Variations on XTAL clock frequency

Is there any published research/report about how much "uncertainity" and/or
variations must be expected on a PC clock frequency ?

 

[X-posted without a F'up2 concentrator. Fixed.]

I doubt there'll be much about PC clocks in particular --- but I'm
sure there's plenty about crystal oscillator stability in general,
which will apply to PCs quite seamlessly.

 

What you are asking would be dependent on the grade crystal materials that
they select for manufacturing them. The crystals have a thermo, and internal
pressure coefficient factor to deal with. They are also a bit voltage
sensitive, as in part of the feedback circuit employed in the design of the
oscillator that they are part of.

For home computers, they do not need to use an expensive high stability type
crystal, as such that is used for precision instrumentation. I would not be
surprised if the clocking frequency in a home PC machine is drifting about
1% to 2%. As long as everything keeps properly synchronized there will be no
problem, no matter what the master clocking frequency is.

The time keeping crystal in a PC is different from the one used for the main
system. The actual real time clock is a separate operation. It is read by
the main system, only at the times where it needs to get the time of day,
and the date data. The time of day accuracy of most computers is about the
same as any low cost quartz watch. I found the time of day on most computers
to drift as much as several minutes a month, if not corrected. A typical
Timex or Casio watch can do better than 15 seconds per month.

There are softwares available to re-set the clock automatically from some of
the various time standard services around the world. I believe that XP comes
with such a software. This can be done over the internet.

If you have the budget you can install a GPS time standard system, and
install the hardware and software in your computer to work with it. The GPS
antenna would have to be installed at a location where it can clearly see
the sky to receive the GPS satellite data. This type of installation would
result in the most possible accurate time of day standard for a PC computer.

--

Jerry G.
=====

<Sunwaesh> wrote in message
Is there any published research/report about how much "uncertainity" and/or
variations must be expected on a PC clock frequency ?

 

Jerry G.,

Thank you for your valuable answer.
In a project I need to have time synchronization between a set of computers
where some of them are networked together on a LAN (no internet) and some
others are running stand alone. I am planning to use the
one-pulse-per-second (1PPS) signal from the GPS receivers. The networked
computers will have one GPS receiver and all the other stand alone computers
will have their own GPS receivers. GPS receivers will generate 1PPS signals
to interrupt the computers to set their internal time clocks. Applications
will use the computer timer (get the time of the day). I want to model (some
how, but I do not know how) the probable variation that a computer clock may
have between 1PPS signals.

Would anyone comment/argue/recommend/suggest/propose how one can model the
variation on a PC clock frequency ?

Regards,

 

What you need is not a 1 pps, but is to have the proper decoder to decoded
the GPS time signal, and then feed the computers with a true time reference.
There are Ethernet, and serial interfaces for this. You should get in touch
with a manufacture of this type of equipment to find out about the necessary
software, and matched GPS system for this.

Using a simple pps is not good, because of one computer jumps off time, it
will not be corrected. It will stay the way is until it is manually
corrected, or it may even jump out again. Sometimes the clock in a computer
goes off time, because it hit a program glitch. The real time clock will not
reset the computer until it is re-started, or until someone comes along and
re-sets the clock.

There are two companies that I have dealt with for the hardware and
software's to do what you want. There is Leitch Video
http://www2.leitch.com/Custserv/products.nsf/WP/Reference , and Torpy Time
http://www.torpeytime.com/main.htm

For our clients, I have used products from both of these manufactures for
time keeping, and have had great satisfaction. We also use many of the
Leitch broadcast systems products, and they are also excellent.

--

Jerry G.
======


<Sunwaesh> wrote in message
Jerry G.,

Thank you for your valuable answer.
In a project I need to have time synchronization between a set of computers
where some of them are networked together on a LAN (no internet) and some
others are running stand alone. I am planning to use the
one-pulse-per-second (1PPS) signal from the GPS receivers. The networked
computers will have one GPS receiver and all the other stand alone computers
will have their own GPS receivers. GPS receivers will generate 1PPS signals
to interrupt the computers to set their internal time clocks. Applications
will use the computer timer (get the time of the day). I want to model (some
how, but I do not know how) the probable variation that a computer clock may
have between 1PPS signals.

Would anyone comment/argue/recommend/suggest/propose how one can model the
variation on a PC clock frequency ?

Regards,

 

You should have mentioned your project in your first post.

For the networked PCs you can use Network Time Protocol. Run an NTP server
on your time standard PC and run clients on all the others. This learns the
error rate of the PC clock and gently slews the time of the client to keep
it in sync. This ensures that the time is monotonic (always goes forwards)
and doesn't have discontinuities. It also has the great advantage that you
just have to install it as everything has been written for it already.

For the standalone PCs the simplest model that I can think of is that the PC
clock runs at fairly constant but slightly incorrect speed. When you have
measured the rate of time drift you can estimate what the offset you need
for the correct time. You will have to average it out over a long time to
get an accurate value. You should always be measuring the drift so that you
can adapt to change in the rate of drift. This should be stable enough for
anything that a PC needs to do.

Peter

 

I'd consider installing a NTP service / daemon. It is able to slowly
adjust the clock division ratio so that the timing stays put. Even
without an external server, NTP is able to use your GPS 1pps tick
as the reference clock.

For more information, Google for NTP or Network Time Protocol.

NTP is also excellent for keeping your networked computers in sync.

 

The specifications of the crystal will tell you.

That would be a gigantic drift for even a very cheap crystal. I think
1/10,000 is more typical of what would be a large drift from a low-quality
crystal.

 

Jerry G.,

For those computers which are LAN'ed together, you should certainly
consider running NTP. The commonly-used "ntpd" daemon for Unix and
Linux and etc. will handle the inter-system coordination, and also has
drivers for most GPS and similar external clock systems (including the
PPS input).

The PPS pulses are infrequent enough, and subject to enough jitter,
that the drivers will need to do a fairly significant amount of
low-pass filtering before using the pulse information to calibrate the
internal clock.

As I understand it, you're going to be dealing with at least two
separate "clocks" per PC.

One is the on-board date/time-of-day clock chip, which is typically
driven by (or drives) a 32.768 kHz quartz low-power "watch crystal".
These usually seem to have accuracies in the 10-15 seconds per month
range, like a cheap wristwatch. This part of the hardware is designed
to provide a coarse setting of the date and time when the system is
booted, and as I understand it the interface to the chip usually does
not provide a way to read or set the time to any precision greater
than +/- 1 second.

The other is the system's main CPU or bus oscillator, which is divided
down and generates interrupts at a predictable rate (many per second)
and/or is used to run a high-speed counter within the CPU itself.
This is also a quartz-crystal oscillator. It has a higher readout
precision than the clock chip, but you have to be careful about using
it... if you try to track the time-of-day by counting clock interrupts
(as I believe Linux does) you can "lose time" if another device
driver, or the BIOS itself, locks out interrupt processing for more
than a few milliseconds.

These two clocks/oscillators are not correlated with one another as
they're driven by separate quartz crystals. Both are subject to error
and drift due to temperature changes, but there's no guarantee that
the two crystals have identical or even similar temperature
coefficients of change.

The Unix "ntp" daemon software is able to estimate a given system's
amount of clock error and keep a record of the amount of drift (in
parts per million), and will load this value and use it to tweak the
clocking when the system is booted.

As to modelling the error you see on a PC's clock: you're going to
have to deal with several sources of error. To a first approximation,
you can consider the PPS data to be "short-term jittery, but
long-term stable". The 60 Hz powerline frequency is similar...
jittery in the short term due to noise, somewhat drifty over the
course of a day, but quite stable in the average over the long term.

The PC's quartz crystal oscillators are probably at about the opposite
end of the spectrum - quite stable in the short term, with a fairly
constant amount of error (in PPM) in the long term, and some amount of
temperature-related drifting around in the middle.

You might find it useful to review the information on Brooks Shera's
page at http://www.rt66.com/~shera/index_fs.htm - he discusses the
construction of a system which uses a GPS receiver's PPS signal to
discipline a high-stability quartz crystal oscillator.

 

Hello Sunwaesh,

If it is a modern PC or laptop keep in mind that clock frequencies are
often purposely dithered. That way the spectral energy spreads and the
manufacturer often gets around an EMC nightmare.

Regards, Joerg

http://www.analogconsultants.com

 

Not true. You can buy a GPS receiver that produces 1 PPS with hardly
any jitter at all. By using a PLL, a high-quality frequency standard
can and has been made. You don't need a lot of low-pass filtering in
that PLL.



-Robert Scott
Ypsilanti, Michigan
(Reply through this forum, not by direct e-mail to me, as automatic reply address is fake.)

 

The PPS pulses are infrequent enough, and subject to enough jitter,

Well, from what I've heard and read... it depends.

For one thing, different GPS receivers have different amounts of
jitter in their PPS pulses. Some (e.g. the Oncore series) are quite
good, while others have a larger amount of variation in timing. This
seems to be due to software/firmware design in the GPS receivers'
controllers. In his discussion on GPS-disciplined oscillators at the
site I mentioned, Brooks Shera notes that many GPS receivers' PPS
pulses are too jittery to use as a source for PLL-disciplining a good
oscillator.

For another thing... most PCs don't have special-purpose PLL or other
capture hardware to "catch" the exact timing of the PPS signal.
Simple approaches usually seem to involve hooking up the PPS line to a
parallel-port pin, to a serial-port CD pin, etc. and generating an
interrupt. The interrupt service routine then captures the
processor's high-resolution clock value. With most operating systems,
there can be a large amount of jitter in the time needed to enter the
ISR, depending on what other interrupt or kernel activity is taking
place. If your PC is down deep in the network card's interrupt
service routine and is handling heavy amounts of packet traffic, it
could require many microseconds, or in some cases a millisecond or
more, to enter the serial/parallel port ISR and sample the CPU clock.
This source of jitter is likely to be much greater than the jitter in
the GPS PPS pulse itself.

Since the original poster asked about the timing error "between PPS
pulses" and didn't mention the use of special-purpose low-jitter
capture hardware, my guess is that this CPU/OS-induced jitter may very
well be relevant to the problem.

 

I agree. I tried to make a high-quality frequency standard using a
Sandpiper GPS receiver. The 1-PPS output was very low jitter on one
edge but very high jitter on the opposite edge. Probably a flip-flop
was being reset by software. I started out using the wrong edge and
couldn't understand why my loop was so unstable. When I switched to
the other edge, then everything worked out fine. This was
special-purpose hardware, not just a digital input pin on a PC, so I
agree also that unless you have special-purpose hardware, you really
can't do much with that low-jitter pulse.


-Robert Scott
Ypsilanti, Michigan
(Reply through this forum, not by direct e-mail to me, as automatic reply address is fake.)

 

The Answer is that without knowing a whole lot of Factors (temperature of
the crystal oscillator, specs of the oscillator etc) it is not easy to model
or estimate the accuracy of a PC's internal clock beyond 'adequate for most
scenarios' If accuracy is required a better clock is required.

The Clock source from the GPS (if properly locked and implemented by the GPS
manufacturer) should be close to a Cesium clock in terms of accuracy and
will be several orders of Magnitude ( around 1 part in 1*10^12) better than
the clock from the PC which will be around 100 Parts per million if the
Temperature is kept constant at around 25c

A few links FYI
http://www.beaglesoft.com/stsystarcardspecs.htm#StarSync
http://www.gpsclock.com
http://www.aelcrystals.co.uk/

 

The battery-powered RTC clock is frequently VERY inaccurate, though.
For reasons I've never had explained to me, they almost always have a
trimcap on that xtal. Their Q/C processes to set that trimcap appear to
be FUBAR.

 

What trimcap? I've never seen one in a PC RTC circuit and, golly, does
it need it!

 

Maybe your motherboard collection differs from mine. But every board
I've owned (except some of the laptops) has had a trimcap.

Mystery why they bother to put such fine-tuning on there, since it is
very common to get as much as 10 seconds drift per day!

 

I would not have expected an un-trimmed Crystal to drift more than 100 Parts
per Million or 86.4 mS per day, the Laptop I am using at the moment is still
within a Second of the Talking clock and I set it 3-4 Months ago.
I am sure I would have noticed if Clocks on any PC's I used were losing or
gaining 10 Seconds a day

 

You're fortunate...
I've only had one laptop out of 4 that didn't drift in time too much.
My current one loses about 4 seconds per day., my previous one
gained about 2 seconds per day.
The PC battery-backed-up clock system was 'adequate' when it was
created back in the mid 1980's, but unfortunately that part of
the PC architecture hasn't changed much since then.

 

"Richard Freeman"

** 100 ppm = 1 part in 10,000.

A day has 24 x 60 x 60 seconds or 86,400 seconds.

So a 100 pm error is 8.64 seconds !!!

You are only 1000 times out.

** Sure - drift is nearly all due the tempco of the crystal.

If the AVERAGE temp value is steady from day to day then there is virtually
no DRIFT.

There may well be fixed setting error though.

** The tempco drift is only a few ppm *per degree C* for most crystals,
over the range from 10 to 35 C.

The indoor temp averaged over a month varies only a few degrees compared
with the previous or next month.

So, the month to month variation should be maybe +/- 10ppm or 0.86
seconds.

Seems about right to me.



................ Phil