DC Accuracy for INAs: Offset, Drift, Gain Error
← Back to:Instrumentation Amplifiers (INA)
DC accuracy in an instrumentation amplifier is not a slogan—it is a four-term budget: offset, offset drift, gain error, and gain drift, each tied to real conditions (gain, temperature span, warm-up, load/headroom). This page shows how to translate datasheet limits into an input-referred error budget, choose the minimum calibration that actually holds, and verify results with pass/fail criteria that survive real wiring, thermal gradients, and production test.
What “DC Accuracy” Means for an INA
DC accuracy is not a single number. It is the sum of four budgetable error terms—two additive (offset-related) and two proportional (gain-related)—each valid only under its stated test conditions and time/temperature window.
The 4 DC-accuracy terms
- Offset (VOS): input-referred DC error at zero differential input. Dominates small-signal and high-gain measurements.
- Offset drift (dVOS/dT or max ΔVOS): change of offset over temperature/time windows. Sets baseline stability across warm-up and ambient sweeps.
- Gain error (ΔG/G): proportional scale error versus nominal gain. Becomes dominant at larger input levels (percentage-of-reading).
- Gain drift (dG/dT or max ΔG/G): gain change across temperature/time. Controls cross-season consistency and recalibration interval.
Conditions and time windows are part of the spec
Any quoted DC-accuracy number must be read together with its conditions (gain setting, common-mode level, load/output swing, supply, source impedance) and its window (power-up warm-up, ambient temperature span, long-term aging). This page budgets the four DC terms; noise, CMRR/PSRR, and EMI protection are treated on their dedicated pages.
Boundary rule: random short-term fluctuations that average down belong to noise; slow, directional changes tied to time/temperature belong to drift.
Not covered here: 0.1–10 Hz and wideband noise (noise metrics page), CMRR/PSRR vs frequency (CMRR/PSRR page), and EMI/ESD protection networks (protection page).
A Practical Error Model: Translate Specs to Input-Equivalent Error
The fastest path to an honest accuracy budget is to convert every datasheet term to input-equivalent error, separate additive vs proportional behavior, and compute each term over an explicit time/temperature window.
Normalize to input-referred error
Budgeting at the input removes ambiguity: additive terms stay constant in volts, while proportional terms scale with the measured signal. Output-domain budgets can be derived by multiplying by gain VOUT,error = G · VIN,error.
Use rules that prevent “paper accuracy”
- Design sign-off: use max over temperature (and stated conditions) plus guardband for wiring, assembly stress, and long-term drift.
- Lab correlation: typical curves can guide expectations, but results must be reconciled to max limits with matched test conditions.
- Windowing: drift must be computed over explicit windows, not averaged away like noise.
Window cheat-sheet: Warm-up (0 → stable), Temperature span (Tmin → Tmax), Life (months/years). Each window needs its own ΔVOS and ΔG/G.
Next steps (covered in later sections): how each term is specified and trapped in datasheets (offset vs drift vs gain), how calibration removes structured terms, and how verification avoids measurement artifacts.
Offset: Where It Comes From and How to Read It in a Datasheet
Offset is not “one number.” It is an input-referred DC bias that can shift with gain setting, common-mode level, output swing/load, and supply headroom—plus PCB-induced leakage paths that look like offset.
Offset contributors that matter in real systems
- Core mismatch: intrinsic input-stage mismatch and internal bias/trim residuals. This is the “part-to-part” baseline term that selection and calibration address.
- Condition-dependent shift: changes driven by VCM, output swing near rails, load current, or supply headroom. The observed offset can change even if the sensor input is unchanged.
- Leakage-induced offset: input protection, contamination, humidity, or high-impedance sources create tiny currents that translate into DC error. This often appears as “mysterious drift” when wiring is touched or humidity changes.
- Residual output stage effects: finite output swing, load-dependent behavior, and recovery from saturation can bias the measured DC value if the operating point is too close to limits.
Datasheet reading: 4 questions that prevent surprises
1) Gain setting? Offset is specified at a particular gain (or across gains). A gain change can shift the reported offset and the available headroom.
2) VCM and output swing? Offset must be interpreted at the stated common-mode and output level. Near-rail operation can introduce nonlinear DC errors that look like offset.
3) Load (RL) and output drive? Heavy loads or capacitive loads can change the operating point or recovery behavior; the observed DC value may shift under real loading.
4) Typical vs max? Accuracy sign-off must use max / guaranteed limits (often “max over temperature”). Typical is useful for expectation-setting, not for guaranteed baseline accuracy.
Boundary note: if offset changes with humidity, cable touch, or cleaning, treat it as leakage-induced until proven otherwise (detailed modeling belongs to “Input Clamp & Leakage Budgeting”).
Not covered here: detailed leakage modeling and protection network leakage tradeoffs (see “Input Clamp & Leakage Budgeting”).
Offset Drift: Temperature, Warm-Up, and Stress-Driven Drift
“Drift” is not one phenomenon. Practical DC stability requires separating temperature drift, power-up warm-up drift, and stress/aging drift—each defined over its own window and verified with a clear pass criterion.
Three drift classes (and what each means)
Temperature drift: systematic offset change with temperature. Typically specified as µV/°C or as max ΔVOS over temperature. It sets the across-ambient baseline accuracy.
Warm-up drift: offset movement after power-up as self-heating and local thermal gradients settle. Often described by time-to-stable or typical curves; if not specified, it must be characterized for the actual board and airflow.
Stress / aging drift: slow shifts from package stress, PCB bending, humidity/contamination, and long-term aging. Long-term specs are often incomplete; robust designs add guardband and validate with sampling plans.
Read drift specs only with windows and conditions
- No window, no number: drift must state a measurement window (warm-up, temperature span, life) and whether thermal equilibrium is required.
- Self-heating matters: confirm whether the drift spec includes internal self-heating or assumes a stabilized thermal environment.
- Typical vs guaranteed: typical curves guide expectations and test planning; guaranteed limits drive sign-off and pass/fail criteria.
- Pass criteria should be a rate or band: use a stability rate (e.g., “|dVOS/dt| < X”) or a settled band (e.g., “within ±X µV”), not just a fixed waiting time.
Boundary note: thermal-gradient control and guard-ring/leakage controls are implementation topics; this section only classifies drift so later checklists can apply the right mitigations.
Not covered here: detailed thermal-gradient layout tactics and guard-ring implementation details (these appear in the engineering checklist and design hooks sections).
Gain Error: What Sets It (Beyond One Resistor) and What It Breaks
Gain error is a scale error (ΔG/G). It is set by internal gain-setting accuracy, any external gain network, and operating-point limits (headroom, load, swing). As input level increases, gain error becomes the dominant DC-accuracy term.
A budget view: convert gain error into input-equivalent error
Treat gain error as a proportional term. For an input differential signal VIN, the input-equivalent error contribution is:
ΔVIN ≈ VIN · (ΔG/G)
This is why gain error is often “quiet” at tiny signals (offset dominates) but becomes decisive near the top of the measurement range (percentage-of-reading).
What sets gain error (beyond one resistor)
- Internal gain-setting accuracy: the INA’s internal resistor network and trim define the baseline “how accurate is G” term, even with perfect external parts.
- External gain network tolerance (if used): any external resistor(s) add tolerance and temperature dependence. If the system relies on ratio accuracy, the external network must be treated as part of ΔG/G.
- Operating-point induced effective gain shift: near-rail output swing, heavy load current, or limited supply headroom can distort the transfer slope and appear as gain error.
- Common-mode and headroom limits: when VCM approaches input-range limits, the slope can compress; the result is a scale error even if offset looks acceptable at small signals.
What it breaks (practical consequences)
- Full-scale accuracy and thresholds: a small ΔG/G produces large absolute error at high readings, shifting trip points and calibration targets.
- Multi-range consistency: when gain settings change, each gain can carry a different ΔG/G. Without per-gain calibration, range switching produces “scale jumps.”
- Channel-to-channel agreement at large signals: small differences in ΔG/G become large mismatches when readings are near the top of the range.
Datasheet focus: always read gain error versus gain setting, temperature, and supply/swing/load conditions. Sign-off uses max/guaranteed limits with stated conditions.
Boundary note: settling behavior, phase margin, and ADC drive stability are not treated as gain error here. If the symptom depends strongly on step response, RC values, or ringing, it belongs to I/O traits and anti-alias filtering sections.
Gain Drift: Tempco, Aging, and Why Two-Point Calibration Exists
Offset-only calibration removes an additive bias, but it does not freeze the scale. Gain drift moves the slope across temperature and time, which is why two-point calibration exists for accuracy that must hold across ambient changes.
Budget view: treat gain drift as a windowed scale change
Gain drift must be computed over explicit windows (temperature span and life). For a given input level:
ΔVIN ≈ VIN · (ΔG/G)window
This makes gain drift especially impactful at high readings and in wide-temperature or long-interval applications.
What drives gain drift
- Resistor network tempco: the internal gain-setting network changes with temperature, shifting the effective gain (slope).
- Reference/bias coupling: internal bias and reference dependencies can translate temperature behavior into a scale shift, depending on architecture.
- Stress and long-term drift: package stress, PCB strain, and aging mechanisms can cause slow slope movement over months or years.
Typical specification formats: ppm/°C (slope per °C) or %FS over temperature (windowed worst-case). Always bind the number to its temperature span and conditions.
When gain drift becomes the dominant problem
- Wide temperature span (industrial/outdoor) and accuracy must hold without frequent recalibration.
- High reading levels (near full-scale), where proportional errors map into large absolute error.
- Cross-season or long-interval consistency is required (calibration cannot be repeated often).
Boundary note: multi-point temperature mapping, LUT design, coefficient stability, and overfitting controls belong to the “Tempco & Calibration Strategy” section. This section only provides selection logic.
Budgeting Workflow: Build a Baseline Accuracy Budget That Survives Reality
A baseline DC-accuracy budget must be windowed, condition-bound, and worst-case driven. It survives reality when every term is expressed in one unit, mapped to a measurement window, allocated between “must be guaranteed” and “can be calibrated,” and protected by guardband.
Step-by-step workflow (engineering SOP)
Step 1 — Define the target in one unit. Choose input-referred (µV, ppm of reading, or %FS) or output-referred—then stick to it. Mixed units create hidden budget gaps.
Step 2 — Define windows. Budget drift only with explicit windows: temperature span (Tmin→Tmax), warm-up (power-on→stable), and life (months/years).
Step 3 — Pull MAX terms with stated conditions. Use max/guaranteed limits, not typical. Bind every number to gain setting, VCM, VS, load, and output swing/headroom.
Step 4 — Translate each term into the same budget language. Use additive vs proportional forms so terms do not get misapplied:
Additive: offset, windowed offset drift
Proportional: gain error (ΔG/G), windowed gain drift
Step 5 — Allocate calibration vs guarantees. Decide which terms are removed by calibration (structured, repeatable) and which must be guaranteed by parts + implementation (floors, non-repeatable effects).
Step 6 — Add guardband for reality. Guardband must cover uncontrolled variation: airflow/thermal gradients, assembly strain, contamination/humidity sensitivity, and long-term drift when vendor data is incomplete.
Step 7 — Close the loop with pass/fail hooks. The final budget must map to a test method: stabilization criteria, temperature dwell rules, and the input levels used for verification.
Failure modes that break budgets in the field
- Averaging drift as if it were noise: windowed drift does not disappear by longer averaging.
- Using typical as a guarantee: typical may match one lab board, but it does not sign off worst-case accuracy.
- Ignoring warm-up: the first minutes after power-up can dominate DC error if stabilization criteria are undefined.
- Ignoring operating point constraints: headroom, swing, and load can change effective gain and offset behavior.
- Missing guardband: contamination, assembly strain, and long-term behavior are rarely “zero.”
A budget survives reality only when every number is tied to a window and conditions, and when uncontrolled variation is explicitly covered by guardband.
Calibration Strategy: What It Can Fix, What It Cannot, and When It Backfires
Calibration is not a universal accuracy upgrade. It removes structured, repeatable error terms under stable conditions. It cannot remove noise floors or non-repeatable drift, and it can backfire when measurement uncertainty or environment changes are written into coefficients.
What calibration can fix vs cannot fix
Can fix (structured & repeatable): static offset, static gain, and repeatable temperature behavior when windows/conditions are stable and measurement uncertainty is well below the target.
Cannot fix (floors & non-repeatable): noise floor, random drift, contamination/leakage-driven shifts, and errors caused by saturation/recovery or operating beyond linear headroom.
Principle: calibration works when the error is stable enough to be measured and re-applied. If the error changes with humidity, touch, or uncontrolled thermal gradients, coefficients will not remain valid.
Strategy selection (actionable, implementation-agnostic)
0-point (no calibration)
Use when target accuracy is loose, temperature span is narrow, and production simplicity dominates.
1-point (offset)
Use when small-signal accuracy is critical and offset dominates, and when the system can reliably create a “zero” condition.
2-point (offset + gain)
Use when readings span a wide fraction of range and proportional errors must be controlled. Two points establish both intercept and slope.
Multi-point / segmented
Use only when temperature span is wide and the error curve is repeatable, and when measurement uncertainty is far below the required improvement.
Boundary note: temperature-sensor placement and thermal-path details are implementation topics. Only the selection principle is stated here: the temperature proxy must represent the error-driving thermal state.
When calibration backfires (common patterns)
- Non-repeatable error: coefficients change with humidity, cable touch, cleaning, or contamination—typical of leakage-driven shifts.
- Uncertainty written into coefficients: stimulus or measurement chain uncertainty is comparable to the target improvement, turning noise into “calibration terms.”
- Unstable conditions: warm-up not settled or temperature dwell too short, so the calibration captures transient thermal gradients rather than a stable behavior.
Practical rule: only calibrate what can be measured with high confidence and re-applied with high stability across the intended window.
Verification: How to Measure Offset/Drift/Gain Correctly Without Lying to Yourself
A measurement is only valid when the test setup’s own errors are smaller than the device term being claimed. Offset, drift, and gain can look “worse than the datasheet” when wiring, thermal gradients, leakage paths, or stimulus uncertainty dominate the result.
Make results comparable to datasheet conditions
Bind the operating point. Record gain setting, input common-mode (VCM), supply (VS), load, and output swing/headroom. Many “bad” results are headroom or loading effects masquerading as DC error.
Define windows. Drift is only meaningful with windows: warm-up (power-on→stable), temperature dwell (per setpoint), and life (days/months). Without windows, “drift” becomes a label, not a number.
Keep the setup below the claimed term. Fixture leakage, thermal EMF, and stimulus uncertainty must be comfortably below the device term being measured; otherwise the setup is being characterized, not the INA.
Measuring offset (shorting without creating new errors)
1) Short at the correct location
Short the true differential input nodes at the front-end (Kelvin short at the connector/fixture). Avoid far-end shorts across long leads where wiring resistance, leakage, and thermal EMF become part of the “offset.”
2) Control thermal EMF
Microvolt-class offsets can be overwhelmed by thermal EMF from dissimilar metals plus temperature gradients. Keep both inputs in the same thermal environment (same airflow, same copper mass, symmetric connections).
Pass criteria (template)
After stabilization, shorted-input reading changes by less than X µV over Y minutes (choose X/Y from the system budget and resolution).
Measuring drift (use slope thresholds, not a fixed wait time)
Stabilization rule: declare “stable” only when the absolute slope falls below a threshold (e.g., < X µV/min). A fixed 2–5 minute wait is not a stability definition.
Temperature sweep structure (practical)
- At each setpoint: wait for slope < threshold → sample → record mean/median and slope.
- Run both directions (up/down) when possible; hysteresis flags stress/thermal-gradient sensitivity.
- Keep airflow consistent; airflow changes are often mistaken as “device drift.”
Pass criteria (template)
Warm-up drift slope < X µV/min after Z minutes, and temperature sweep return difference < X µV across the specified window.
Measuring gain (avoid measuring the source)
Stimulus uncertainty must be smaller than ΔG/G. If the reference source accuracy or stability is comparable to the claimed gain error, the result is dominated by the stimulus.
Practical checks that prevent self-deception
- Use at least two input levels (near-zero and near operating amplitude) to separate intercept vs slope.
- Hold common-mode and headroom consistent; near-rail operation can shift “effective gain.”
- Repeat after changing load/swing if the system will see multiple output conditions.
Boundary note
Full production data schema and automated screening belong in the “Self-Test & Production Test” page. This section focuses only on lab verification without setup-induced bias.
Common traps (symptom → root category)
- Thermal EMF: microvolt steps when touching connectors, or direction-dependent temperature sweeps.
- Leakage: humidity/cleanliness sensitivity, high-Z nodes drifting with time even under “short.”
- Source accuracy: gain changes when changing the stimulus range or instrument.
- Wiring R: gain/offset shifts with cable length or wiring topology.
- Warm-up: large initial drift that disappears after a stabilization slope threshold is met.
- Airflow: slow drift tied to airflow changes rather than device physics.
Engineering Checklist for DC Accuracy (Layout, Leakage, Thermal, Guardband)
DC accuracy is built on symmetry, leakage control, thermal control, and explicit guardband. The checklist below focuses only on actions that protect offset, drift, and gain from real-world board-level effects.
Layout (symmetry + Kelvin + avoid thermal gradients across the input pair)
P0 — Must do
- Route IN+ and IN− as a symmetric pair: same length, same layer, same surroundings.
- Use Kelvin routing where the sensor/bridge or shunt requires true sensing (do not “share” current paths).
- Keep both input nodes in the same thermal environment (avoid one side near heat sources).
P1 — Strongly recommended
- Keep digital switching and self-heating resistors away from the input symmetry region.
- Maintain matched copper/thermal mass around both inputs (vias, pours, nearby parts) to reduce thermal EMF.
Leakage (high-Z nodes, surface paths, cleanliness)
P0 — Must do
- Use guard rings around high-impedance input nodes to intercept surface leakage.
- Define a cleaning strategy; flux residue and contamination often dominate microvolt-level stability.
- Avoid moisture-sensitive surface paths near inputs; maintain spacing and keep sensitive areas dry.
P1 — Strongly recommended
- If conformal coating is used, verify it does not introduce stress/thermal gradients that increase drift.
- Treat protection components as leakage sources; place and validate with the leakage budget.
Thermal (couple what must track, isolate what must not)
Thermal consistency beats absolute temperature. Keep IN+ and IN− on matched thermal paths. Avoid placing one input near heat sources, copper cuts, or airflow edges that create differential gradients.
Where multiple precision blocks exist (INA, reference/excitation, sensor interface), ensure the thermal design does not create “moving gradients” across the measurement path. Temperature proxies must represent the error-driving thermal state.
Guardband + pass criteria (templates for sign-off)
Guardband must cover temperature gradients, assembly stress/board flex, contamination sensitivity, and long-term behavior that is not fully specified. Any un-modeled contributor is treated as non-zero and must be bounded.
Pass criteria examples (replace X with budget-driven values)
- After stabilization, drift slope < X µV/min.
- Temperature sweep return difference < X µV across the specified window.
- Gain shift under representative VCM/load/swing < X ppm of reading.
Application Mapping (DC Accuracy Angle Only)
Different applications amplify different DC error terms. This map ranks which term usually dominates: offset, offset drift, gain error, or gain drift. It intentionally excludes noise, CMRR, filtering, and stability topics.
How to use this map (fast rules)
- Smaller signals → offset and drift usually dominate.
- Larger signals / wide range → gain error and gain drift become more visible (proportional error).
- Wide temperature / long intervals → drift and gain drift weight increases.
Bridge / Weighing
Dominant term: offset + offset drift (small signal).
Why: warm-up and temperature gradients create microvolt-level baseline movement.
Verify hook: stabilized drift slope < X µV/min (X from the system budget).
RTD (low-frequency, long-term)
Dominant term: offset drift + gain drift (wide-temp consistency).
Why: accuracy is judged over seasons; two-point calibration is common, but drift repeatability sets the limit.
Verify hook: temperature sweep return difference < X µV across the defined window.
Bio-potential (ECG / EEG / EMG)
Dominant term: offset + drift (baseline stability).
Why: baseline movement corrupts low-frequency content; noise and CMRR are intentionally not covered here.
Verify hook: baseline slope under the target VCM and supply stays below a budgeted threshold.
Industrial (4–20 mA / ±10 V)
Dominant term: gain error + gain drift (proportional error on large readings).
Why: proportional error grows with reading; calibration often targets slope as much as intercept.
Verify hook: two-level (or multi-level) linearity check under representative load/headroom.
Seismic / Low-frequency monitoring
Dominant term: drift + long-term stability (trend fidelity).
Why: slow trends can be mistaken as real signals; 1/f noise is intentionally excluded here.
Verify hook: long-window slope stability (hours) + sweep hysteresis consistency.
IC Selection Logic (Fields → Risks → Vendor Questions)
The goal is to turn DC-accuracy knowledge into a repeatable buying workflow: spec fields → risk weighting → vendor questions. Priority should be driven by signal level, temperature range, and recalibration interval—not by typical graphs alone.
A) Must-have datasheet fields (DC accuracy only)
Offset: Vos max (with gain/VCM/VS/RL conditions). Record the exact test conditions—numbers without conditions are not comparable.
Offset drift: drift max over temperature (window and method). Separate warm-up behavior from steady-state drift when possible.
Warm-up: time-to-stable or warm-up drift (if not specified, treat as a risk and demand characterization).
Gain error: gain error max across gain settings and output conditions (headroom and load can change effective gain).
Gain drift: ppm/°C or %FS over temperature. Confirm which gain setting(s) the drift applies to.
Long-term drift: if not specified, request distribution or screening capability. Treat “unknown” as non-zero in guardband.
B) Risk weighting (system conditions → spec priority)
Small signal (bridge / microvolt-level): prioritize Vos max and drift max.
Wide temperature or long recalibration interval: increase weight on drift and gain drift.
Large swing / wide range (±10 V, scaled DAQ): increase weight on gain error and gain drift.
C) Vendor questions (copy/paste template)
- Provide Vos max and the full test conditions (gain, VCM, VS, RL, output swing).
- Provide Vos drift max over temperature and define the measurement window (includes warm-up or steady-state only).
- Provide warm-up behavior (time-to-stable or warm-up drift) and the characterization method.
- Provide gain error max across gain settings and temperature range (and any headroom/load dependencies).
- Provide gain drift over temperature across gain settings.
- Provide any distribution / production capability (screening, 3σ data, guardband practice) for DC-accuracy terms.
D) Reference examples (part numbers as starting points)
These part numbers are provided only to speed datasheet lookup and lab validation. Final selection must follow the field list and risk weighting above (max values, conditions, and guardband).
Small-signal / bridge-style starting points
TI INA333 · TI INA188 · ADI AD8237
General precision INA starting points
ADI AD620 · ADI AD8221 · TI INA821
Low-noise precision starting point (DC accuracy perspective)
ADI AD8421
Request a Quote
FAQs (DC Accuracy: Offset, Drift, Gain Error)
These FAQs collect long-tail issues without expanding the main text. Each answer uses a fixed 4-line format: Likely cause / Quick check / Fix / Pass criteria.
Scope is limited to DC accuracy terms: offset, offset drift, gain error, and gain drift. Noise theory, EMI, filtering, and stability are not expanded here.
