Author: Robert Hyatt
Date: 13:06:15 08/12/05
Go up one level in this thread
On August 12, 2005 at 14:27:23, Matthew Hull wrote:
>On August 12, 2005 at 14:23:29, Robert Hyatt wrote:
>
>>On August 12, 2005 at 11:52:26, Darrel Briley wrote:
>>
>>>On August 11, 2005 at 20:57:09, Robert Hyatt wrote:
>>>
>>>>On August 11, 2005 at 14:37:56, Uri Blass wrote:
>>>>
>>>>>On August 11, 2005 at 14:00:24, Robert Hyatt wrote:
>>>>>
>>>>>>On August 11, 2005 at 13:57:36, Madhavan wrote:
>>>>>>
>>>>>>>On August 11, 2005 at 13:53:37, Robert Hyatt wrote:
>>>>>>>
>>>>>>>>On August 11, 2005 at 12:55:10, Thorsten Czub wrote:
>>>>>>>>
>>>>>>>>>1. Fruit
>>>>>>>>>2. Shredder
>>>>>>>>>3. Junior
>>>>>>>>>4. Crafty
>>>>>>>>>5. Zappa
>>>>>>>>>6. Diep
>>>>>>>>>7. Sjeng
>>>>>>>>>8. Jonny
>>>>>>>>>9. IsiChess
>>>>>>>>>10. My_fute
>>>>>>>>
>>>>>>>>
>>>>>>>>That will be a miracle. A single opteron CPU (single core) is 2x as fast as a
>>>>>>>>3.0ghz pentium. I, at least, will be running on 8 of 'em. That is a tough
>>>>>>>>disadvantage to overcome...
>>>>>>>
>>>>>>>Don't forget Fruit beta has got another 100 elo improvement.
>>>>>>
>>>>>>
>>>>>>Doesn't matter at all. a speed advantage of 16x is huge. Not insurmountable,
>>>>>>but _huge_ and very difficult to overcome.
>>>>>
>>>>>I guess that it does not gives more than 150-200 elo improvement because of
>>>>>diminishing return that means that being twice faster give only 40-50 elo
>>>>>improvement at long time control and not 70 elo.
>>>>>
>>>>>It seems that public fruit is about 150-200 elo better than public Crafty.
>>>>>I expect both WCCC fruit and WCCC Crafty to be better than the public version so
>>>>>it seems to me that fruit and Crafty have equal chances.
>>>>>
>>>>>Maybe Crafty has better chances because you could learn from Fruit's evaluation
>>>>>so you improved Crafty more than Fabien improved Fruit but only you can tell if
>>>>>looking at the source of fruit helped you to improve Crafty.
>>>>
>>>>That I can answer easily. I've not looked at the source of fruit. So there's
>>>>no way it could have helped. In fact, I haven't even seen a game fruit vs
>>>>crafty so I have no idea how the thing plays.
>>>>
>>>>My comment was solely about speed. a factor of 16 (over 3.0ghz) is 4 doublings,
>>>>each doubling is certainly worth something. Whether it will make Crafty
>>>>stronger than fruit, I won't speculate about since I have no idea how they
>>>>compare on equal hardware with reasonable opening book moves. But given two
>>>>programs that are within a hundred rating points of each other, I'd be
>>>>hard-pressed to not pick the one that is suddenly 12-16X faster...
>>>>
>>>>That was my only point. I don't know that I'll win a single game. But I do
>>>>know that it is going to be hell to beat the thing. I've seen 15 ply searches
>>>>in 5 3 games on ICC in the middlegame. I saw 16-17 ply middlegame searches
>>>>against Junior in the next-to-last CCT which Crafty won on a quad cpu box. This
>>>>one is 2x faster...
>>>>
>>>>>
>>>>>
>>>>>>
>>>>>>Also I don't believe the +100 elo stuff anyway, otherwise all the programs would
>>>>>>be rated 3600 and up by now...
>>>>>
>>>>>I also do not believe the +100 elo stuff for fruit in this version but the stuff
>>>>>is only about fruit so I do not see how you get your conclusion about other
>>>>>programs.
>>>>
>>>>Simple. Every year, every new version, "the new version is 60-80-100 elo
>>>>stronger than last year's version." Seen that over and over and over. :)
>>>>
>>>>
>>>>
>>>>
>>>>>
>>>>>The +100 elo stuff was correct for Fruit in the past and it is a fact that fruit
>>>>>improved faster than other programs
>>>>>
>>>>>Fruit2.0 was already above Crafty level on one cpu and Fruit2.1 is slightly more
>>>>>than 100 elo better than fruit2.0 based on the CEGT rating list.
>>>>>
>>>>>Uri
>>>
>>>Bob may not have noticed (or payed it much attention since it was a blitz game),
>>>but I used Fruit 2.1 against 8x Crafty in one game. Score follows.
>>>
>>>
>>>[Event "ICC 5 3"]
>>>[Site "Internet Chess Club"]
>>>[Date "2005.08.10"]
>>>[Round "-"]
>>>[White "Frutit 2.1"]
>>>[Black "crafty"]
>>>[Result "0-1"]
>>>[ICCResult "Crafty"]
>>>[WhiteElo "3020"]
>>>[BlackElo "2813"]
>>>[Opening "Sicilian: Richter-Rauzer, Rauzer attack, 7...a6"]
>>>[ECO "B66"]
>>>[NIC "SI.29"]
>>>[Time "23:22:40"]
>>>[TimeControl "300+3"]
>>>
>>>1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 d6 6. Bg5 e6 7. Qd2 a6 8.
>>>O-O-O h6 9. Be3 Be7 10. f4 Nxd4 11. Bxd4 b5 12. Qe3 Bb7 13. Bxf6 gxf6 14.
>>>Bd3 b4 15. Ne2 Qa5 16. Kb1 Rc8 17. Rhe1 Qc5 18. Qg3 Bf8 19. Nd4 Qa5 20. f5
>>>e5 21. Nb3 Qb6 22. Nd2 Qc5 23. h3 Qc7 24. Bc4 a5 25. Bb5+ Bc6 26. Bxc6+ Qxc6
>>>27. Rc1 a4 28. c3 b3 29. axb3 axb3 30. Nxb3 Rb8 31. Nd2 Qb5 32. Rc2 h5 33.
>>>c4 Qa5 34. Qc3 Qa6 35. Rd1 Ra8 36. b3 Bh6 37. Nf3 O-O 38. Qb2 Rfb8 39. Rc3
>>>Rb6 40. g4 Qa7 41. g5 fxg5 42. h4 g4 43. Ng5 Ra6 44. Rcd3 Ra1+ 45. Kc2 Ra2
>>>46. Rxd6 Bf8 47. R6d5 Bb4 48. Rd8+ Kg7 49. f6+ Kg6 50. Rg8+ Rxg8 51. Nh3
>>>Rxb2+ 52. Kxb2 Ra8 53. Nf4+ exf4 54. Rd5 Qa1+ 55. Kc2 Ra2+ 56. Kd3 Rd2#
>>>{White checkmated}
>>>0-1
>>>
>>>Fruit showed a pretty good advantage in the eval for most of the game, that is
>>>until those Q-side pawns came crashing down.
>>>
>>>It's only a blitz game, but as a result I've revised my thinking on Fruit's
>>>chances. Lack of multi-processor support and using the hardware provided
>>>locally (PIV 3Ghz) will keep Fruit from being a serious contender.
>>>
>>> DB
>>
>>
>>In that game crafty should have been whispering/kibitzing probably. I don't
>>remember the game myself since I have played several over the past few nights
>>and am generally pretty well burned-out before I start the games, so I tend to
>>not remember who the opponents were (other than Blargh, shredder on a dual
>>opteron that I have played several longer games against recently to test
>>things). Did Crafty think it was worse during the game? Or did both think they
>>were better, which is not that uncommon...
>>
>>And remember, that was a fast game. Longer time controls tend to narrow the
>>search gap significantly. But speed can be a killer no matter what... of
>>course...
>
>
>Were you able to get that last .5 factor of NPS scaling?
yes and no. Yes with aggressive internal tuning. But I have turned it down a
bit for more consistent SMP performance. For example, would you prefer a 7.5X
SMP speedup on one position, 4.5 on the next, or a more reasonable 5-6 on many
positions from middlegame to endgame?
I've gone for the latter to attempt to reduce those
not-so-common-but-all-too-frequent cases where the speedup is bad. This version
seems to play consistently in blitz, in long games, in endgames, in middlegames,
etc...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.