Resumen
In this paper, we consider the implementation of the three-pass method of fast automatic differentiation. Adding one more pass to the standard 2-pass method leads to a significant decrease in the consumption of RAM, with an insignificant (25%) increase in the number of calculations. Estimates of performance and requirements for the amount of RAM of 2 and 3-pass methods are given. A numerical experiment is performed on the example of the inverse problem of reconstructing the initial data for the transport equation.