Main paper can be found here.

An optimized GPU version of TSP solver using Iterative Hill Climbing (IHC) technique
Paper limits itself to 100-cities with 100,000 random restarts.
However, the techniques suggested here can be generalized

each thread works on one of the climbers.
- continues to perform the 4851 2-opt moves in each IHC step, until convergence
- FYI, $$4851 = \frac{(100 - 1) * (100 - 2)}{2}$$
after convergence of each climber, a global atomicCAS is performed in order to update the best solution found so far
they only compute the incremental TSP cost instead of computing the full path cost inside IHC step and only compute the global path cost after convergence
all threads continue to fetch their next climber from a global queue until it is exhausted
distance matrix is assumed to be pre-computed (for edge weights)