On July 31st 2017 at 2:31:00am CEST, I synced the time on a DS3231N to GPS via NTP (+/- a few microseconds). I had previously tuned the clock by setting the offset register to 0x09 and verifying the stability against my Trimble Thunderbolt using an oscilloscope.
At the time I set it, my oscilloscope-and-Thunderbolt measurements indicated it had a short-term stability (over the course of a few minutes) of 1.38 ppm, which is within the +/- 2 ppm specs. I then removed the module (it had a coin cell battery backup) from the NTP server Raspberry Pi that had set it and transferred it the Raspberry Pi I have setup as a strictly “offline” system to store some PGP keys away from the prying eyes of the internet.
Since this system would never see the internet, having an accurate RTC meant that I wouldn’t need to set the system clock from my wristwatch when I turned it on — that’d be inconvenient.
Slightly over three years later on August 12th 2020 I turned on the offline Pi and logged in via a serial link (no network connectivity at all) to the terminal. At 11:18:47 PM PDT as measured by my GPS-backed NTP server, I compared the time between the NTP server and the Pi. Adjusting for time zones, the offline Pi reported the time as 11:19:13 PM, or 26 seconds fast.
95,838,467 seconds of actual time elapsed between the two measurements and the clock only gained 26 seconds. That’s a long-term stability of 0.27 ppm.
I’m impressed: considering the clock was running off a CR2032 battery as opposed to regulated power from the Pi (which had not been plugged in at all during that time). During this time it had undergone several flights and car journeys as part of an international move, and had not been in any sort of specifically-regulated thermal environment (e.g. ordinary residential rooms).