Skip to content

Analog-to-Digital Converters

SAR ADCs: How Successive Approximation Works

Twelve yes-or-no questions, and every one of them has to settle before the next.

  • Lesson 2 of 2
  • Advanced

6 min read

Give a successive-approximation converter one microsecond and it does not measure your signal for a microsecond. It takes a snapshot in a few tens of nanoseconds, disconnects from your circuit entirely, and then spends the rest of the time interrogating that snapshot with a sequence of yes-or-no questions.

Almost everything surprising about SAR converters follows from that separation. The input behaves one way while it is sampling and a completely different way while it is converting.

Quick answer. A SAR converter performs a binary search. It compares the sampled input against half the reference, records whether it is above or below, then against a quarter, an eighth, and so on, one comparison per bit. Twelve bits means exactly twelve decisions, each of which has to settle before the next one starts. The input is only connected to your circuit during the acquisition window at the beginning.

The converter compares the sample against half the range, records the bit and repeats on the remaining half. Each decision halves the uncertainty, so exactly as many decisions are needed as there are bits.
Figure 1. One comparison per bit, no more and no fewer.

The figure follows four bits of a conversion. The first decision asks whether the sample is above or below half the reference, which splits the range in two. The second splits the remaining half. Each decision halves the uncertainty, so after N decisions the uncertainty is the range divided by 2 to the N, which is one LSB.

That is why a SAR needs exactly one comparison per bit, and why its conversion time scales linearly with resolution rather than exponentially. It is also why every extra bit costs the same again in time: the twelfth decision takes as long as the first.

It is a capacitor array, not a resistor ladder

The textbook drawing of a SAR shows a DAC and a comparator. Most real parts implement the DAC as a binary-weighted capacitor array that does double duty: it is both the sampling capacitor and the digital-to-analog converter.

During acquisition, all the capacitors in the array are connected to the input and charge to the input voltage. When the sample is taken, the input switch opens and the array holds that charge. The conversion then proceeds by switching individual capacitors between the reference and ground, which redistributes the stored charge and moves the comparator's input by a calculated amount. The comparator reports which way it moved, and that is the bit.

Three consequences that matter to the circuit outside:

  • The input capacitance is real and it is switched. The array has to be

charged from your circuit at the start of every conversion. Lesson 5 works out what that costs.

  • The reference is in the signal path, not just a parameter. Every

decision connects capacitors to the reference, so the reference supplies charge during the conversion and has to recover between decisions. Lesson 7 is about what that demands.

  • There is no averaging. The output is a measurement of one instant. Any

noise present during the acquisition window is in the answer at full strength, which is the opposite of how a delta-sigma converter behaves.

The time budget

A conversion has two parts, and they compete:

PhaseWhat happensWhat it needs
Acquisitioninput switch closed, array charges to the inputthe source must settle the array to within half an LSB
Conversioninput switch open, N comparisonseach decision must settle before the next

For the 12-bit converter this course uses, suppose the datasheet allows 300 ns of acquisition. Settling to half an LSB of 12 bits means settling to one part in 2 to the 13:

n = ln(2^13) = 9.01 time constants
τ ≤ 300 ns / 9.01 = 33.3 ns

Thirty-three nanoseconds of RC, for everything between your amplifier and the array. That single number is the whole design constraint of lesson 6, and it is why the obvious reflex of putting a large filter capacitor at the ADC input causes problems rather than solving them.

Note where the 9.01 came from: settling to half an LSB of an N-bit converter needs about ln(2^(N+1)) time constants, which is roughly 0.69 × (N + 1). A 16-bit part needs about 11.8, a 10-bit part about 7.6. More bits costs more settling time even when nothing else changes.

What it is good at, and what it is not

SAR converters sit in a particular place:

  • One conversion, one answer, no latency. The result corresponds to a

single known instant and arrives immediately. That is what makes them the default in control loops and in multiplexed measurement.

  • Easy to multiplex. Because each conversion is independent, you can

switch the input between channels between conversions. A delta-sigma cannot do this without discarding results while its filter refills.

  • Resolution has a practical ceiling. The capacitor array has to be

matched to the resolution you want, and matching gets expensive quickly above sixteen to eighteen bits.

  • It does not average for you. If you want the noise reduced, you have to

oversample and average in software, and then you have given up the latency advantage you came for.

That last pair is the trade the next lesson is about.

The failure that looks like noise

When a SAR gives readings that jump by far more than a code, the instinct is to blame noise. Often the real cause is that the acquisition window was not long enough for the array to charge, so each conversion starts from a slightly different residue of the previous one.

The tell is that the error depends on what the previous channel or sample was. A multiplexed system reading a 0 V channel right after a 3 V channel, with an input that does not settle, will read the 0 V channel high. Swap the order and the error changes. Noise does not behave that way.

What comes next

Lesson 3 takes the other common architecture, which solves the noise problem by spending time, and explains exactly what that time buys and what it costs.

Sources and further reading

Where to go next

¿Te ha resultado útil?

Reddit LinkedIn

The XKEMA newsletter

Hardware engineering worth keeping

New courses, engineering notes and hardware guides, whenever there is something genuinely worth reading. Nothing else.

We will email you to confirm. You can unsubscribe from any issue.

XKEMA, in one sentence

Engineering knowledge management software for companies that develop electronic products: the product structure, its bills of materials, its documentation and its revisions, all in one place.