In this study, the effect of predictor dimensionality reduction using Principal Component analysis (PCA) on the accuracy of Direct Genomic Values (DGV) for of 2,822 Italian Holstein bulls was tested. A subset of animals (916) were genotyped with high density (HD, 800K) beadchip; the remaining were imputed from medium density (MD, 50K) to HD. Several 617,166 markers were retained after data editing. A MD panel was simulated for all animals by merging the HD panel with the BovineSNP50v2. A total of 40,669 markers were retained for the analysis. PCA were carried out both genome and chromosome wide for the MD (MD_GW and MD_CHR respectively), only by chromosome for the HD (HD_CHR) panel. Several PC explaining 90% of the total variance was retained: 1,436 (MD_GW) 4,829 (MD_CHR) and 5,321 (HD_CHR) respectively. PC score effects for 32 productive, functional and conformation traits were estimated on 2,301 training bulls born before 2004. DGVs were computed for the remaining 521 validation bulls born after 2003. Accuracy of prediction (rDGV) was computed as correlation between DGVs and phenotypes. Average rDGV across 32 traits were 0.29 ± 0.13, 0.31 ± 0.13, 0.36 ± 0.14 for MD_GW, HD_CHR and MD_CHR, respectively. Such an increase may be explained by the reduction of asymmetry between the number of predictors and observations. In particular, the shrinkage of the total variance of different SNP panel size in a quite similar number of PCs occurring during PCs extraction, could be seen as possible explanation of a better repartition of variance that resulted in a gain of rDGVpassing from genome to chromosome wide and from high to medium density. This research was supported by Italian Ministry of Agriculture, grant INNOVAGEN.