STM is not a great fit for this kind of problem, there's no need for all the transaction machinery if the problem is embarrassingly parallel. In an ideal world what you want is threads which just split the work sections like they do in C/C++/Haskell.