Genomic selection (GS) could make more efficient the two-stage phenotypic breeding scheme used for rubber production in Hevea brasiliensis. It was evaluated using two trials in Côte d'Ivoire comprising 189 and 143 clones of the cross PB 260¿×¿RRIM 600, genotyped with 332 simple sequence repeat markers. The effect of statistical genomic prediction methods, training size, and marker data on GS accuracy was investigated when predicting unobserved clone production within and between sites. Simulations using these empirical data assessed the efficiency of replacing current first stage of phenotypic selection (evaluation of seedling phenotype) by genomic preselection, prior to clone trials. Genomic selection accuracy in between-site validations using all clones for training and all markers was 0.53. Marker density and training size strongly affected accuracy, but 300 markers were sufficient and using more than 175 training clones would have marginally improved accuracy. Using the 125–200 markers with the highest heterozygosity, between-site GS accuracy reached 0.56. Prediction methods did not affect GS accuracy. Simulations showed that genomic preselection on 3000 seedlings of the considered cross would have increased selection response for rubber production by 10.3%. Hevea breeding programs can be optimized by the use of within-family GS. Further studies considering other crosses and traits, consecutive breeding cycles, more contrasted environments, and cost-benefit ratio are required.