Overclocking the ATtiny 2 ADC. With graphs and stuff.

[burnsb]

Per the data sheet, the ADC on the tiny 2s (and I would assume the other post Microchip AVRs) is rated for 375ksps max. However, if you set:
ADC0.CTRLB = 0; // prescaler div/2
ADC0.CTRLE = 0; // sample duration of 1/2clock 0.25us @20mHz

You can get an ADC rate of around 615kHz. Which gives a 300kHz or so max frequency of a sampled wave given Nyquist, not that I'm any sort of expert on Nyquist.

So far, I have looked at 125kHz triangle waves. (not very triangular, my function generator is a AD9833 chip ran through a LM358 op amp with a couple of pots to adjust amplification and offset. I have also looked at a 10k pot. I think that the pot measurements are probably going to be the most useful, and to those ends I've ordered a pot assortment from 100ohms and up. I imagine that values from 10k and lower will be most useful as I expect impedance to quickly make any results past 10k not very useful.

My tests have been done on the same ATtiny3224. I'm using a Rigol DS1202Z-E scope for my reference measurements. My power supply is just the 5V rail off a Uno clone that is running the AD9833. So, these are definitely less than optimal test conditions/equipment. However, I'm getting meaningful results.

Also, I'm looking at measuring SMPS output, so impedance isn't going to be a big problem anyhow. My application will also tolerate errors of several percent, probably 5% error won't be too much.

The sketch running on the ATtiny 3224 takes 100 freerunning samples as well as count from TCB1.

I used a 125kHz wave as the lm358 couldn't keep up with 300kHz and give me the amplitude I wanted. A typical result (2990mV avg with 1010mV peak to peak):

And here are the results tabulated with a nice curve representing error:

So, 1v is the sweet spot. It measures less than 1% deviation around 1V. This is also backed up by the test series on just a 10K pot. (more on that in a minute). At 3000mV error has grown to more than -17%. Strangely, at levels below 1V, the ADC error quickly grows in a positive direction. I.E. it reads higher than what it's measuring.

Anyhow, A 10k pot gave results that were remarkably consistent with the triangle wave error. Here's a typical series, 2500mV:

The following results page has two charts, the larger chart is results from 1V and up, At 59mV the deviation was 360%. The smaller chart shows all the results and you can see it going to the moon as the voltage gets small. So, I think I can pretty conclusively say that an overclocked ADC isn't going to give you usable results below 1V. Sample duration maybe?

Now, a flat 3v voltage gives an error of -18.6%. Pretty close to the -17.6% of a triangle wave with an average of 3V. With that wave being 1V peak to peak, I bet it accounts for the 1% difference as error is growing more negative as the voltage increases.

If anyone wants it, I can share the sketch used for sampling or my spreadsheets. I think I'm going to test a 100ohm pot at 1/4V increments. I may do some other series on higher value dividers.

I think it will be possible to create a lookup table for ADC correction at max overclock and get usable results if the source is low impedance and high accuracy isn't needed. I'll probably find the curve that best fits the data and generate the table from it, doing math with only 28 cycles is probably not going to work. If you want to use this data at these speeds, it needs to be simple and quick.

I welcome questions/comments/suggestions!

[burnsb]

Pots came in. When taking initial measurements, I realized the 5V power from a Arduino clone was just too noisy for this. Ended up using a LM317 and a 9V battery. While it's very dated and inefficient, it was what I had onhand that would do the job. Produces voltage flatter than the noise on my scope. There's 20-40mV of noise that I can't get rid of.

However, it was plenty flat for these measurements. With a better power source and a 100ohm pot, each series of samples showed very little variation. They also repeat for for a given voltage. For instance, with the divider set to 2.5V, I get 10mV of variance in the samples. Many of the samples show even less. (Not including the first sample, it's 20mV or higher than the rest of the samples. )

I am discarding the first sample from each series. I assume it's high as a result of the capacitor in the ADC having longer to charge as it is initialized.

The results from 1/4V increments with a 100ohm pot:

As I have time, I'll take a look at some higher value pots. While the 10k ohm pot results aren't as clean, they aren't terribly different from the 100ohm. Quickly comparing without doing any math, the results are within 3% of each other.

This is only one chip's ADC. I don't even have another one soldered up to try it out on a different chip of the same model. I'll release source and schematic soon. I'd really appreciate it if others would run some test series. I'm curious as to how consistent this is.

[burnsb]

Took a look at a 500k pot. As expected, I don't think its very usable, however, it was instructive. Results:
.

.
The results aren't really the interesting part. The initial sample that is 'high' that I've been throwing away was now several samples long and around 100mv:
.

.
It was the worst at 2.5V, right in the middle of the pot with 250k of resistance on either side. While I can check the voltage of the pot setting with a multimeter, I have the scope and it's a nifty toy. I noticed a drop during sampling:
.

.
Not sure if my terminology is technically correct, but you can basically see the impedance as well as the capacitor in the ADC charging and discharging. The voltage drops about 270mv during sampling. Each little sawtooth is the ADC sipping current each cycle.
.

.
Now, I don't think this actually does anything for calibrating an overclocked ADC. However, I thought it was pretty freaking cool that you could actually see an effect like this.