Between the two issues, there may be a strong connection.Curvefit can occur due to the small number of bars used. As I wrote in another post "FSB generates too many "good strategies" that are very difficult to verify without having a method of measuring their robustness (WFE). I mean the old principles of Robert Pardo.My proposal is to use a genetic optimizer with settable parameters.(population size, max generations, fitness functions, etc) , and most important , a WFA (walk forward analyser) module. .Until this proposal will be considered, no choice but to do WFA "manual".
It is a long and painful task.
When you have to make 10 - 20 optimizations, each with 4-8 parameters,on 200000-500000 bars charts, speed really does matter!.
To write an application that can load 8 (16) cores at> 80% capacity, is a task that could not be solved neither by giant companies such as TradeStation.(BTW: only, whom I know, who have solved the problem are the guys from Multichatrs).
In conclusion: In order to find robust strategies is needed WFA. But WFA requires higher computing speed (in parallel for all core).
FSB-Pro does not succeed load my CPU more than 12-15% (probably like yours, as well).
So no need for GPU, but it requires a software written "otherwise".
But, as  Mr. Popov said, parallel programming is very complicated. This means much more money to spend...Free is too cheap!
For better strategies we`ll have to pay ( I presume...)