduce noise. It cannot average away bias.
If noise is sufficiently uncorrelated, averaging multiple samples reduces random variation. Oversampling plus decimation can sometimes improve effective resolution when adequate noise or dithering lets the signal explore code boundaries.
But averaging does not repair a bad reference, insufficient acquisition time, offset error, gain error, saturation, or a ground shift that moves every sample in the same direction. Know which error term you are attacking before adding more samples.
CALIBRATION
Calibration belongs to the whole measurement chain.
Single-point offset correction can remove a fixed zero error. Two-point or multi-point calibration can estimate gain and nonlinear behavior. Temperature compensation may be needed when sensors, references, resistors, amplifiers, or the converter drift materially over the operating range.
Calibration values need provenance too: which hardware unit, which reference instrument, which conditions, which firmware algorithm, and when. A calibration constant copied from another unit is not calibration. It is a guess wearing decimal places.
FIELD METHOD
Test the measurement chain as a measurement chain.
1. Drive known inputs from a characterized source. Sweep across the range.
2. Measure Vref at the ADC pins. Include dynamic system load.
3. Vary source impedance and acquisition time. Look for settling dependence.
4. Repeat one channel after large changes on another. Detect multiplexer memory.
5. Histogram a constant input. Separate random spread from systematic offset.
6. Test temperature. Track reference, sensor, and conversion drift.
7. Compare against an external calibrated instrument. The firmware code is not its own reference standard.
BOTTOM LINE
The ADC is one component in an analog instrument.
Useful accuracy comes from reference integrity, source settling, grounding, protection, noise control, calibration, and an error budget tied to the actual mission.
Twelve bits of output can still contain eight bits of information and four bits of confidence theater.
SOURCE TRAIL
Technical starting points.
Analog Devices — Understanding ADC specifications
Texas Instruments — Understanding Data Converters