In short: SNR, SNDR, SFDR and ENOB all describe an ADC's dynamic (AC) accuracy while INL/DNL describe its static (DC) accuracy — and ENOB is simply SNDR rewritten as a bit count, so a "16-bit" converter can easily behave like an 11-bit one once real noise and distortion are counted.
Every datasheet proudly prints a resolution — 12-bit, 16-bit, 24-bit. That number only tells you how many output codes exist; it says nothing about how many of them are actually trustworthy. The gap between the nominal resolution and the usable resolution is where ADC performance metrics live. If you are preparing for an analog or mixed-signal interview, or specifying a converter for a real data-acquisition front end, you need to read these numbers the way a designer does. Let us walk through them with concrete figures.
Static vs. dynamic metrics: two different questions
ADC specifications split cleanly into two families. Static (DC) metrics — offset, gain error, DNL and INL — answer "if I apply a slowly-changing voltage, how close is the code to ideal?" Dynamic (AC) metrics — SNR, SNDR, SFDR, THD and ENOB — answer "if I feed a fast sine wave, how clean is the reconstructed spectrum?" A converter can be excellent at one and poor at the other, which is why both appear on the datasheet.
DNL and INL: the staircase under a microscope
An ideal ADC transfer curve is a uniform staircase: every code is exactly 1 LSB wide. Differential nonlinearity (DNL) is the deviation of each step width from that ideal 1 LSB. A DNL of +0.4 LSB means that code is 1.4 LSB wide; a DNL of −1.0 LSB means the code has zero width and is missing — a serious fault. Integral nonlinearity (INL) is the running sum of DNL errors, i.e. how far the whole curve bows away from a straight line end-to-end.
Rule of thumb: DNL > −1 LSB everywhere guarantees no missing codes and monotonic behaviour. A 16-bit SAR with ±0.5 LSB DNL and ±2 LSB INL is typical; the INL bow is what limits usable accuracy in a weigh-scale or sensor bridge.
SNR, THD, SNDR and SFDR: reading the FFT
Dynamic metrics are measured by driving the ADC with a pure sine wave and taking an FFT of the output. Four numbers fall out:
- SNR — ratio of signal power to total noise power, excluding harmonics. For an ideal N-bit ADC, SNR = 6.02·N + 1.76 dB. A perfect 12-bit part tops out at 74 dB.
- THD — total harmonic distortion: the summed power of the harmonics relative to the fundamental.
- SNDR (or SINAD) — signal to noise-and-distortion: noise and harmonics together. Always ≤ SNR. This is the honest "everything bad" number.
- SFDR — spurious-free dynamic range: the gap from the fundamental to the single largest spur (often a harmonic). It sets how small a tone you can detect next to a big one — critical in RF and communications receivers.
ENOB: where everything converges
Effective Number Of Bits rearranges the ideal-SNR formula to turn SNDR back into a bit count:
ENOB = (SNDRdB − 1.76) / 6.02
Consider an "ADC" that advertises 16 bits but measures an SNDR of 90 dB at the Nyquist input frequency. ENOB = (90 − 1.76)/6.02 ≈ 14.6 bits. The lower ~1.4 bits vanished into noise and distortion — jitter, quantiser imperfections, reference noise. This is why experienced designers quote ENOB, not resolution, when sizing a signal chain.
ENOB falls as input frequency rises, mostly because of aperture jitter. With tjitter = 1 ps RMS and a 70 MHz input, jitter alone caps SNR near 67 dB (≈10.8 ENOB) — regardless of how many physical bits the core has.
| Metric | Type | What it bounds | Typical 16-bit SAR |
|---|---|---|---|
| DNL | Static | Missing codes / monotonicity | ±0.5 LSB |
| INL | Static | Absolute linearity | ±2 LSB |
| SNR | Dynamic | Noise floor | 92 dB |
| SNDR | Dynamic | Noise + distortion | 90 dB |
| SFDR | Dynamic | Largest spur | 105 dB |
| ENOB | Derived | Real usable bits | 14.6 |
How the metrics relate to each other
They are not independent. SNDR combines SNR and THD through the power relation 10−SNDR/10 = 10−SNR/10 + 10−|THD|/10. ENOB is a direct restatement of SNDR. SFDR is usually larger than SNDR because it counts only the worst single spur, not the integrated distortion. DNL and INL, meanwhile, set a floor on SNR at low frequencies: a bowed INL curve injects low-order harmonics that degrade THD, coupling the static and dynamic worlds.
Measuring them in practice
In a lab you capture a coherent sine-wave record and run an FFT with proper windowing (or coherent sampling to avoid leakage). Simulation tools do the same on a testbench: Cadence Spectre or Virtuoso ADE for transient FFT analysis of a schematic-level converter, Synopsys VCS/CustomSim for mixed-signal co-simulation, and Siemens AFS/Analog FastSPICE for fast behavioural sweeps. For silicon bring-up, an ATE or a bench board with a high-purity source and a notch filter removes the generator's own distortion so you measure the ADC, not the stimulus.
If you want the architectural context behind these numbers — why a SAR converter trades ENOB for speed while a sigma-delta trades speed for ENOB — read that companion piece, and see our guide to Nyquist sampling and aliasing for why ENOB is always quoted at a specified input frequency. Teams that design these front ends for clients, like the mixed-signal group at Avecas design services, live and die by the ENOB-vs-frequency curve.
Want to measure ENOB on a real converter and read an FFT like a designer? Our hands-on analog course walks through it end to end. Explore CourseTron courses →
Frequently asked questions
Is higher ENOB always better?
For accuracy yes, but ENOB costs power and bandwidth. A sensor datalogger needs high ENOB at low speed; an RF digitiser may happily accept 8–10 ENOB at gigasample rates. Match ENOB to the signal chain, not to bragging rights.
Can SNDR exceed SNR?
No. SNDR adds distortion power to the noise power, so it is always less than or equal to SNR. If a datasheet shows SNDR > SNR, one of the numbers is defined differently — check the test conditions.
Why does ENOB drop at higher input frequencies?
Clock aperture jitter converts timing error into voltage error that scales with input slew rate. Faster inputs slew harder, so the same jitter produces more noise, lowering SNR, SNDR and therefore ENOB.
Does good INL guarantee good ENOB?
Not by itself. INL is a DC metric; ENOB is dominated by noise and jitter at speed. A converter can have superb INL yet mediocre ENOB at Nyquist if its sampling clock is noisy.