ADC Resolution vs Accuracy: Why More Bits Won’t Fix Noisy Readings

Resolution vs accuracy, in one picture

Think of the ADC reference as a ruler. Resolution is how finely the ruler is marked. A 24-bit ADC has very fine marks. But if the ruler itself stretches and shrinks with every bit of supply noise, those fine marks just measure the wobble.

How small is one step?

Here is one step (LSB) on a 3.3 V reference, and what 20 mV of ordinary charger ripple looks like at each resolution:

The same ripple that barely moves a 10-bit Uno turns the bottom half of a 24-bit result into random numbers.

Your usable bits are capped by supply and reference noise

The number of bits you can actually trust is set by the ratio of the reference voltage to the noise on it: usable bits ≈ log₂(Vref ÷ noise p-p). That limit is the same whatever ADC you plug in.

Why supply noise goes straight into the reading

Path 1: the supply is the reference

On an Arduino Uno the default reference is AVCC, the board’s 5 V. On HX711 load-cell boards the bridge excitation and the reference come from the same analog supply. Many sensor modules output a fraction of their own supply. In all these cases, a 1% wobble on the supply is a 1% wobble in the result. For bridge and divider sensors the wobble partly cancels, but only if the sensor and the reference share exactly the same node.

Path 2: limited PSRR, and it falls with frequency

ADCs with an internal reference (ADS1115, ESP32) are better, but not immune. Their power-supply rejection ratio (PSRR) is strong for slow mains ripple and much weaker at the hundreds of kilohertz where phone chargers and buck converters switch.

So noise has to be stopped before it reaches the ADC’s supply pin. That’s the job of a low-noise linear regulator with ferrite filtering: it rejects the switching noise the ADC itself can’t.

Measure your real usable bits in 2 minutes

Don’t trust the number on the box. Put a perfectly steady voltage on the input, a single AA cell (≈1.5 V, + to the analog pin, − to GND), read it a few hundred times and see how many counts it jumps:

// Real usable bits = log2(full scale / peak-to-peak counts)
const long FULL_SCALE = 1024;     // Uno 10-bit: 1024 · ESP32: 4096 · ADS1115 (+ only): 32768

void setup() { Serial.begin(115200); }

void loop() {
  long lo = 2147483647L, hi = -2147483647L;
  for (int i = 0; i < 256; i++) {
    long v = analogRead(A0);       // replace with your ADC's read function
    if (v < lo) lo = v;
    if (v > hi) hi = v;
    delay(2);
  }
  long pp = hi - lo + 1;           // +1: a perfectly steady reading still spans one code
  Serial.print("p-p counts: "); Serial.print(pp);
  Serial.print("   usable bits: "); Serial.println(log(FULL_SCALE / (double)pp) / log(2.0), 1);
  delay(1000);
}

Run it twice: once with the board on a phone charger, once with the analog side on a clean supply. The difference is what your supply is costing you.

Fix these before you buy more bits

1. A low-noise analog supplyPower the ADC and sensors from a dedicated low-noise linear regulator, after any switching converter, with ferrite filtering. Never share it with Wi-Fi, motors or LEDs.
2. A clean referenceIf the ADC uses its supply as the reference, the supply is your accuracy. For about 16 usable bits a low-noise LDO is enough; for 20+ bits use a dedicated precision reference IC.
3. Grounding and layoutKeep digital and load currents out of the analog ground. Join grounds at one point, keep analog wires short and away from switching parts.
4. Source impedanceHigh-impedance sensors can’t charge the ADC’s sampling capacitor in time. Add a 100 nF capacitor at the pin or a buffer amplifier.
5. Filtering and data rateAn RC filter at the input stops noise above your signal’s bandwidth. On sigma-delta ADCs (HX711, ADS1115, ADS1256), a lower data rate gives more usable bits. Averaging helps random noise only.
6. Calibration and driftCalibrate offset and gain against known inputs, and let things warm up. Resolution doesn’t fix a wrong reading, only a coarse one.

Which ADC for which job?

Frequently asked questions

© Ramanu Systems · Precision electronics modules, Made in India · ENOB and noise-free-bit definitions follow common ADC-vendor conventions. Regulator figures from the TI TPS7A20 and AMS1117 datasheets; other noise values are typical, for illustration. Found an error? Tell us and we will fix it and credit you.