A single-nucleotide polymorphisms-based genomic relationship matrix (G_SNP) discriminate less identity by state from identity by descent (IBD) alleles compared with a multi-locus haplotype-based relationship matrix (G_HAP), which can better capture IBD alleles and recent relationships. We aimed to compare the prediction reliability and prediction bias of genomic best linear unbiased prediction (GBLUP) using either G_SNP or G_HAP in Holstein cattle. Therefore, a total of 57 traits with a wide range of heritability values were analyzed. Classical validation tests were done using a validation dataset comprised of 50k genotype records of 561–669 proven bulls born in 2010–2011 with an official estimated breeding value (EBV) in 2016 and a training set of 5314–19 678 bulls born before 2010, depending on the trait. The method for building the genomic relationship matrix (G) had significant, but small effect on observed reliability (r²_GEBV) (p < 0.0001) and bias (p < 0.0001). A significant interaction between G and the level of trait heritability on r²_GEBV and bias was also observed (p < 0.0001). The small gains in r²_GEBV and small reductions in the bias by using G_HAPBLUP were increased when predicting moderate to high-heritability traits compared with low-heritability traits.