As we all know, a coder with A rating has a 1/(1+10^((B-A)/400)) probability of accepting a B difficulty problem.

At present, the following situations may occur: Two equally difficulty problems have a difference of 14.3% in acceptance rate; Two problems with a difficulty difference of 100 have the same acceptance rate. Accurating difficulty to one places can reduce errors to 0.1%.

Codeforces Beta Round #23 has only 765 participants, Codeforces Global Round 9 has 21150 participants, so we can definitely make difficulties more accurate by using participants that have increased by more than 25 times.

By the way, would not it be better to have the same accuracy of rating between the problems and coders?