This study aims at characterizing the asymptotic behavior of genomic prediction R2 as the size of the reference population increases for common or rare QTL alleles through simulations. Haplotypes derived from whole-genome sequence of 85 Caucasian individuals from the 1,000 Genomes Project were used to simulate random mating in a population of 10,000 individuals for at least 100 generations to create the LD structure in humans for a large number of individuals. To reduce computational demands, only SNPs within a 0.1M region of each of the first 5 chromosomes were used in simulations, and therefore, the total genome length simulated was 0.5M. When the genome length is 30M, to get the same genomic prediction R2 as with a 0.5M genome would require a reference population 60 fold larger. Three scenarios were considered varying in minor allele frequency distributions of markers and QTL, for h2 = 0.8 resembling height in humans. Total number of markers was 4,200 and QTL were 70 for each scenario. In this study, we considered the prediction accuracy in terms of an estimability problem, and thereby provided an upper bound for reliability of prediction, and thus, for prediction R2. Genomic prediction methods GBLUP, BayesB and BayesC were compared. Our results imply that for human height variable selection methods BayesB and BayesC applied to a 30M genome have no advantage over GBLUP when the size of reference population was small (<6,000 individuals), but are superior as more individuals are included in the reference population. All methods become asymptotically equivalent in terms of prediction R2, which approaches genomic heritability when the size of the reference population reaches 480,000 individuals.