The availability of dense molecular markers in many crops has made possible the use of genomic selection (GS) for plant breeding. Genomic selection involves estimating marker effects from a training population with known phenotypes and genotypes and predicting genomic estimated breeding values (GEBVs) of untested lines in a breeding population. Selection of superior lines will then be based on the predicted GEBVs instead of actual phenotypes. We evaluated several models and genomic prediction scenarios using a wheat diversity panel and a RIL population that were phenotyped in multiple environments for six agronomic traits and genotyped with the wheat 90K SNP array. Within-population, across-population and across-environment genomic predictions were made to simulate various GS strategies for wheat breeding. Highly variable accuracies were obtained depending on the trait, statistical model and prediction scenario. BayesB resulted in higher accuracies when there were large effect QTL, but there was no difference among the models for traits controlled by many small effect QTL. Moderate to high accuracies (ranged from 0.44 to 0.77) were obtained when predictions were made within populations. In contrast, across-population prediction accuracies were close to zero, even when genetically related lines were included in the training population. Across-environment prediction using the parental generation as training population to predict phenotypes of progenies resulted in moderate to high accuracies (ranged from 0.56 to 0.90). Taken together, our results support deployment of GS strategies in early generations using unreplicated, nest experimental designs in multiple locations representing target agroecological environments.