Is that an NTP synchronized clock thats synchronized to a GPS 1pps via kernel discipline, or is it running off the GPSDO 10 MHz?
They would likely be different it sounds -
With the caveat that I have still been too lazy to recompile my Raspberry Pi's kernel to get an extra two orders of magnitude with my plain GPS time server, I fully do expect that when i do, I will get much more accurate time from my GPSDO and its 1PPS than I do from my GPS. Even though its a good one.
However, lacking the so called kernel discipline its still accurate to better than a microsecond so I cant complain.
NTP connected clocks do have error. Various kinds of wander occur and can be characterized. So what you see is / are likely the combined effect of a number of different sources of error common to computers in that situation. Including clock and network latency. Thats far more likely.
I think "Chrony" is better than NTP in that respect. I'm planning to hook my GPSDO up to a machine I have that has a real ethernet card not a USB one, and real GPIOs and I'm also planning to switch my RPI to Chrony one of these days. Then I'll have two or when my desktop is running three sources of network time to serve as sanity checks for one another.
Did anyone ever do a comparison between their bg7tbl gpsdo and a different reference to see if the output frequency is 10MHz? I'm using mine (version 2017-03-29) to feed a clock (date + time). Because that clock wanders compared to an ntp synchronised clock I measured the output frequency and it turns out my bg7tbl gpsdo is 1.5Hz too low! Needless to say this thing starts to look more and more like a waste of money.