Comments (2)
  • Add a Comment
  • Edit
  • More Actions v
  • Quarantine this Entry

1 msoos commented Permalink

The 3666x sounded ridiculous. For the price that a D-Wave machine costs, buying many regular machines will do the jobs faster and for less. Also, this "up to X faster" sounds fishy to me: from my own experience, solving time is probably non-deterministic for both solving systems (though less so for UNSAT, which is needed for proving the optimality). This allows for taking the best from one and the worst from the other. Also, with both systems being non-deterministic, reporting speedup to the last digit and not simply saying e.g. "more than 3 orders of magnitude" gives the impression as if these speedups can be measured to that precision.

2 JeanFrancoisPuget commented Permalink

As said, my focus was on getting better performance for CPLEX. I don't think we can draw conclusions on the usefulness or not of D-Wave machine from my experiments.

For instance, history shows that hardware prices drops dramatically as usage generalizes. I would not draw conclusions from today's levels.
It is not obvious either that buying lots of machines would help CPLEX solve the problem faster. Currently, CPLEX does not run on a distributed memory system.
Last, I don't know whether solving time is non deterministic with D-Wave machine. From McGeoch paper it is not. For CPLEX there is some variability as shown by my experiments but this variability is quite small when using multi threading.