Grain yield and end-use quality are the major target traits in durum wheat breeding. In traditional pedigree breeding, phenotypic evaluation for these traits is conducted in late generations and in replicated field plots due to both low heritability and seed quantity requirements. The inclusion of more breeding lines in yield trials can increase the selection effectiveness but is constrained by the high phenotyping cost. In genomic selection (GS), breeding values of complex traits are predicted using low cost genome wide markers, which allows selection of promising lines from a larger number of early generation breeding lines prior to expensive yield trials, and therefore enhances selection efficiency. The aim of this study was to explore the potential application of GS in the North Dakota State University (NDSU) durum wheat breeding program. Genomic prediction ability was evaluated using over 2,000 breeding lines from 2012-2018 preliminary and advanced yield trials of the NDSU durum wheat breeding program. Prediction accuracies of 0.51-0.70 based on cross-validation were obtained for grain yield and quality traits. High forward prediction accuracies were found when using breeding lines from previous breeding cycles as the training population to predict a new breeding cycle. This can be explained by the narrow genetic diversity of the NDSU durum breeding population and the close genetic relationships between breeding cycles, where elite cultivars were commonly used as parents. Our results suggest that the developed GS models could predict new crosses within the NDSU durum breeding germplasm. For some quality traits, the optimal training population was identified as a subset of breeding lines with close genetic relationships to the selection population rather than including all the breeding lines. Given the important role of genetic relationship, designing a cross-specific training population may achieve a better prediction. Genotyping cost is a major limiting factor to implement GS, so we are exploring the feasibility of genotyping only a few hundred markers in the selection progenies and utilizing imputation power to achieve high prediction accuracy.