This question has the same answer as previously given. If you can do that with a parallel search, you should be able to do that with a multiplexed serial search.Dann Corbit wrote:What if it is possible to use the information gathered to reduce the branching factor by 0.15? By 0.25?syzygy wrote:He certainly did not see a superlinear speedup.Dann Corbit wrote:Don Homan does the same thing in his parallel search and he saw a speedup. But people do not do that in the normal alpha-beta serial search.
I have no problem accepting that there might be good approaches to multi-threaded searching that to some extent leave the "try to mimick the serial search as best as we can"-approach. If a particular smp implementation happens to reach max Elo at 55 cores, because with 56 cores the extra synchronisation overhead outweighs the benefit of that extra core, then doing anything else with that core that makes remotely sense can only be an improvement. But you're just not going to get the performance of a single core search at 56x the speed.
I am not saying that this can be done. But I don't know why it can't be done either.
1.4^56 = 152464944
1.2^56 = 27174
5611 times faster from a change in the exponent of 0.2 for a 56 ply search.
Note that I just chose 56 arbitrarily. It has nothing to do with your 56.
This is also why I think that any really large Elo gains will come from search enhancements rather than all the tweaking of the evaluation function.
A simple example might put this to rest. Suppose that for some reason, your move generation and ordering is completely backward, where you search the worst move first and the best move last. A parallel search will produce a super-linear speedup at times, because it will search all moves at a ply at once, and get to that best move sooner than the serial search. But if you multiplex the search, you will still do a simulated split at that node, and search one node for each sub-tree, then another node for each subtree, and you get that same effect. Even though the serial search was completely broken initially because move ordering was backward.