Data dependences are a major limitation to the amount of instruction-level parallelism (ILP) that current processors can achieve. Data value speculation is a promising emerging approach that can eliminate the ordering imposed by data dependences. The objective of this work is to establish the performance potential of data value speculation. First, we study the performance of data value speculation for an ideal machine with infinite resources. Our results indicate that most performance benefit for the ideal configuration is derived by the correct prediction of arithmetic instructions. We then evaluate the performance of a more realistic superscalar processor configuration by predicting separately load and arithmetic instructions. The results for this configuration indicate, unlike an ideal machine, that load and arithmetic speculation have similar performance potential. Finally, we study the effect of combining arithmetic and load speculation. The average speedup is about 16%, and in the best case 71%, for Spec95 benchmarks.