Hacker News Reader

Comment by fsh | Hacker News Reader

It is usually much simpler to put a GPS receiver in the device and to make use of the dozens of atomic clocks flying over our heads. Or to use NTP over the internet.

View on HN

You can have hundreds of ms of error with these approaches, which is hard to compensate for.

With GPS you can measure time with <100ns person.

You should tell that to meta, since they are putting atomic clocks into each one of their servers, which is, according to you, unnecessary, because same or better result can be achieved by just constantly syncing with GPS.

I think that's a fallacious argument. On the face of it, that they're not using GPS doesn't necessarily mean that it's because of a lack of precision. There could be other factors than precision that influence their choice.

You're assuming that Meta developed the Time Card [0] because GNSS or NTP aren't accurate enough. That's not the case, Meta tells us exactly why they developed it, in the "Why do we need a new time device" section of [0]. TL;DR: openness and software support.

And they're not putting one in every server, they're using it as a reference for their new NTP-based timing protocol [1], with which they showed an uncertainty of only a few hundred microseconds.

[0]: https://engineering.fb.com/2021/08/11/open-source/time-appli...

[1]: https://engineering.fb.com/2020/03/18/production-engineering...

The atomic clock on their card performs much worse than the built-in GPS receiver. In normal operation, it serves no real purpose. I guess some bored engineers wanted to have something fun to play with.

They're also working on the metaverse, so they've got some weird choices overall. That's it's meta doing it doesn't magically make it the best .

The big weaknesses of GPS are when you're accelerating and way up north or way down south where satellite coverage is poor

Usually the reason for including the atomic clock is reliability: the atomic clock can keep good enough time for a short period if the GPS signal is lost. Their time cards still primarily run off of GPS (in fact there are 3 nested clocks in the standard design of these systems: an oven-controlled crystal oscillator which is already quite stable by itself, but more importantly gives a very clean sine wave output, then an atomic clock (usually rubidium) which is more stable but higher noise which is used to tune the crystal oscillator, then the GPS signal which is even more stable with even higher noise used to tune the atomic clock. You can often just skip the atomic clock in the arrangement and you'll get the same performance while you have a GPS signal.)

Hundreds of milliseconds? A receiver like [0] is specced for single-digit nanoseconds. Even the standard-precision modules [1] are double-digit.

My understanding is that GPS wouldn't function if the timing was inaccurate to several hundred milliseconds.

[0]: https://www.u-blox.com/en/product/lea-f9t-module

[1]: https://www.u-blox.com/en/product/neo-m9n-module

Reply on news.ycombinator.com