The heritability is an important determinant of genomic selection success. Striving for a heritability as high as possible when designing a multi-environment preliminary breeding trials quickly hits the limit of available resources. To allow for using these more efficiently for hybrid performance prediction, we investigated several ways of allocating a fixed number of phenotyping plots onto different testing environments in a sparse fashion. We estimated the success of different such ways and genomic prediction models by subsampling a published genomic and phenotypic dataset of 1,604 winter wheat (Triticum aestivum L.) hybrids, estimating hybrid performance with each model and comparing the predictions of the sparse scenarios to those of the full data set as reference values. The results showed that models which had a notion of the relatedness of the genotypes yielded predictions were more capable in predicting nearer to the reference values in unbalanced settings than models without that notion. The adequate characterization of each environment by balanced sampling with genotypes also was associated with predictions that were nearer to the reference values. As a possible application, one advantageous design was testing groups of genotypes in different, but overlapping sets of environments. That layout provided accurate predictions and low variability depending on sampling while covering a larger range of environments without higher resource demand, showing the utility of optimized trial design for hybrid breeding.