The 0.5 Millivolt Offset: Three ADC Architectures and the Calibration That Decides
Archi5 min read
Contents
The 0.5 Millivolt Offset That Decides Which ADC You Buy#
On a bench in a small test lab, an engineer hooks a 10 mV signal to three analog-to-digital converters. All three are 16-bit parts. All three cost within 20 percent of each other. The first reads 9.6 mV. The second reads 10.1 mV. The third reads 10.0 mV but takes 40 times longer to produce the number. The difference is a few tenths of a millivolt, and it is the whole ballgame.
When teams choose an ADC, they usually compare resolution, sample rate, and price. The spec that actually decides whether the product works is input offset voltage and its drift over temperature. A 0.5 mV offset on a 10 mV signal is a 5 percent error before any gain stage. Calibration can remove it, but calibration costs time on the line and needs a reference you trust.
Three Architectures, Three Ways to Be Wrong#
The three common choices for precision work are successive-approximation (SAR), delta-sigma, and pipelined converters. Each handles offset and noise differently, and each makes a different trade.
A SAR converter is fast and simple. It compares the input against a reference one bit at a time, so a 16-bit part needs 16 comparison cycles. That speed comes with offset from the comparator and the internal capacitor array. Typical parts specify offsets in the low millivolts and drift of a few microvolts per degree Celsius. On a 10 mV signal, a 1 mV offset is 10 percent. You can trim it in firmware, but the trim has to track temperature or it walks away.
Delta-sigma converters trade speed for accuracy. They oversample the input and push noise out of the band with a digital filter. Offsets are often in the microvolt range, and the input-referred noise is low enough that a 10 mV signal sits well above the noise floor. The cost is latency: a settled reading can take tens of milliseconds, and multiplexing several channels means each one waits its turn. For a weigh scale or a thermocouple, that is fine. For a motor current loop, it is not.
Pipelined converters sit in the middle. They split the conversion into stages, so they run fast, but each stage adds gain error and offset that must be corrected. Many pipelined parts include digital calibration that runs at power-up. That calibration can remove most of the offset, but it assumes the temperature stays near where it ran. In a chassis that heats from 20 to 60 degrees Celsius, the residual offset can drift by a millivolt or more.
The Detail That Mattered: One Resistor#
In a recent comparison for a battery-monitoring board, the team measured all three architectures on the same 10 mV source with the same 2.5 V reference. The SAR part showed 0.9 mV offset. The delta-sigma part showed 4 microvolts. The pipelined part showed 0.2 mV after its internal calibration. On paper, the delta-sigma wins. On the board, the delta-sigma lost.
The reason was a single resistor in the input divider. The delta-sigma part needed a low-impedance source to hit its noise spec. The divider was 100 kOhm, which added thermal noise and a settling error that showed up as 0.6 mV of apparent offset. The SAR part had a much higher input impedance and did not care. The fix was to buffer the input, which added an op-amp and two capacitors and changed the board layout. That one resistor decided the architecture.
This is the pattern in most ADC comparisons. The datasheet offset is a starting point. The input impedance, the reference quality, and the layout decide the result. A 0.5 mV offset on a 10 mV signal gets worse when the source impedance is high, when the reference is noisy, and when the ground return shares a path with a switching supply.
How to Compare Them Without Getting Fooled#
Start with the signal, not the part. Write down the smallest input you must resolve and the error you can tolerate. If the signal is 10 mV and the error budget is 1 percent, you have 100 microvolts to work with. That budget must cover the ADC offset, the offset drift, the reference error, and the noise from the source. In practice, a 16-bit part with 1 mV of offset cannot meet that budget without calibration.
Next, check the input structure. A SAR part with a switched-capacitor input draws charge from the source on every conversion. If the source impedance is above a few kilo-ohms, the input does not settle, and the reading is low. A delta-sigma part with a similar input has the same problem, plus its digital filter needs time to settle after a channel switch. A pipelined part often has a differential input with a defined common-mode range, which can be easier to drive but leaves less headroom.
Then look at the reference. A 2.5 V reference with 0.1 percent initial accuracy and 10 ppm per degree Celsius drift contributes 2.5 mV of error at 10 degrees Celsius. That is larger than the ADC offset you are worried about. Many teams spend weeks choosing the converter and ignore the reference, which is the bigger error term.
Finally, plan the calibration. If you can calibrate at one temperature, you remove the offset but not the drift. If you can calibrate at two temperatures, you can estimate the drift, but only if the part's drift is linear. Some SAR parts have a drift curve that bends, so a two-point fit leaves a residual. Delta-sigma parts often have lower drift, which makes the calibration simpler.
The Choice, Stated Plainly#
For a slow, high-impedance sensor with a 10 mV signal, a delta-sigma converter with a buffered input and a good reference is the safe pick. For a fast, low-impedance signal where latency matters, a SAR part with a factory trim or a one-point calibration is usually enough. For a high-speed pipeline where the signal is large, a pipelined part with power-up calibration works, but check the drift over the temperature range you will actually see.
The 0.5 mV offset is not a footnote. It is the number that tells you whether the part fits the board. Compare the offset, the drift, the input impedance, and the reference together, and the architecture choice becomes clear. Compare only the resolution and the price, and you will find the problem later, on the bench, with a soldering iron in your hand.
More articles