NTP is reliant on a reference clock and all clocks on the NTP network are synchronised to that time. It is therefore imperative that the reference clock is as accurate as possible. The most accurate timepieces are atomic clocks. These large physics lab devices can maintain accurate time over millions of years without losing a second.
An NTP server will receive the time from an atomic clock either from across the internet, the GPS network or radio transmissions. In using a atomic clock as a reference an NTP network will be accurate to within a few milliseconds of the world’s global timescale UTC (Coordinated Universal Time).
NTP is a hierarchical system. The closer a device is to the reference clock the higher on the NTP strata it is. An atomic clock reference clock is a stratum 0 device and a NTP server that receives the time from it is a stratum 1 device, clients of the NTP server are stratum 2 devices and so on.
Because of this hierarchical system, devices lower down the strata can also be used as a reference which allows huge networks to operate while connected to just one NTP time server.
NTP is a protocol that is fault tolerant. NTP watches out for errors and can process multiple time sources and the protocol will automatically select the best. Even when a reference clock is temporarily unavailable, NTP can use past measurements to estimate the current time..