[time-nuts] Chooses for a desktop/server NTP external 1PPS reference

[Alexander Sack]

On Tue, Dec 8, 2009 at 7:10 PM, Hal Murray <hmurray at megapathdsl.net> wrote:
>
> pisymbol at gmail.com said:
>> Yes but in fact I would like to play with IEEE 1588 as well for work
>> reasons!
>
> Recently (month or two), somebody mentioned that one of the recent Intel
> Ethernet chips supports it.
>
> Another rathole I've avoided, so far.

Yes the next generation of Intel 10 GIGE chipsets do offer timestamp
registers for incoming packets.  I know a couple of partners building
cards around it.  However part of the issue is that very FEW companies
actually spend the time to design a timestamp framework that makes any
sense.  I can tell you today I spoke with one vendor who was expecting
to piggyback off of Intel's reference IEEE1588 design and believes
that is enough.  I have found that anything that involves an active
network connection to achieve timestamp accuracy is inferior to cards
that DO have higher quality crystal oscillators and can HARDWARE LATCH
onto a high quality 1PPS (relative to UTC) with zero network
transactions (no taking avergae deltas across several transactions
over a heavy utilized network).

There are some vendors out there that do just that (include a clock as
well as the ability to PLL against a 1PPS signal, though what's
interesting is that this 1PPS is not standardized across the industry
- DOES ANYONE KNOW WHY?  It kinda blows my mind.).  Of course these
cards are an order of magnitude more expensive than your commodity
Intel design.

-aps