Core formulas
| Step size Δ (LSB) | = V_FS / 2^N (full-scale / 2^N) |
|---|---|
| Quantization noise | σ_q = Δ / √12 |
| SQNR (ideal) | ≈ 6.02 · N + 1.76 dB |
| ENOB (from SINAD) | = (SINAD − 1.76) / 6.02 |
Resolution / dynamic range
| Bits | Levels | SQNR (dB) | Use |
|---|---|---|---|
| 8 | 256 | 49.9 | Basic MCU ADC |
| 10 | 1 024 | 61.96 | Arduino (10-bit) |
| 12 | 4 096 | 74.0 | General MCU ADCs, audio for noise floors |
| 14 | 16 384 | 86.0 | Precision instrument |
| 16 | 65 536 | 98.08 | CD audio |
| 20 | 1 048 576 | 122.2 | Precision sigma-delta |
| 24 | 16.7 M | 146.2 | Audio DAW / weigh-scale |
| 32 | 4.3 G | 194.2 | Rare — limited by noise floor |
Practical notes
- ENOB is always less than advertised resolution due to noise and non-linearity.
- Oversampling + averaging: 4× oversampling gives 1 extra bit of effective resolution (up to the noise floor).
- Dithering: add small noise before quantization to decorrelate quantization error from signal.
- Input referred noise sets practical floor — beyond ~22 effective bits, thermal noise dominates.
Worked examples
Step size and noise: a 12-bit ADC with a 3.3 V full scale has an LSB of 3.3 / 4,096 = 0.806 mV. The quantization noise is Δ / √12 = 0.806 / 3.464 = 0.233 mV RMS, and the ideal SQNR for a full-scale sine is 6.02 × 12 + 1.76 = 74.0 dB. A sensor signal of 100 mV is only about 124 counts, so resolution is about 0.8% of the reading; if the signal fills only a tenth of the range, the SQNR falls by 20 dB.
ENOB: if a 16-bit converter measures a SINAD of 80 dB, ENOB = (80 − 1.76) / 6.02 = 13.0 bits. The bottom three bits are noise, so additional resolution is wasted unless the noise falls.
Oversampling: each doubling of the oversampling ratio adds 3 dB, so each factor of 4 adds one effective bit, provided the input has at least about 1 LSB of noise or dither. Averaging 16 samples of the 12-bit ADC gives about 14 bits, at the cost of a sample rate 16 times lower (100 ksps becomes 6.25 ksps).
Choosing the right resolution
- Match the bits to the signal, not the datasheet. If the sensor and amplifier have 0.1% noise, a 10-bit converter is already matched and a 16-bit one adds cost without improving the measurement.
- Use the full range. Scale the signal with a gain stage so the interesting part uses most of the input range. Every halving of the used range costs one bit.
- Check reference accuracy. A 16-bit ADC with a reference that drifts 0.1% across temperature is accurate to about 10 bits in practice.
- Audio: 16 bits (98 dB) covers most finished recordings, while 24 bits gives headroom for recording, because the analog noise floor of real microphones and preamps, not the converter, usually limits quality.