> > Of course smart playouts are still very important. But when I read > things like fuego running on experimental hardware with 100+ threads > and shared memory I know my chances. I would say currently the > smartest programs running on ordinary machines could probably take one > or two doublings, but not much more. My impression is that in the past > the cluster programs still suffered from significant mpi code bugs, > but the good teams should have solved that by now, and of course none > of that matters for monsters with shared memory... > > I must say that the situation around that is not clear to me. MoGo with MPI is now stable (thanks to, I think, progress in OpenMpi, as the code is the same...), and wins with huge success rate against the non-MPI version. But against humans, or even against other computers, I am not sure that this makes such a big difference.
I'm not sure - not enough numbers for checking this. But playing against oneself is really misleading... For this year, we'll use clusters, but we'll prefer easily available clusters on which we can test easily before the competition, rather than extremely strong clusters - I'm not sure looking for huge hardware is worth the candle. Olivier
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
