IMO, I have mostly seen mislabeling, rebinning, and passing off obvious QC rejects.

example from many years ago: https://www.youtube.com/watch?v=e6DfBuPwAAA

1. https://righto.com

That the ADS1115 costs <$1 on LCSC means they buy millions from them every year. They are one of the biggest trustable players in Asia.

I have access to our internal STM32 pricing. You'd be shocked.

(also interestingly the STM32 clones I've seen had stacked die flash because they didn't fab them in a technology that could also do flash, so you can easily tell the counterfeit from sanding down the package and looking for an extra set of bonding wires; it's also a cool place to access the internal flash bus if you wanna bypass some readout protection :) )

https://youtu.be/upTgM_S5rAQ

And if you want to see CERN's multimeter, check this out.

https://youtu.be/D28uSzCs7-k

Marco Reps is a treasure.

There are a bunch of reasons but the primary reason is that good ADCs are made using a different mixed signal process than microcontrollers. MCU ADCs are capacitive charge-balancing successive-approximation type which limits their sensitivity and precision.

Standalone ADCs also eliminate significant sources of noise like temperature fluctuations and electronic noise (the digital logic on the chip often runs at less than 1Mhz for example)

The chip has a 12bit SAR ADC. Layout and board design mattered a lot, but even the worst ones had 10 bits worth, and the best one had nearly 12 bits effective.

That was without doing too much on the software side, meaning the other modules weren't running, besides a single serial output. On the bad boards the serial affected it, but on the good board very little.

Also, and perhaps more importantly, the test rig is a lot simpler and a lot cheaper if you can generally trust manufacturer data. Sure, send off a few samples (likely prototypes with parts from Digikey instead of LCSC) to run extended testing in an environmental chamber with thermal imaging, build an endurance test rig that pushes the button once a second for four weeks to simulate once-daily use for years, whatever you want to do...but after that, if TI says it's good from -40 to +125, you're going to trust them on a lot of the edge cases.

Do 100% testing of the things you can test in-circuit if you can - power it up at room temperature and make sure it works once - but that doesn't mean you get the actual rated performance across all published environmental conditions.

I recall reading about a project at CERN to design a 12bit ADC chip that could sample at tens of GHz, maybe 50 or more.

I was perplexed at how they could achieve this.

Turned out it was the same we programmers do. Parallel processing.

They had taken a 12bit SAR unit which ran at like MHz rates, and just cloned it many times. They then had a large analog multiplexer in front to route the signal to the active ADC unit in a round-robin fashion.

That takes a lot of chip real-estate, and the analog muxer had to be carefully designed.

For a simpler approach to speed there is Flash ADCs[1], which kinda brute-force it.

For precision I know multi-slope ADCs[2] are often used.

Sadly I don't know much about the history, and would also love to learn more about it. Bound to be some fascinating stories there.

[1]: https://en.wikipedia.org/wiki/Flash_ADC

[2]: https://www.analog.com/media/en/training-seminars/tutorials/...

What about the AD9226? It only has a single channel but can do up to 65 MSa/s at 12 bits. I bought one as a module for around $12 on AliExpress to experiment with software decoding of analog video. I only run it at 20 MSa/s and only use 8 bits because, funnily enough, the limiting factor is the speed at which I could get the data into my laptop. I connected it to a Raspberry Pi Zero and use the SMI peripheral as described here: https://iosoft.blog/2020/07/16/raspberry-pi-smi/

They need a lot of pins to be toggled. Otherwise they spit out no data.

And a lot of manual stuff means it's super DMA unfriendly. And you need DMA for high-speed stuff.

Natively/internally, it runs at 860 samples per second, and you can configure it to provide that data at a lower sample rate at lower noise levels by averaging multiple readings together internally.