What's the highest precision achieved for an ADC?

I was browsing Digikey the other day (don't you?) and I stumbled across some 32bit ADCs, there were offerings from Linear, TI and Analog. One stood out, the AD7177 from Analog which states in Table 7 on page 19 of the datasheet that at 5 samples per second it's got a staggering 27.5 effective number of bits (and an RMS noise of 50 nanovolts). On the other hand of course, it's accuracy is significantly worse, but still.

This got me wondering, if a relatively cheap off-the-shelf ADC can hit an ENOB of 27.5 bits...

What's the highest ENOB ever achieved? Be it in some super integrated IC, some piece of stupidly expensive lab gear, a lock-in amplifier? Has anyone ever beaten 27.5 bits of precision?

 A bit of clarification I'm not looking to buy/build or otherwise aquire such a device I'm just curious what the current state of the art is, modern atomic clocks have hit 3x10-18 (3 quintillionths) uncertainty, where do modern scientific voltmeters sit on the scale?

• @PlasmaHH The fact that modern atomic clocks are measured to parts per quintillionth would imply that there are uses for insanely precise devices.
– Sam
Dec 12, 2016 at 20:38
• What kind of sample rate do you need? Precision is unlimited if sampling time is unlimited. Keep integrating for an hour and the result is pretty accurate.
– PkP
Dec 12, 2016 at 20:38
• @PkP I'm not looking for any sample rate, I'm just curious what the record is. Although presumably in reality there comes a point when you're limited by the device itself?
– Sam
Dec 12, 2016 at 20:40
• 'effective resolution' (the 27.5) is not the same thing as ENOB. Short the input, measure RMS noise, divide by full scale voltage range. Dec 12, 2016 at 20:52
• @SpehroPefhany That's how they say the value was generated "The numbers given are for the bipolar input range with an external 5V reference. These numbers are typical and are generated with a differential input voltage of 0 V when the ADC is continuously converting on a single channel." So 5V reference, 50nV RMS noise, that's 100 million to 1 (ok, so that's 26.5 bits but still)
– Sam
Dec 12, 2016 at 20:58

Definition from Wiki: -

Effective number of bits (ENOB) is a measure of the dynamic range of an analog-to-digital converter (ADC) and its associated circuitry. The resolution of an ADC is specified by the number of bits used to represent the analog value, in principle giving 2^N signal levels for an N-bit signal

Quote from Atmel: -

In most cases 10-bit resolution is sufficient, but in some cases higher accuracy is desired. Special signal processing techniques can be used to improve the resolution of the measurement. By using a method called 'Oversampling and Decimation' higher resolution might be achieved, without using an external ADC.

Oversampling - take 4 consecutive samples and combine them to get one more bit of resolution; take a fairly standard 18 bit ADC and oversample by 256 to get a 22 bit ADC. Oversample by another 256 times to get a 26 bit ADC...

Do you see where this is going?

If noise is present and causes dithering of the signal, you can make any ADC have one extra bit by averaging/decimating 4 samples so, average as many as you like to get a higher resolution but clearly the price to pay is proportionally lower bandwidth and accuracy.

What's the highest ENOB ever achieved?

What do you want it to be?

Footnote - a sigma delta ADC does exactly what I've described above except, it manages out of band noise much better and therefore gets a better yield on increased bits per converted samples averaged (or decimated).

It only uses a 1 bit ADC (a comparator) so clearly this technique works but it doesn't have to use a 1 bit ADC. It's all about noise filtering: -

The noise from a sigma delta ADC is progressively higher at higher frequencies due to the use of an integrator in the signal path - this forces noise to be low at low frequencies and, after decimation this yields a net benefit in resolution compared to just a conventional ADC that has been over-sampled and decimated.

• Unless there is a typo in TI docs, probably an exponentiation was lost in copy-paste: in principle giving 2N signal levels for an N-bit signal. 2N should be 2^N. Jan 8, 2017 at 9:11
• @lorenzo well spotted. Jan 8, 2017 at 9:53
• @sam are we done with this question and answer now? Mar 4, 2020 at 23:50

It's possible to get around 32-bit effective resolution with an LTc (Low critical temperature) DC SQUID (Superconducting Quantum Interference Device) using well-executed hybrid digital/analog techniques. A few $\mu\Phi_0$ RMS noise and, say, +/-10000 $\Phi_0$ range gives 32 bits with 1Hz bandwidth and 0.3Hz corner frequency. Actual critical current is at least an order of magnitude higher so a few more bits might be possible.

Good for making picovoltmeters and such like. Kind of expensive and inconvenient because of the 4K environment.

• Sorry, but what do you mean with that uppercase PHI greek letter? Is it a typo or some constant related to the specific application? Dec 12, 2016 at 21:11
• @LorenzoDonati It's a constant- the magnetic flux quantum = h/2e where h is the Planck constant and e is the charge on an electron. Dec 12, 2016 at 21:13
• Yep! I was following the links in the article you linked to. I just found it here: Magnetic flux quantum. Ouch! Tough stuff! :-D Dec 12, 2016 at 21:14
• This surpasses the Turbo Encabulator's capability by varying the swaging angle of the dingle-arm and rendering the tremie pipe obsolete.
– user98663
Dec 12, 2016 at 21:21
• @Wossname The hardest part was aligning the hydrocoptic marzel vanes inside the cryostat. Dec 12, 2016 at 21:23

Texas Instruments has an ADC with 31-bit resolution, ADS1282, with up to 4000 samples per second, in industrial temperature range (-40 C + 85 C). Just ~\$40 in qty. 1000. However, one needs to work really, really hard to get the analog front-end noise down to that level of resolution, although some sliding averaging might help at the expense of sampling rate and/or bandwidth.