[ntp:questions] Can NTP sync within 1ms

William Unruh unruh at invalid.ca
Sun Apr 27 15:52:47 UTC 2014

On 2014-04-27, Rob <nomail at example.com> wrote:
> Paul <tik-tok at bodosom.net> wrote:
>> I don't know what "terribly accurate" might be to you but in the real world
>> sufficient accuracy depends on the circumstance.
>> Someone should conduct an experiment.
> I am in a group that works on a project that needs synchronous audio
> on geographically distributed PCs.
> Of course there are several hurdles.
> First, we sync all machines to locally connected GPS receivers with
> PPS output.  We use ntpd and kernel PPS.  This is wellknown territory.
> In the ntpq -p stats this appears to bring the systems within 10us,
> often within 2us, of the PPS signal.  We still have to find out if this
> is reality or just output of a program.

reality. You can test the interrupt latencies on the machines (a few us) 

> The next problem is to send output to a soundcard and making it send
> a sample at the sampling clock edge closest to a specified time.
> (48kHz sampling rate corresponds to a sampling clock period of 20.8us)

It will certainly depend on the sound card. AFAIK most have their own
internal oscillators, that are not adjustable. I suspect you will have
to build your ownsoundcard to accomplish this. 

> When that has been achieved, it of course is easy to wire up two of
> those systems, place them closely together, and check on a scope.
> It could be further improved when we can somehow lock the sampling
> clock to the PPS signal, so another +/- 10uS uncertainty/jitter is
> removed.

48KHz is 20us. Why are you worried about a few us? What are you trying
to do?

More information about the questions mailing list