A nonlinear optimization method based on first-order generalized perturbation theory (GPT) and mathematical programming has been extended to three dimensions in the code OPTEX and applied to a realistic problem in the physics design of Canada deuterium uranium (CANDU) reactors. The choice of three-dimensional linear GPT for computing the cost coefficients is justified, and the optimization approach is discussed in reference to methods used for light water reactor fuel manage-ment. The design problem consists of simultaneously adjusting the fueling rate distribution and the grading of the adjuster rods in the core, while satisfying limits on the maximum bundle and channel powers at full power equilibrium refueling. Passage to three dimensions is a requirement for a real-istic modeling of equilibrium refueling in CANDU. It has a significant effect on the system equations, which become nonlinear with the inclusion of the axial dimension. The nature of the constraints is also affected: Separate limits on channel and bundle powers must now be accounted for. These problems are addressed, and a practical optimization scheme is proposed that can handle realistic CANDU core and fuel management design problems.