What this page solves (linearity errors in real systems)

Static DAC accuracy fails in predictable patterns. This page maps common field symptoms (wrong setpoint, uneven steps, poor repeatability) to the specific error families that create them: gain/offset (absolute accuracy), DNL/monotonicity (code-to-code consistency), INL (curve shape), and drift (time/temperature stability).

The goal is not to memorize definitions. The goal is to assign ownership and pick the first measurements that can separate “DAC core” effects from reference/load/board effects, without jumping straight into calibration and accidentally masking the real problem.

Static error taxonomy: offset, gain, INL, DNL, monotonic

Clear language prevents bad decisions. The same DAC can look “great” or “bad” depending on definitions, reference lines, excluded codes, and measurement uncertainty. The taxonomy below is written in engineering terms: what the error means, how it is expressed, and the most common pitfall.

How datasheets specify INL/DNL (endpoint vs best-fit, codes excluded)

INL and DNL numbers are only meaningful when the reference line, code range, and measurement conditions match the intended system definition. The same DAC can report very different INL depending on whether it is referenced to endpoint, best-fit, or a segmented-fit line. “Good-looking” typical plots can also hide excluded codes, restricted code ranges, or temperature limits that matter in precision setpoint systems.

This section provides a practical reading checklist that prevents three common failures: budgeting errors (wrong accuracy roll-up), guarantee errors (assuming monotonicity or worst-case behavior without proof), and reproducibility errors (lab results that do not match the vendor’s test method).

Where nonlinearity comes from (architecture-agnostic root causes)

Nonlinearity is not a single mechanism. It is the visible output of physical mismatches, parasitics, coupling paths, and thermal gradients. Understanding root causes helps predict risk during selection and prevents wasted calibration effort on problems that move with load, timing, or temperature.

The most useful level of detail is: root cause → which error terms it tends to create → one fast observation that can separate it from other causes. The cards below follow that exact pattern.

Monotonicity assurance: what it really means and how to guarantee it

Monotonicity is a guarantee about behavior, not a one-time observation. It means that increasing code never produces a decreasing output within a stated window: temperature range, warm-up state, load range, update conditions, and measurement uncertainty. Without these conditions, “monotonic” becomes an ambiguous claim that can be broken by noise, drift, or incomplete settling.

DNL is closely related because it describes code-to-code step width. A common rule of thumb is DNL > −1 LSB, but the practical guarantee still needs margin: measurement noise and temperature drift can hide a true negative step or create an apparent one. For multi-channel systems, “system monotonicity” must also be defined: monotonic per channel is not the same as monotonic for a derived function such as comparisons, thresholds, or channel-to-channel subtraction.

Temperature drift: separating offset drift, gain drift, and INL shape drift

Drift is not one number. It is a combination of three behaviors that look similar in a single-point check but behave very differently across the code range. Separating them early prevents the wrong calibration strategy and improves how recalibration intervals are planned.

Offset drift shifts the whole curve, gain drift changes the span, and shape drift changes the curvature of INL. Offset and gain are often corrected with one-time 1-point or 2-point calibration. Shape drift is harder because it moves the residual error across codes, often with thermal gradients and self-heating.

Long-term drift & aging: what moves, how fast, and how to plan recalibration

Long-term stability is defined by which error terms move over time and whether that motion can be controlled by calibration or must be contained by design. Aging mechanisms usually appear as slow changes in offset and gain, while some environments can also change INL shape and invalidate a previously good multi-point correction.

The most reliable planning language is a recalibration window: define an allowed drift threshold in system units, estimate drift rate with margin, and choose a recalibration interval that balances maintenance cost and accuracy risk. Production planning also must separate batch spread (unit-to-unit variation at time zero) from time drift (change of a given unit over time).

Error budgeting: rolling up INL/DNL + gain/offset into end accuracy

Budgeting turns isolated specifications into system performance. End accuracy is not determined by INL or gain alone; it is the combined effect of DAC static errors, reference accuracy and drift, output driver regulation, load interaction, and the measurement chain used for calibration and verification.

A practical budget should produce three outputs: absolute worst-case accuracy, short-term repeatability, and post-drift accuracy. Offset and gain are often reducible by calibration. INL shape and configuration-dependent errors usually remain and must be controlled by component choice and system conditions.

Measurement & test: how to measure INL/DNL correctly (and not fool yourself)

Measuring INL/DNL is a system problem: the result depends on the stimulus, the settling and sampling rule, the reference line definition, the measurement chain uncertainty, and the thermal state during the sweep. A “good looking” curve can still be wrong if noise, timing, or baseline nonlinearity is being fitted into the result.

A reproducible approach starts with a stable test chain and ends with statistics: repeat readings at the same code to quantify noise, compare forward vs reverse sweeps to detect code-order coupling, and keep the reference line definition consistent (endpoint vs best-fit). When results depend on sweep style, the measurement chain is writing artifacts into the data.

Calibration strategies: what calibration can and cannot fix

Calibration is not a magic eraser. It reduces error terms that are low-dimensional and repeatable under the stated condition window. Offset and gain usually calibrate well. INL shape and configuration-dependent errors often remain unless multi-point compensation stays stable across temperature, time, load, and operating modes.

Calibration also injects new limits: measurement uncertainty can be written into coefficients, coefficients can drift, and interpolation and storage quantization can create residual errors. When noise or drift dominates, more calibration points can simply fit the measurement chain rather than the DAC.

Recommended topics you might also need

• Error Budgeting
• Matching & Calibration
• Reference & Buffering
• Code-to-Code Consistency / Major Carry
• Supply & Grounding

FAQs: INL/DNL, gain/offset, drift, measurement, and calibration

Short, actionable answers for static linearity work: definitions, risk judgement, drift planning, measurement repeatability, and calibration boundaries.