I totally agree. But if you are a non-profit you might be able to get a SCIP license for free, you just have to contact the people behind it. But if a moderate performance improvement is not that important for you then GLPK is definetly the best option.
I believe they are all free for academic use. There is also SCIP, one of the fastest academic solvers but you also need a license for commercial use, although I think the license is much cheaper than CPLEX and Gurobi.
I think reducing the TSP to a SAT problem is not a very efficient approach. MIP solvers have become extremely sofisticated (and implement many of the techniques of SAT and much more) and offer a much more powerful modeling paradigm. it's hard to imagine a SAT solver outperforming them on such problems.
I wonder if there is any good resources on hardware details for someone with a backgroud in applied math.
Whenever I find an article talking about cache/branch prediction, I find it hard to follow.