Approximation method is slower than the exact method

I am comparing the computational performance of approximation method vs. exact method over a few benchmark datasets. It is my impression that we should expect less computation using approximation method as a less number of split points are evaluated in each tree splitting. This is the case for most of the benchmark datasets but one. This dataset is for a binary classification task, with around 2.2M training samples and 434 features (both continuous and encoded categorial features). In this dataset, the approximation method (setting tree_method='approx') is about 20% slower than the exact method. I cannot share the data as it is proprietary, but would like to possible ways to investigate what is the reason here.