# Measure sensitivity of ADC

I would like to benchmark my ADC.

I can calculate a sensitivity value based on the information in the datasheet. However, I would like to measure/benchmark this figure.

• Is there a standard method for doing this?
• How would you go about measuring accuracy of any ADC?

Considering this is a 24bit chip that can measure down to few µV, it is not very straight forward to plug in a signal generator.

• Would high ratios of voltage divider help to generate a low noise low voltage signal in µV range to feed into the ADC?
• Wouldn't Johnson noise stop one from using high value resister in this case?
• You can make multiple measurements to characterize the noise, and compare that to the expected Johnson noise. You're still going to be chasing your tail as to whether the problem is the ADC or your circuit, but IMHO simple is better, when you can. Commented May 30, 2019 at 15:27
• Yep, also, not many companies produce 24 bit ADCs – and you're already aware of the few µV resolution being dominated by thermal noise – so, I'd start out by trusting your datasheet, especially in terms of ENOB / noise floor. If you have a problem with the performance, I'd strongly suggest looking at other sources of noise or nonlinearity in your circuit first. How noisy is the power supply, how well decoupled the ADC? What do you do for ADC buffering, and what's the noise figure of the active device you use for that? Do these numbers up to give you the same noise floor as your ADC measures? Commented May 30, 2019 at 15:31
• oh, and not to forget: you should really define what "sensitivity" means for your overall application, not just the ADC component! I can't think of many use cases where one would even make use of a 24 bit ADC resolution easily; many overall applications get many more bits out of the data coming from the ADC than the ADC has ENOB, because: processing gain! Commented May 30, 2019 at 15:36
• Feed the ADC an appropriate sinusoid, divided down by 100,000:1. By stimulating only 10 or 20 codes, the "distortion" will be interestingly detected. Use 62 ohm to ground (has 1nanoVolt/rtHz noise density), preceeded by 4.7Meg + 1.5Meg ohms. Establish signal bandwidth with a large capacitor across 62 ohms. Use a Ground plane. Place a metal shield above the 100,000:1 divider, and tie shield to the Ground plane. Keep black-brick switching supplies far away. Given the output code stream, examine the distortion, and the noise. Realize some "distortion" may be tonal input noise, aliased down. Commented May 30, 2019 at 15:58
• Connect a parallel RC to the ADC inputs, charge the C, then let them run down. This will generate a very low noise, low interference, non-linear ramp with a very predictable rate. Use to find missing codes, integral linearity, differential linearity, idle tones (as it's a sigma delta), and how noise varies with level. Commented May 30, 2019 at 17:01

Is there a standard method for doing this?

Yes (each manufacturer has slightly different methods), if you want to measure noise, then ground the inputs. INL and DNL will require a much more involved setups (shown below). Most methods involve a signal generator with a ramp signal. The ramp signal is compared to the output of the ADC and the INL and DNL are calculated.

How would you go about measuring accuracy of any ADC?

Depends on the parameter that needs to be measured and the application. In my case, I mainly care about INL. I use a 24bit ADC for a control application, INL can create error in the feedback loop. If you need absolute precision, in my opinion INL is one of the better things to look at. If relative precision is important DNL might be a better thing to look at. Some applications require low noise.

But you don't need to measure the accuracy of the ADC, you only need to understand how the manufacturer measured/tested the performance as it is guaranteed in the datasheets of most manufacturers. If you design with an ADC and don't get the performance listed in the datasheet, that would be a good point to talk to the manufacturer.

If I did have to measure the performance, I would go to the manufacturer and get the application note for testing ADC's and use their methods and circuits.

Would high ratios of voltage divider help to generate a low noise low voltage signal in µV range to feed into the ADC?

If you want to design a uV to nV system, the noise of all components needs to be considered, this means doing a noise analysis. A good resource for this is Noise: three categories: Device, Conducted and Emitted. Noise bandwidth and filtering are essential concepts to understand how to attain low noise.

Wouldn't Johnson noise stop one from using high value resister in this case?

A 24-bit ADC with a 5V span has 298nV per bit, A 500k resistor with a 5Hz low pass filter has 201nV-rms of noise. That same resistor with a 500Hz filter has 2uV-rms of noise. So filtering becomes important.