There is something I don't understand. They say that the winner used Speedment. If I understood the example on github correctly, Speedment also uses SQL interna

There is something I don't understand. They say that the winner used Speedment. If I understood the example on github correctly, Speedment also uses SQL internally. So, why is the same query over JDBC so much slower than with Speedment? Could it be that Speedment avoids the nested select and does the calculations of the statistics by itself? Would that mean that I should use Speedment only when the result set is very small? How would the performance look like if the task were to calculate the average film length of 10000 films from a table of 100,000 films?