National laboratories usually operate a number of clocks. These are run independently of one another and their data are combined to generate a perennial time scale. This scale is more stable and more accurate than that of any individual contributing clock. The scale is based on the results of local clock comparisons in the laboratory, and often has an uncertaint of less than 100 ps. These time scales are generally designated TA(k) for laboratory k.

The synchronization of clocks operating in widely separated laboratories is an important concern for time metrology. It calls for accurate methods of clock comparison that can be operated everywhere on Earth, at any time. The satellite system of the Global Positioning System (GPS) provides a satisfactory solution to this problem: made up of twenty-four non-geostationary satellites, this system is designed for positioning, but has the particular feature that the satellites are equipped with caesium clocks which broadcast time signals. These signals are used in the following way. Clocks in two distant laboratories are compared individually with a clock on board a satellite which is visible simultaneously from both laboratories and the difference is calculated. For a comparison extending over ten minutes, the uncertainty thus obtained may be a few nanoseconds, even for clocks which are separated by several thousand kilometres. To reduce these uncertainties to this limit the data must be considered very carefully: results obtained from views that are not strictly simultaneous must be systematically rejected and a correction must be applied to take account of the exact position of the satllite, data known only a few days later.

The GPS is used on a regular basis to link national laboratories in many countries and it will shortly be complemented by a similar Russian system: the Global Navigation Satellite System (GLONASS). Among other methods under study are bidirectional techniques based on the transmission of an optical or radiofrequency signal from one laboratory to another and back, via a satellite. Such methods should lead to subnanosecond accuracy before the end of the century. All these methods of time comparison are subject to relativistic effects which may exceed 100 ns, so corrections must be applied to take them into account.

Optimal combination of all the results of comparisons between the clocks maintained in the national time-service laboratories results in a world reference time scale, International Atomic Time (TAI), approved by the 14th CGPM in 1971 (Resolution 1). The first definition of TAI was that submitted by the then CCDS in 1970 to the CIPM (Recommendation S 2; PV, 38, 110 and Metrologia, 1971, 7, 43):

In the framework of general relativity TAI must be regarded as a time coordinate (or coordinate time): its definition was therefore completed as follows (declaration of the CCDS, BIPM Com. Cons. Déf. Seconde, 1980, 9, S15 and Metrologia, 1981, 17, 70):

This definition was amplified by the International Astronomical Union in 1991, Resolution A4:

Responsibility for TAI was accepted by the CIPM from the Bureau International de l'Heure on 1 January 1988. TAI is processed in two steps. A weighted average based on some 200 clocks maintained under metrological conditions in about fifty laboratories is first calculated. The algorithm used is optimized for long-term stability, which requires observation of the behaviour of clocks over a long duration. In consequence, TAI is a deferred-time time scale, available with a delay of a few weeks. In 1997, the relative frequency stability of TAI was estimated to be 2 parts in 10¹⁵ for mean durations of two months. The frequency accuracy of TAI is evaluated by comparing the TAI scale unit with various realizations of the SI second of primary frequency standards. This requires the application of a correction to compensate for the relativistic frequency shift between the location of the primary standard and a point fixed on the rotating geoid. The magnitude of this correction is, between points fixed on the surface of the Earth, of the order of 1 part in 10¹⁶ per metre of altitude. In 1997, the difference between the TAI scale unit and the SI second on the rotating geoid was +2 x 10^–14 s, and was known with an uncertainty of 5 x 10^–15 s. This difference is reduced by steering the frequency of TAI by the application of corrections, of magnitude 1 part in 10¹⁵, every two months. This method improves the accuracy of TAI while not degrading its middle-term stability.