Genomic selection is a promising approach for reducing the length of the selection cycle in forest tree breeding. Its efficiency must be evaluated across generations for this purpose, but such studies have been performed for multi-generational breeding programmes in only a few forest tree species to date. We analysed a subset of the Eucalyptus globulus breeding population from the Portuguese company Altri Florestal. In total, 412 genotypes from three successive breeding generations were genotyped with 14,716 SNP markers. A comparison of pedigree-based and marker-based relationship coefficients allowed to correct several documented pedigree errors. Deregressed breeding values were estimated from phenotypic records for growth traits (height and diameter) and survival for 31 field trials distributed in one breeding zone in Portugal and used as pseudo-phenotypes for genomic prediction models. Accuracy was assessed by cross-validation according to two main scenarios: (i) a scenario based on a five random fold number, not taking generation into account, and (ii) scenarios investigating progeny validation using parental generations to predict the progenies. Accuracy was highest after pedigree correction and ranged from 0.46 to 0.60 for the first scenario, from - 0.56 to 0.72 for parent/progeny scenarios and from 0.33 to 0.79 when progenies were added to the calibration population. This genomic selection study provides promising insight for the Altri Florestal Eucalyptus breeding programme.