Noise & Dynamic Budget (SNR/ENOB) for Active Filters
← Back to: Active Filters & Signal Conditioning
Noise & Dynamic Budget is about translating every noise source (front-end, filter shaping, ADC/quantization, reference) into one consistent scale (RMS over ENBW and dBFS at the ADC), then verifying the chain meets the target SNR/ENOB without clipping.
When bandwidth, gain code, or source impedance changes, the budget must be updated first to reveal the true dominant term—only then should the circuit or parts be adjusted.
H2-1 · What “Noise & Dynamic Budget” means in an active-filter chain
Noise & dynamic budgeting aligns headroom to full-scale with the in-band noise floor, ensuring target SNR/ENOB without clipping inside the specified bandwidth.
“Dynamic” does not mean “maximum possible range.” It means a contract between two distances measured on the same ruler: (1) how close the largest expected signal gets to full-scale (headroom), and (2) how far the integrated noise stays below full-scale (noise floor). A correct budget predicts performance before layout and lab time, and it remains verifiable with measurements.
Three equivalent coordinate frames keep every specification comparable. Switching frames is allowed; mixing frames is not.
- Input-referred (at the sensor/source): best for deciding whether the dominant limit is source impedance, amplifier (en/in/1/f), or bandwidth.
- ADC-referred (at the ADC input pin): best for checking whether driver/noise-gain placement makes the ADC, reference, or post-filter dominate.
- dBFS (full-scale as 0 dBFS): best for exposing wasted ENOB when signal utilization is low or headroom is oversized.
Vn_rms ≈ en_rms_density × sqrt(ENBW)Vn_total = sqrt(Σ Vn_i²) (RSS)Noise_dBFS = 20·log10(Vn_total / VFS_rms)ENOB ≈ (SNR_dB − 1.76)/6.02The page deliverable is intentionally practical: a single budget table that converts every contributor to one reference point, plus a repeatable validation recipe (FFT noise floor and time-domain RMS with controlled bandwidth).
Common failure modes (budget looks “clean,” lab looks “ugly”):
- -3 dB bandwidth ≠ noise bandwidth. ENBW depends on the filter response and measurement windowing.
- Current noise is invisible until source impedance makes it voltage noise (in × Zs), especially at low frequency.
- ENOB is often wasted by low full-scale utilization. A quiet chain can still measure poorly in dBFS if the signal uses only a small fraction of FS.
H2-2 · Step 0: Freeze the system contract (bandwidth, full-scale, impedance, gain map)
Noise budgeting becomes deterministic only after the chain’s “contract” is frozen. Without a frozen contract, the budget attempts to predict a moving target: bandwidth changes ENBW, full-scale changes dBFS, source impedance changes the dominance of in × Zs, and gain map changes which stage masks the next.
BW (signal band + noise-integration basis), FS (ADC full-scale in Vrms or Vpp), Zs (source impedance, at least a usable approximation), Gain map (stage-by-stage, including programmable ranges), Headroom (3–6 dB typical).
1) Freeze the input signal contract (worst-case driven)
- Max peak (including bursts/steps/over-range behavior): defines clipping risk and required headroom.
- Nominal RMS: defines target SNR/ENOB and expected noise floor requirements.
- Offset + drift: consumes swing even when AC signal is small; must be included in headroom.
- Common-mode range (for differential chains): treated as a constraint only; detailed CM control belongs to converter/filter pages.
2) Freeze the bandwidth contract (separate signal BW from noise BW)
- Signal BW: the passband content that must preserve amplitude/phase within spec.
- Noise integration BW: represented by ENBW, not by the -3 dB corner. ENBW will be used later to integrate noise density consistently.
3) Freeze the full-scale contract (define dBFS unambiguously)
- Define full-scale as Vrms or Vpp and keep that definition consistent in all conversions.
- State the intended utilization (example: signal RMS at -3 dBFS) and the peak safeguard (example: worst peak stays below -1 dBFS).
- Headroom is not “free”: extra headroom reduces dBFS utilization and can reduce effective SNR even if the analog noise is low.
4) Freeze source impedance (Zs)
- If Zs is frequency dependent, record a usable approximation inside the passband; that is sufficient for a first-order budget.
- High Zs shifts dominance toward current noise and 1/f at low frequency; low Zs shifts dominance toward voltage noise and resistor thermal noise.
5) Freeze the gain map (stage-by-stage, including ranges)
- Record each stage gain and bandwidth; treat programmable gain steps as separate “modes.”
- Budget must pass at the worst-case mode: smallest gain often exposes ADC/reference noise; largest gain often exposes headroom/clipping constraints.
BW_signal = 20 kHz, ADC_FS = 2.0 Vpp (define Vrms equivalent), Zs ≈ 1 kΩGain_total = +20 dB (split across stages), Headroom = 6 dBH2-3 · Thermal noise: resistors, source impedance, and ENBW in one page
Most “noise budgets” fail because they mix bandwidth definitions. Thermal noise starts as a flat density, but the final RMS depends on how the transfer function weights it and on the correct ENBW (not the -3 dB corner).
Thermal noise is the baseline floor: every real resistor and every real source impedance contributes. In budgeting, the goal is not to re-derive physics, but to convert “noise density” into “in-band RMS noise” using a consistent ruler. That ruler is ENBW (Equivalent Noise Bandwidth).
1) From resistor/source to noise density (nV/√Hz)
- Resistors contribute thermal noise density proportional to the square-root of resistance.
- Source impedance should be treated as an equivalent noise generator in-band (use a practical approximation of the resistive part).
- Frequency-dependent Zs matters: if Zs changes across the passband, current-noise conversion (next chapter) and the thermal floor both shift.
2) ENBW: same fc, different response ⇒ different integrated noise
ENBW replaces a shaped weighting (|H(f)|²) with an equivalent rectangular bandwidth that produces the same total noise power. Two filters with the same “corner frequency” can have different ENBW because their magnitude responses distribute gain differently across frequency.
Vn_rms ≈ en_density × sqrt(ENBW)ENBW ≈ K × fc (K depends on response definition)3) Practical ENBW strategy (avoid coefficient arguments)
- First pass: use ENBW ≈ K × fc with a conservative K range suited to the chosen response family.
- Second pass: compute ENBW from the measured/simulated magnitude response so the budget matches the implementation.
- Validation rule: for white-noise dominance, doubling ENBW increases RMS noise by about √2 (≈ +3 dB).
Common traps to call out explicitly
- -3 dB bandwidth is not ENBW. Using fc as the noise bandwidth often underestimates RMS noise.
- Zs is not a single DC number for many sensors and front-ends; in-band behavior can change the noise floor materially.
- FFT “per-bin” noise is not total noise unless the bandwidth and window ENBW are correctly accounted for (handled again in the validation chapter).
H2-4 · Amplifier noise model: en, in, 1/f corner, and how source-Z turns current noise into voltage
Amplifier noise is not a single number. The input-referred model has three key parts: en, in, and 1/f. Current noise becomes voltage noise through the impedance seen at the input (Zseen), and low-frequency bandwidth definition sets how much 1/f is integrated.
The budgeting target is a consistent conversion: all amplifier noise contributors should be expressed as an equivalent input-referred noise density, then integrated over the same ENBW and referred to the same point as thermal noise. This prevents “good-looking datasheets” from producing mismatched lab results.
1) The three-part noise model (budget view)
- en: input voltage-noise density (dominant in many mid-band, low-impedance systems).
- in: input current-noise density (becomes voltage noise through impedance).
- 1/f: flicker region; its contribution depends strongly on the low-frequency integration limit (fL).
2) The key conversion: current noise × impedance = voltage noise
Current noise does not directly appear as a voltage until it flows through the impedance seen at the input. In budgeting, define a usable Zseen approximation inside the passband (often close to the source impedance plus any explicit input network). Then compare magnitudes:
if (in · |Zseen|) > en ⇒ current-noise term becomes dominant3) How 1/f is handled in a budget (without expanding into servo design)
- Define a low-frequency bound fL for integration (measurement window, system high-pass/servo corner, or application limit).
- If a datasheet provides 0.1–10 Hz RMS noise, treat it as a low-frequency noise block and RSS it with the white-noise result.
- If only curves are available, approximate the 1/f region with a piecewise density and integrate from fL to fH using the same ENBW concept.
4) Dominance guide (quick decision table)
| Condition | Most likely dominant term | Budget action |
|---|---|---|
| High Zs / high Zseen e.g., kΩ–MΩ sources |
in · Zseen current-noise conversion |
Model Zseen carefully; select lower in; avoid unnecessary input resistance. |
| Low Zs + mid/high band white region dominates |
en + thermal floor | Select lower en; control ENBW; ensure gain placement masks ADC noise later. |
| Very low fL (near DC) long integration window |
1/f | Define fL explicitly; include 0.1–10 Hz spec or integrate flicker region conservatively. |
Common traps to call out explicitly
- Picking an amplifier by en only while ignoring in × Zseen (high-impedance chains fail quietly).
- Ignoring 0.1–10 Hz noise in low-frequency systems (the budget looks great, drift/noise looks bad in the lab).
- No fL definition (1/f can be undercounted or overcounted by orders of magnitude).
H2-5 · Filter noise shaping: pre-filter vs post-filter, and why “where you filter” matters
A filter shapes noise as much as it shapes signal. With the same total gain and nominal bandwidth, moving the filter earlier or later changes which noise sources get weighted by |T(f)|² and which noise terms bypass filtering entirely.
A correct budget is position-sensitive: each noise contributor must be referred from its injection point to the observation point (typically the ADC input). Noise density does not “just add up” by DC gain; it is weighted by frequency response. For contributor i, the output noise power is shaped by |Ti(f)|², then integrated over the relevant bandwidth.
Vn_out² = ∫ Sni(f) · |Ti(f)|² dfPre-filter vs post-filter: why they differ
- Pre-filter (filter first): source noise and any upstream noise are shaped early, but noise injected after the filter (later stages) is not removed by that filter.
- Post-filter (filter later): upstream noise and amplifier noise (if injected before the filter) can be shaped by the post-filter, reducing in-band noise at the ADC.
- Same gain ≠ same noise: equal DC gain can still produce different integrated RMS noise because noise weighting depends on injection location and magnitude response.
What must be accounted for (common traps)
- Noise gain vs signal gain: some active structures amplify internal noise differently than the signal path; budgeting must follow the noise path, not just DC gain.
- Stopband noise can still matter: in sampled systems, out-of-band noise can fold into band if anti-alias filtering is insufficient. This page flags the risk; detailed AAF strategy belongs to the AAF page.
- “Filter order” is not the whole story: the shape of |H(f)| and where noise is injected are both required to predict in-band RMS noise.
H2-6 · ADC quantization and input noise: turning everything into dBFS and ENOB
Convert every noise contributor into RMS noise at the ADC input, then express it as dBFS. Once signal and noise share the same full-scale ruler, SNR and ENOB become direct outcomes rather than guesses.
The most common budgeting failures in mixed-signal chains come from inconsistent full-scale definitions. Before converting to dBFS, full-scale must be defined unambiguously (Vrms vs Vpp) and used consistently across signal, noise, and datasheet conversions.
1) Quantization noise in engineering form (no “ADC textbook”)
LSB = VFS / 2^N (keep VFS definition consistent)Vq_rms ≈ LSB / √12 (uniform quantization model)2) ADC input noise / datasheet SNR → equivalent RMS noise
- If the datasheet provides SNR under defined conditions, it can be treated as an equivalent noise floor for that mode.
- Convert SNR to an equivalent RMS noise at the ADC input using the same signal and full-scale conventions used in the datasheet.
- If only SINAD/ENOB is provided, treat it as a conservative “noise+distortion equivalent floor” unless a noise-only SNR is available.
Noise_dBFS = 20·log10(Vn_total_rms / VFS_rms)SNR_dB ≈ Signal_dBFS − Noise_dBFSENOB ≈ (SNR_dB − 1.76) / 6.023) Full-scale utilization: the most common “wasted ENOB”
When signal amplitude uses only a fraction of full-scale, SNR drops by the same fraction in dB, even if the analog noise floor is unchanged. This is not a subtle effect; it often dominates real systems.
SNR loss (dB) = 20·log10(k), where k = signal / full-scalek = 0.2 (20%FS) ⇒ 20·log10(0.2) ≈ −14 dBCommon traps to call out explicitly
- Vpp vs Vrms confusion: formulas can be correct while results are off by a large margin.
- SINAD treated as SNR: distortion is included; this is conservative but must be labeled.
- Low utilization: strong ADC specs do not help if the chain only uses a small fraction of full-scale.
H2-7 · Reference noise and supply-coupled noise: when “ref” sets the real noise floor
The ADC’s code scale is set by Vref. Noise on Vref behaves like a moving ruler: even with a quiet analog front end, reference noise and supply-coupled noise can become the dominant limit on SNR/ENOB.
A noise budget that stops at amplifier and resistor noise is incomplete. Reference noise maps directly into code jitter because it perturbs the conversion scale. This becomes most visible in high-resolution or narrowband systems, especially when signals use a large fraction of full-scale.
1) Budget the reference as a first-class noise source
- LSB size depends on Vref: LSB = Vref / 2^N. Any Vref noise effectively modulates that step size.
- Reference noise can dominate once the analog chain noise is pushed very low (a common “why ENOB won’t improve” symptom).
- Low-frequency behavior matters: if the measurement bandwidth is very low, 1/f-like reference noise can integrate into a larger RMS value.
CodeNoise_rms ≈ VrefNoise_rms / LSB = VrefNoise_rms · 2^N / VrefRefNoise_dBFS ≈ 20·log10(VrefNoise_rms / Vref_rms)2) When reference noise is more dangerous than front-end noise
| Situation | Why ref noise becomes visible | Budget action |
|---|---|---|
| High resolution (large N) | LSB is small, so Vref ripple/noise can span multiple codes | Convert Vref noise to dBFS/codes and RSS it with the analog noise floor |
| Narrow bandwidth | Low-frequency noise integrates into a larger RMS value | Use the same bandwidth/ENBW definition for Vref noise as for input noise |
| High full-scale utilization | Scale noise maps directly into output codes when near FS | Keep the dBFS ruler consistent: signal_dBFS and ref_noise_dBFS must share the same FS definition |
3) Supply-coupled noise (budget view only)
Supply ripple can couple into the reference and the ADC input path. The budgeting goal is to express rail noise as an equivalent input/code noise using a coupling factor (often represented by PSRR or a measured transfer function), then integrate over bandwidth.
Veq_in(f) ≈ Vrail_n(f) · Kcouple(f)H2-8 · Gain staging & headroom: allocate noise vs clipping margin like a CFO
More gain improves full-scale utilization (better dBFS and ENOB), but it reduces clipping margin and can increase distortion risk. A robust dynamic budget allocates margin explicitly: signal level, peak headroom, and noise dominance across all programmable gain codes.
Gain staging is an allocation problem. The aim is to keep typical signals close to full-scale without violating peak limits under worst-case input, offset, and transient overshoot. At the same time, the front-end noise should remain above downstream/ADC floors so that later stages do not set the overall noise performance.
1) Three practical gain-staging rules
- Use full-scale efficiently (keep headroom): set a target signal level (in dBFS) and reserve peak margin for crest factor, offsets, and overshoot.
- Keep noise dominance upstream: ensure downstream/ADC noise floors stay at least several dB below the front-end in-band noise (avoid “back-end-limited” budgets).
- Guarantee worst-code compliance: check both extremes—high gain (clipping risk) and low gain (utilization/SNR loss).
Vin_peak_max = (signal_peak_max + offset_max + overshoot_est) · gain_maxVin_peak_max < FS_peak − margin2) Programmable gain codes: what “worst case” usually means
| Gain code risk | Failure mode | Budget check |
|---|---|---|
| High gain | Clipping / compression / distortion | Peak headroom under max input + max offset + overshoot |
| Low gain | Wasted dBFS → reduced SNR/ENOB | Signal utilization vs noise floor (dBFS) and required SNR |
| Mid gain | Usually best balance | Confirm upstream noise dominance and stable margins |
Common traps to call out explicitly
- RMS-only thinking: peak/crest factor and transient overshoot can invalidate an otherwise “clean” RMS budget.
- Typical-only assumptions: max input + max offset + tolerance must be checked explicitly for clipping risk.
- Back-end-limited noise: if ADC/driver/reference floors dominate, front-end improvements will not move the measured ENOB.
H2-9 · Build the full noise budget table (worked template + sanity checks)
A complete budget table that converts every noise contributor into a single reference point (Referred-to-ADC + dBFS), combines them by RSS, and closes to SNR/ENOB. Includes sanity checks that catch “fake budgets.”
A credible noise budget is not a list of numbers. It is a consistent pipeline: choose a reference point, keep bandwidth definitions consistent (ENBW), convert each contributor into RMS noise at that reference point, then combine by root-sum-square (RSS).
1) Budget table field template (fillable schema)
| Field | Meaning | Why it must exist |
|---|---|---|
| Stage / Block | Where the noise is generated (source Z, resistor, amp, filter, ADC, ref, rail-coupled) | For traceability and dominance decisions |
| Gain (lin + dB) | Gain map for each stage or gain code | Prevents hidden scaling errors and makes code-to-code checks possible |
| Noise density | nV/√Hz or pA/√Hz (clearly labeled) | Integrated noise is meaningless without density + bandwidth |
| Shaping term | |T(f)|, noise gain, or a practical ENBW factor (K) | Captures where filtering/gain placement changes the integrated noise |
| ENBW (Hz) | Equivalent noise bandwidth used for integration | Defines the “noise collection window” (must match measurement) |
| Integrated noise (Vrms) | Per-line RMS noise after shaping + ENBW integration | Required input to RSS summation |
| Referred-to-input (RTI) | Optional cross-check column | Sanity checking and front-end comparison |
| Referred-to-ADC (RTA) | Primary reference-point column | All contributors must land here before summation |
| dBFS | Noise relative to full-scale at ADC | Unifies analog noise, quantization, ref noise, and rail-coupled noise |
| Dominance flag | Dominant / Secondary / Negligible | Prevents “every item equals” fake budgets |
2) Combination rules (the only allowed math flow)
Vn_total_rms = sqrt( Σ Vn_i_rms² )Noise_dBFS = 20·log10( Vn_total_rms / VFS_rms )SNR_dB ≈ Signal_dBFS − Noise_dBFSENOB ≈ (SNR_dB − 1.76) / 6.023) Dominance flag (mechanical, not subjective)
Power share ≥ 50%: Vn_i² / Vn_total² ≥ 0.5
Secondary: 10–50%. Negligible: <10%.
4) Sanity checks (catch fake budgets fast)
- Dominance looks real: one or two items should typically dominate per gain code; “everything equal” usually indicates a unit/ENBW/reference-point error.
- Gain-step behavior makes sense: stepping gain should shift utilization and dominance (back-end-limited vs front-end-limited regimes).
- White-noise scaling holds: if ENBW doubles, integrated white noise should rise by about √2; large deviations point to 1/f, interference, or bandwidth mismatch.
H2-10 · How to validate: FFT noise floor, RMS noise, bandwidth control, and common traps
Measurement must match the budget’s bandwidth and scale. Validate noise using both FFT (with correct ENBW handling) and time-domain RMS under controlled termination, and avoid common artifacts that create false agreement.
1) Field validation triad (repeatable workflow)
Short input or use a known termination impedance (document the method and value).
Fix FFT length, window, and integration band; align ENBW/RBW with the budget table.
Integrate FFT power over the same band and compare to time-domain RMS under equivalent filtering.
Gain code, Fs, N, window type, ENBW, and band limits must be written in the test log.
2) FFT floor vs total RMS noise (what must be true)
A single FFT bin is not “total noise.” Total RMS noise requires bandwidth integration (or power summation across bins) using a consistent scaling and window ENBW. Without ENBW alignment, FFT plots can look “better” or “worse” while the true in-band RMS noise is unchanged.
Vn_rms_band ≈ sqrt( Σ Vbin_rms² )3) Six common traps (symptom → cause → fix)
| Trap | Typical symptom | Fix |
|---|---|---|
| Bin ≠ total noise | FFT floor looks low but time RMS remains high | Integrate over bandwidth (power sum). Log band limits and ENBW. |
| Window ENBW ignored | Noise floor changes when window type changes | Fix window choice; apply ENBW correction; keep the test contract constant. |
| Input not truly terminated | 50/60 Hz and harmonics dominate; results drift with environment | Use verified short/known impedance termination; document the method. |
| Sampling-rate unit confusion | Changing Fs moves the FFT floor and is misread as “noise changed” | Compare in consistent units (Vrms over band or dBFS over band). Avoid per-bin comparisons. |
| Clipping makes noise look smaller | RMS noise seems to improve while waveforms show limiting | Verify peak headroom first; ensure linear operation before noise measurements. |
| Ref/rail code drift misattributed | Slow code wandering or low-frequency “noise” persists across front-end changes | Compare to budget lines for ref/rail; monitor ref/rail spectra during measurement. |
H2-11 · Design checklist: quick rules-of-thumb for active-filter chains (with example parts)
Convert the target ENOB/SNR into an allowed noise floor (dBFS), then force a clear dominance story (front-end vs back-end). Treat bandwidth changes as budget changes first, and only then change the circuit.
Rule 1 — Target ENOB → allowed total noise (dBFS)
Start from the performance contract. For a given ENOB target, compute the required SNR, then derive the maximum allowed in-band noise floor in dBFS (relative to the ADC full-scale, using one consistent Vrms definition and a defined signal utilization).
SNR_target(dB) ≈ 6.02 · ENOB_target + 1.76
Noise_target(dBFS) ≈ Signal_dBFS − SNR_target
Keep Signal_dBFS fixed per gain code (e.g., −3 dBFS or −6 dBFS) to preserve headroom.
Use datasheet SNR/SINAD to back-calculate equivalent noise when building the “ADC intrinsic” line item in the table.
Rule 2 — If the front-end must dominate, enforce a margin over back-end noise
Enforce a dominance margin so the front-end contribution (referred to ADC) stays measurably above the back-end lumped noise (ADC intrinsic + reference + late-stage/driver), across the worst gain code.
Vn_front_end_RTA ≥ Vn_back_end_RTA · M
Choose M ≈ 2 (≈6 dB) for robust dominance, or larger if gain codes/temperature spread are wide.
A clean reference and a stable ADC driver prevent “ref-limited” or “drive-limited” surprise floors.
Rule 3 — Source impedance decides whether in/1-f or en/4kTR is the first worry
High source impedance converts current noise into voltage noise (in·|Zs|) and exposes low-frequency 1/f behavior. Low source impedance shifts priority to voltage noise (en) and resistor thermal noise (4kTR).
Vn_in_equiv ≈ in · |Zs|
Vn_en_equiv ≈ en (+ resistor 4kTR contribution)
Whichever term wins at the target band should get explicit line items and dominance attention in the table.
Rule 4 — When bandwidth changes, update the budget first (then change the circuit)
Bandwidth changes alter integrated noise and can flip dominance. Treat bandwidth as a budget change: update ENBW, re-run Vrms lines and RSS/dBFS closure, then modify filter order, gain staging, or parts only after the new dominant term is identified.
(1) Update ENBW → (2) Recompute Vrms lines → (3) RSS + dBFS → (4) Identify dominant term → (5) Change circuit
For white-noise-dominated cases, doubling ENBW should raise RMS noise by about √2. Large deviations indicate 1/f or interference/band mismatch.
- Budget table complete: Stage, Gain, Noise density, ENBW, Integrated Vrms, RTI/RTA, dBFS, Dominance flag.
- Coordinate locked: full-scale definition and ENBW are consistent across budget and measurement.
- Dominance story: each gain code yields 1–2 dominant contributors (not “everything equal”).
- Closure: Noise_dBFS closes to SNR/ENOB with margin for temperature and tolerances.
- Validation loop: termination, FFT window/ENBW, band integration, and time RMS cross-check are documented.
- Parts documented: each stage lists the chosen part number(s) and the spec that justified them (en/in/1-f, SNR, ref noise, drive requirement).
Request a Quote
H2-12 · FAQs (Noise & Dynamic Budget)
Answers are written in a consistent budgeting/validation format (dBFS + ENBW + RSS). Example part numbers are illustrative anchors, not a full selection guide.
1Why can the same circuit show very different noise after changing the measurement bandwidth?
Because “total noise” is an integrated quantity, changing bandwidth (or window ENBW) changes the integration window. FFT floor per bin is not total RMS noise; the spectrum must be integrated over the same band with the same window scaling. Validate by fixing Fs, N, window, and band limits, then cross-check FFT-integrated RMS vs time-domain RMS. Example ADC: AD7177-2.
2Should a noise budget be done input-referred or in dBFS?
Use both, but do not mix them. Input-referred (RTI) is best for comparing analog front-end blocks, while dBFS at the ADC is best for unifying analog noise, ADC noise, quantization, and reference noise in one scale. A practical table keeps Referred-to-ADC (RTA) + dBFS as the primary columns and RTI as a cross-check. Example ADCs: AD4003, ADS8881.
3When source impedance increases, what usually causes the sudden noise degradation?
Higher source impedance converts amplifier current noise into voltage noise (in·|Zs|) and often exposes low-frequency 1/f behavior. It can also raise thermal noise if series resistance increases. Budget it explicitly by adding separate lines for in·Zs and resistor 4kTR, then compare their integrated RMS within the target ENBW. Example front-end parts: OPA140, ADA4522-2.
4Why does “putting gain later” often reduce ENOB?
Later gain cannot undo noise already added by the ADC, reference, and late-stage blocks; instead, those back-end noises become dominant when early gain is too small. A robust chain enforces a dominance margin so front-end noise (referred to ADC) stays above the back-end lumped noise across the worst gain code, while maintaining headroom. Re-check dominance after any gain-map change. Example drivers: THS4551, ADA4945-1.
5How to tell whether reference noise is the system bottleneck?
Reference noise directly modulates the ADC code scale, so it can set the real dBFS noise floor even when the analog front end is quiet. Convert the reference noise into an equivalent Referred-to-ADC RMS noise (or dBFS) over the same ENBW and compare it to the RSS total; if it dominates, front-end improvements will not help. Validate by monitoring code noise while swapping only the reference. Example refs: ADR4550, LTC6655.
6FFT noise floor looks great, but time-domain RMS noise is bad—what is usually wrong?
The most common errors are treating “per-bin” FFT noise as total noise, ignoring window ENBW, or integrating the wrong band. Another frequent cause is an input that is not truly terminated, injecting 50/60 Hz and environmental pickup that inflates time RMS. Fix the contract: termination method, Fs, N, window type, ENBW, and band limits must match the budget. Example ADC: AD7768-1.
7If only 30% of full-scale is used, how much SNR is lost, and how can it be recovered?
Using only 30% of full-scale reduces signal utilization by 20·log10(0.3) ≈ −10.5 dB, so achievable SNR (and effective ENOB) drops by roughly the same amount if the noise floor is unchanged. Recover it by moving gain earlier (while keeping headroom), optimizing gain codes for worst-case input, and preventing back-end dominance. Recompute Signal_dBFS and Noise_dBFS per gain code. Example: AD4003 with THS4551 drive.
8How should 1/f noise be represented inside the budget table?
1/f noise cannot be captured by a single flat noise density across the band; it must be budgeted with either segmented bands or a dedicated low-frequency line item. A practical method is to split the integration into a low-frequency region (where 1/f dominates) and a white-noise region, then RSS the resulting RMS values at the same reference point. This prevents “white-noise-only” budgets from underestimating low-frequency chains. Example low-1/f parts: OPA2188, ADA4522-2.
9ENBW is unclear—are there practical engineering approximations?
ENBW depends on the transfer function shape and is not equal to the −3 dB bandwidth. For quick budgeting, use an approximation of the form ENBW ≈ K·fc, where K is response-dependent; keep K consistent for the same response family and validate it by measurement. During validation, the FFT window also has its own ENBW, which must be accounted for when integrating spectral noise. Example workflow: AD7177-2 noise test with fixed window and band integration.
10How to separate and combine quantization noise vs ADC intrinsic noise?
Quantization noise is a theoretical limit for an ideal converter, while real ADCs include intrinsic noise from internal circuits and reference sensitivity, often summarized by datasheet SNR/SINAD. Convert each contributor into the same unit (Referred-to-ADC Vrms over ENBW or dBFS), then RSS-combine them. If the datasheet gives SNR, back-calculate the ADC equivalent noise and add it as a distinct budget line next to quantization. Example ADCs: LTC2500-32, ADS8881.
11How to define “dominant contributors” in the budget without self-deception?
Use a mechanical dominance metric based on noise power share: Vn_i² / Vn_total². Mark a contributor as dominant if it exceeds ~50% of total noise power, secondary if ~10–50%, and negligible below that. A realistic chain usually shows one or two dominant items per gain code; if every line looks equal, there is typically a unit mismatch (Vrms vs Vpp), an ENBW mistake, or a reference-point mistake. Example back-end suspects: ADR4550 reference, THS4551 driver.
12How to design a minimal-cost noise validation experiment (especially input handling)?
Start with input termination discipline: true short (or a documented known impedance) at the input node that defines the budget reference. Next, lock the measurement contract (Fs, N, window, ENBW, and band limits), then compute both FFT-integrated RMS and time-domain RMS under the same bandwidth control. Finally, repeat at two bandwidths; white-noise-dominated chains should scale by about √ENBW. Example measurement chain: AD7177-2 with a clean reference such as LTC6655 to avoid ref-limited confusion.
