The availability of molecular markers has made possible the use of genomic selection in plant breeding. Several proof-of-concept experiments on genomic selection had been implemented in CIMMYT maize program to identify the best untested inbreds within bi-parental populations and to carry out rapid cycle recurrent selection for population improvement within multi-parental populations. Twenty three bi-parental tropical maize populations were used to test genomic selection prediction accuracy for selecting best untested inbreds as parents, which were comprised around 200 F2 individuals each, genotyped with around 200 SNPs and phenotyped in several drought and optimal environmental conditions. Results indicated that increasing the population size of training set and marker density had a positive impact on prediction accuracy. When 90 F2 individuals and all SNPs were used to predict the rest, prediction accuracy for grain yield under optimal condition was up to 0.58. But it dropped to around 0.40, when 30 individuals were sampled. Predictive ability for all the traits under drought condition was worse than that under optimal condition, which confirmed the importance of improving field evaluation under stress conditions. Rapid cycle genomic selection for population improvement was implemented with 1000 S1 lines formed by 18 parents after two diallel cross generation and one self-pollination generation. RKHS and GBLUP methods showed similar overall prediction accuracy for grain yield and anthesis date prediction, prediction accuracy was up to 0.547 and 0.418 for grain yield and anthesis date, respectively. Results of these proof-of-concept experiments from CIMMYT encourage us to use genomic selection as an effective strategy to select untested lines from bi-parental populations and to do rapid cycle recurrent selection for tropical maize improvement.