Allen has provided a good summary. This GE chart from the mid-1950s provides an illustration:
With just 8 notches, there needs to be judicious allocation of the increments, and typically notch 5 is at or near the half-power point. As may be seen, as well as graduating running power, the same 8 control notches are also used to graduate starting and low-speed accelerating tractive effort. Also, the designated power for each notch is delivered only over the hyperbolic section of each notch curve; above and below this, the power taken from the engine decreases.
Where larger number of notches are used, as found to some extent in European and Japanese practice, it is possible to have some separation between the two functions. For example, the 1950s Sulzer viewpoint was to use the initial notches to graduate out to maximum starting tractive effort with the engine remaining at minimum speed, with the remaining notches then graduating running power with steadily increasing engine speed.
The GE case illustrated appears to be that of a 1600 hp (tractive) Alco 12-244 engine fitted with the GE 17MG electro-hydraulic governor. That allowed pretty much an independent choice of each of the 8 engine speeds. With the Woodward PG governor fitted with electro-hydraulic speed control, it was customary to use 8 equally-spaced engine speeds. This was not mandated by the governor speed control system, although the speeds were necessarily interrelated in some way. That patented Woodward speed control system arrived, I think, with the 1945 new version of the SI governor, and was segued to the PG in 1948. It could provide up to 15 engine speeds, and was also configured to replicate (using the same trainwire protocol) the 8-speed control provided by the 4-unit variant of the GE 7MK3 throttle operator (and its EMD derivative), as used with the UG8 and earlier SI governors fitted with rod shutdown.
As Allen says, this kind of information is not easy to find, and at best one usually has, to use an analogy, an incomplete jigsaw puzzle from which interpolations and extrapolations are needed to construct an estimate of reality. There are different ways of approaching the whole subject; my preference is to use Lemp load control as the foundation for study.