Maize ear cob diameter (CD) and kernel ratio (KR) are controlled by multiple quantitative loci and both traits associate with the yield production. The genomic selection in conjugation with information identified by genome-wide association study (GWAS) is an effective method to improve the prediction accuracy. By taking use of the phenotypic datasets of CD and KR and the genotypic data derived from genotyping-by-sequencing in 309 maize inbred lines, here we investigated the genomic prediction accuracy using three GS models (genomic best linear unbiased prediction, GBLUP; Bayes A; reproducing kernel Hilbert space, RKHS) and different marker subsets (GWAS-derived markers: fixed and random model circulating probability unification, FarmCPU; compressed mixed linear model, CMLM; randomly selected markers, and all markers). By taking use of FarmCPU- and CMLM-derived markers at both traits, only slight difference (0 to 0.03) on the prediction accuracy using three prediction models was observed. For random markers, GS using RKHS model represented higher prediction accuracy of CD (3.57-15.91%) if compared to two other models, whereas no difference for KR was detected. Except for 50 and 100 markers, the prediction accuracy of CMLM-derived marker using three models were higher than that of FarmCPU-derived markers. Compared to random markers, GWAS-derived markers were able to increase the prediction accuracy (15.52%-88.37% for CD; 1 to 5.89-fold for KR). The prediction accuracy by deployment of GWAS-derived marker subsets was higher than that of all markers. Collectively, these results indicated that genomic selection using GWAS-derived markers could improve the prediction accuracy of CD and KR in maize.