Abstract Fatigue life of gearwheels made from commonly used gear materials has been studied extensively. The results can be expressed in the form of S- N curves, and more recently also in R- S- N curves. For a typical gear application, the load spectrum representing gear service loads can be obtained from industrial standard data or through direct measurement. Based on the above load spectrum and R- S- N curves, the reliability design of gears can be carried out by applying one of the cumulative damage hypotheses. This paper investigates the applicability of the Palmgren-Miner hypothesis and two other modified linear cumulative damage hypotheses in gear reliability design through experimental tests. The results show some improvement of the calculated lifetimes with the modified hypotheses. In the higher reliability range, however, a relatively significant discrepancy between the experimental and calculated lifetimes still exists. Hypotheses with progressive damage accumulation are required to give more precise lifetime predictions in this important range.