I am well aware that ntp needs 3 or more sources of time. In the experiments with multiple GPS sources, there were also several network time sources defined so ntp always had at least 4 sources to choose from. Those sources were a mix of other GPS ntp servers on my 1G network and my ISP's ntp servers which are only a few hops away.
I think I said you only need one GPS and if you have two, put one on another machine. It was an experiment to see what happens.
As the post was about using cheap GPS receivers, there are no cheap GPS receivers with good hold over performance.
You will be hard pressed to find any hardware with USB 1 that is still running and I have already posted data about the jitter and accuracy that can be expected using USB 2.
The current USB GPS receivers are already very close to the limit of jitter that can be had using only NMEA sentences and going to USB 3 can not improve that in any significant way. The advantage of a USB 3 GPS receiver would be that it could have a PPS signal which WOULD significantly reduce jitter, but there are currently none available. Maybe next year or so.
As I have said several times now, I have a Raspberry Pi with an Ultimate HAT GPS that keeps time in the microseconds. If one wishes to home brew, the simplest thing to do is get such a HAT or an evaluation board, build a simple logic level to RS-232 converter and connect it to a serial port. Using USB for serial data and something else for PPS would require hacking the OS.
Using the $15 VK-162 G-Mouse USB GPS I recommended at the start of this thread provides a long term (several days) timing accuracy of about 5 milliseconds to a 95% confidence level and will settle to well within 100 milliseconds a few minutes after power on. There are lots of others that will produce worse results, but ironically some are more expensive.