Part 2 of The Analog Edge, a 14-part field guide to the analog half of a GopherTrunk installation. Part 1 drew the line at the ADC and inventoried the tracker bugs that lived on the analog side of it. This part picks up the first instrument on that side of the line — the dBFS meter — for our running reader with the marginal system, because the very first thing they’ll do is look at a level number and draw exactly the wrong conclusion from it. dBFS answers one question with total authority and a different question not at all, and telling those apart is the whole part.
TL;DR: dBFS is level relative to the ADC’s full scale — 0 dBFS is the rail, everything real is negative, and the only thing the number measures is headroom. GopherTrunk exports it as
gophertrunk_sdr_iq_power_dbfs(RMS over a ~1 s window: idle ≈ −45, healthy ≈ −25) withgophertrunk_sdr_iq_clip_ratiobeside it, because RMS averages away peak clipping — the clip ratio is the authoritative overload signal. And dBFS says nothing about quality: the two #764 captures both peaked ≈ −48 dBFS, yet one locked at ~19.7 dB demod SNR and the other sat undecodable at ~9.5 dB. Same meter reading, 10 dB of usable signal apart.
Key takeaways
- 0 dBFS is a cliff, not a target. Full scale is where the ADC stops being a measuring instrument; every dB of margin below it is headroom against clipping, and you want plenty.
- Peak and RMS are different numbers, and GopherTrunk’s gauge is RMS. A high-crest signal can pin the rails on peaks while the average still reads a merely-hot −5 dBFS — which is why the clip ratio exists as a separate, authoritative metric.
- dBFS cannot rank two signals by decodability. It measures how loud the digitized waveform is, not how much of that loudness is your channel versus noise, neighbors, and front-end artifacts.
- Learn the regimes, not a single magic number. Idle ≈ −45, healthy ≈ −25, > −3 clipping — those anchors come straight from the metric’s own help text, and the table below fills in the space between.
Cheat sheet
| Concern | What it tells you | Where it lives |
|---|---|---|
| Mean level, per control SDR | RMS dBFS, ~1 s window, labelled per system | gophertrunk_sdr_iq_power_dbfs (internal/metrics/prom.go) |
| Overload, authoritatively | fraction of samples pinned to the rail | gophertrunk_sdr_iq_clip_ratio — sustained > ~0.002 is overload |
| Whole wideband capture | pre-DDC level per dongle serial | gophertrunk_sdr_wideband_input_iq_power_dbfs |
| Wideband overload | a strong site burying weaker taps | gophertrunk_sdr_wideband_input_clip_ratio (issue #749) |
| DC contamination | DC-bin power vs total; ≤ −20 dB is clean | gophertrunk_sdr_iq_dc_ratio_db |
| The definition itself | full scale, headroom, clipping | dBFS in the Field Guide |
In this post
- What full scale actually is — the ADC’s rail and the sign convention.
- Peak vs RMS — crest factor, and why the gauge needs a partner metric.
- Where GopherTrunk shows the number — the gauges and their anchors.
- What dBFS cannot tell you — the two −48 dBFS captures of #764.
- The regime table — reading → likely condition → action.
What full scale actually is
An ADC measures voltage against a fixed reference. Drive it with exactly that reference voltage and every bit of the sample is used — that’s full scale, and we call it 0 dBFS. Everything smaller is expressed as decibels below that ceiling, so real-world readings are always negative: −20 dBFS is a waveform reaching a tenth of the rail voltage, −40 dBFS a hundredth. (If decibels themselves are shaky ground, the decibels lesson in the learning module is a fifteen-minute fix.)
Two properties follow immediately. First, dBFS is a digital unit: it says where the waveform sits relative to this converter’s rail, and nothing about absolute RF power at the antenna — the same station moves up and down the dBFS scale as you turn the gain knob. Second, 0 dBFS is not a maximum you approach for best results; it’s the point where the ADC stops measuring and starts lying. A sample that would have been larger than full scale is recorded at the rail, its true value gone. That’s clipping, and Part 4 is about how catastrophically it spreads across a band.
Peak vs RMS: why the gauge needs a partner
“The level” is actually two numbers. Peak is the largest single sample in a window; RMS is the energy average. The gap between them is the signal’s crest factor, and it varies enormously: a single constant-envelope carrier has a small crest factor, while a wideband capture full of independent carriers adds up like noise, with rare peaks far above its average.
GopherTrunk’s gophertrunk_sdr_iq_power_dbfs gauge is RMS over a roughly
one-second window, and its own help text carries the warning that matters:
gophertrunk_sdr_iq_power_dbfs Mean IQ power on the control SDR in dBFS
(window ~= 1 s). Idle ≈ -45; healthy signal ≈ -25. This is RMS and
averages away peak clipping — watch iq_clip_ratio for the
authoritative overload signal.
gophertrunk_sdr_iq_clip_ratio Fraction of IQ samples pinned to the ADC
rail. 0 = no clipping; a sustained value above ~0.002 means front-end
overload — reduce gain or add attenuation, do not raise gain.
Read that pair as designed: a high-crest capture can clip on peaks while the RMS reads a merely “hot” −5 dBFS, so the clip ratio, not the power gauge, is the overload verdict. This is not a hypothetical — the Nineteen Dibits postmortem turned on a capture that was 50% rail-pinned while every downstream symptom pointed, convincingly, at software.
Where GopherTrunk shows the number
Four gauges cover the level story end to end. Per decoded system, the
narrowband iq_power_dbfs (labelled "<system> @ <freq> MHz") tracks the
channel actually feeding a decoder. Per dongle, the wideband pair —
wideband_input_iq_power_dbfs and wideband_input_clip_ratio — measures the
whole capture before channelization, which matters because every tap on a
wideband dongle shares one gain: a strong site can overload the shared ADC
and bury its weaker siblings, the issue #749 failure mode the
config.example.yaml comments walk through. A tap sitting far below the
wideband input power, or any sustained non-zero clip ratio, is also logged as
a throttled WARN so you don’t need a Prometheus dashboard to catch it.
The anchors to memorize come from the help text: idle ≈ −45 dBFS (an antenna hearing only the noise floor), healthy ≈ −25 dBFS, and anything above −3 dBFS means clipping. Near-zero readings in the other direction — down at −70 and below — mean the chain is delivering essentially nothing: no antenna, gain at zero, or a dead cable. The DMR two-slot work from Part 1 is still blocked on exactly such a capture: ~−75 dBFS RMS, no frame sync anywhere in it.
What dBFS cannot tell you
Here is the part that saves you a week of mis-aimed debugging. In #764, the reporter supplied two captures of the same site from the same antenna: one at 2.5 MS/s, one at 10 MS/s. Both peaked at approximately −48 dBFS — neither clipped, neither was weak enough to suspect a dead branch. On the level meter, they are identical twins.
Decoded, they are nothing alike. The 2.5 MS/s capture replays at ~19.7 dB demod SNR (EVM 7.4%) and locks; the 10 MS/s capture manages ~9.5 dB (EVM 22.5%) and never locks. Ten decibels of usable signal separate two files the dBFS meter cannot distinguish — because the missing 10 dB wasn’t level, it was signal replaced by front-end phase noise, sample for sample, at the same amplitude. The full detective story is in Ten Megasamples, and Part 5 of this series retells its physics for operators.
The moral generalizes: dBFS is a headroom meter. It tells you whether the ADC has room to represent the waveform without lying. It does not tell you what fraction of that waveform is your channel, whether the modulation survived the front end, or whether a decoder will lock. Quality questions belong to quality numbers — demod SNR, decode error rate, and ultimately CRC yield — which is a thread that runs all the way to Part 13’s coherence-over-dBFS argument.
The regime table
What to do with a reading, from the top of the ladder down:
iq_power_dbfs reads… |
Likely condition | Action |
|---|---|---|
| above −3 | clipping — the ADC is saturated | reduce gain or add attenuation now; never raise gain (Part 4) |
| −3 to −15 | hot; RMS may hide peak clipping | check iq_clip_ratio; if sustained > ~0.002, back the gain off |
| −15 to −35 | the healthy band (anchor: −25) | leave it alone; judge by decode quality, not by pushing level higher |
| −35 to −55 | workable but lean — both #764 captures peaked ≈ −48 here | fine if it decodes; if not, the deficit is SNR, not headroom — sweep gain against decode quality (Part 3), then look at antenna and rate (Parts 5–7) |
| −55 to −70 | very weak; decoders will be marginal | more signal, not more software: gain staging first, then antenna/LNA (Parts 7, 9) |
| below −70 | effectively dead: no antenna, gain 0, broken coax | fix the chain before reading anything else — no capture from here answers any question |
Two habits complete the table. First, always read the clip ratio with the power gauge — the pair is designed as an instrument, and either one alone can mislead. Second, treat the healthy band as a plateau, not a slope: once you’re clear of both the noise floor and the rails, adding level buys nothing and eventually costs plenty. The SDR gain & overload Field Guide entry condenses this into one page you can keep open.
Where this goes next
If dBFS is a headroom meter, what should you actually turn the gain knob
against? Part 3
answers with a cautionary tale from the tracker — an operator who raised
front-end gain 17 dB to push a reading 0.2 dB past a software constant — and
the method that replaces threshold-chasing entirely: sweep the ladder and
score each rung by decode quality, the way AutoGainSweep does it.
FAQ
Is a stronger dBFS reading always better? No. Between the noise floor and the rails there is a wide plateau where level changes decode quality not at all, and above it every extra dB eats headroom until peaks clip. The gauges’ anchors — healthy ≈ −25, clipping > −3 — bracket that plateau. Once you’re on it, the knob that matters is covered in Part 3, and it isn’t level.
My waterfall looks strong but nothing decodes. How? Because an FFT display and a decoder measure different things. A carrier can be tall in the spectrum while its modulation is degraded — by phase noise, intermod, or multipath — and #764 is the canonical case: the wideband FFT carrier SNR was actually higher in the capture that failed. Trust demod SNR and error rate over any picture of the band.
What’s a normal reading with no antenna connected? The gauge should drop toward its idle anchor or below — you’re measuring the receiver’s own noise. If it doesn’t move when you disconnect the antenna, you were never measuring the antenna in the first place: suspect a stuck gain stage, a strong local interferer getting in past the connector, or a mislabeled device.
Why does GopherTrunk use RMS for the gauge instead of peak? RMS over a ~1 s window is stable enough to alert on and matches how energy (and therefore SNR) accumulates. Peak alone is jumpy and dominated by single samples. The design splits the job: RMS for level, and a dedicated clip ratio for the rail — the one question where peaks are exactly what matters.
Does dBFS relate to dBm? Only through your entire gain chain. dBm is absolute power at a reference point; dBFS is relative to one converter’s full scale, after antenna, coax, LNA, and tuner gain have all had their say. Change any of them and the same dBm at the antenna lands at a different dBFS. That’s why this series treats dBFS as a staging instrument, never a field-strength meter.
Series navigation
Part 2 of 14 · ← Part 1: Where the Software Ends · Next → Part 3: Gain Staging — Never Chase a Software Threshold