hardware doubling number for Crafty
Posted: Wed Sep 15, 2010 5:02 am
Since a couple of posters keep hijacking every thread, I thought I would at least publish the just-completed test on 23.4.
I repeated the same idea, although only for 1/2, that I did for 10.18. If you remember, with 10.18, cutting the search time by 1/2 (leaving opponents at normal time, which simulates cutting cpu speed by 50%) reduced 10.18 by 70 Elo, and halving that reduced it by another 80.
For 23.4, here are the two numbers:
23.4-1 and 23.4-2 are my normal "control" runs, normal everything. 23.4-3 was played using a search time of 1/2 the normal value, slowing the program effectively by 50%. Cost was almost exactly 100 Elo. Very possibly caused by the difference in branching factor between 23.4 and 10.18. 23.4 is pretty close to 2 or even better at times. 10.18 is around 5 for the few positions I tried, could be a little more at times. For the starting position, it is about 5, where 23.4 is right on 2. So for 23.14 we get a whole ply, for 10.18, a fraction of a ply that is < .5...
I did not do the 1/4 speed run (yet). Got side-tracked trying to get gnuchess 5.07 to run on the cluster. Seems (so far) it is not quite protocol version 2 complaint, refusing to send "move xyz" and insisting on "my move is xyz". I don't want to hack my referee so am looking at gnu to see if I somehow overlooked something. It is responding to the protover 2 correctly, but it isn't sending the move correctly...
I repeated the same idea, although only for 1/2, that I did for 10.18. If you remember, with 10.18, cutting the search time by 1/2 (leaving opponents at normal time, which simulates cutting cpu speed by 50%) reduced 10.18 by 70 Elo, and halving that reduced it by another 80.
For 23.4, here are the two numbers:
Code: Select all
Crafty-23.4-1 2666 4 4 30000 61% 2576 20%
Crafty-23.4-2 2664 4 4 30000 61% 2576 21%
Crafty-23.4-3 2567 3 3 30000 50% 2576 21%
I did not do the 1/4 speed run (yet). Got side-tracked trying to get gnuchess 5.07 to run on the cluster. Seems (so far) it is not quite protocol version 2 complaint, refusing to send "move xyz" and insisting on "my move is xyz". I don't want to hack my referee so am looking at gnu to see if I somehow overlooked something. It is responding to the protover 2 correctly, but it isn't sending the move correctly...