Where are the funs of Leela?

[Dann Corbit]

If they figure out how to get GPU cards the do alpha beta, or they get GPU cards to do full recursion without wait, then the Alpha beta searcher engines will mop the floor with the NN engines, laughing all the way to the bank.
The compute cycles of the GPU cards are orders of magnitude higher than the CPU approach.
It looks like a miracle.
It's a giant nuclear pile strapped to a 1950 Ford pickup.
Beautiful, brutal, and it goes.

[Nay Lin Tun]

Please check current standing of Leela in TCEC and CCC, Leela is back to no 1.
As I said before, previous tournment results were due to small sample size.

In TCEC,

[Nay Lin Tun]

In chess.com,

[chrisw]

If they figure out how to get GPU cards the do alpha beta, or they get GPU cards to do full recursion without wait, then the Alpha beta searcher engines will mop the floor with the NN engines, laughing all the way to the bank.
The compute cycles of the GPU cards are orders of magnitude higher than the CPU approach.
It looks like a miracle.
It's a giant nuclear pile strapped to a 1950 Ford pickup.
Beautiful, brutal, and it goes.

a chess engine running on GPU was written some time ago, I would guess it was AB, there should be a link somewhere.

why do you suppose AB algorithm would be so good on GPU? AB is an inherently sequential algorithm, sure you can do tricks with parallelisation, or run lots of threads at once, but why would a GPU excel over, say a CPU with multi-thread capability?

For example, a standard GPU operation could be to multiply a 1000 square matrix with another 1000 square matrix, one instruction. Whoosh.
But then let’s say you’ld like to add all the 1000 cells together to give one output. That requires, or used to, last time I looked, an algorithm which add 500 cells together, then 250 cells, then 125 cells, then 64 cells, then 32 cells, then 16, then 8, then 4, then 2, then you got the answer. Not so whoosh.
GPUs are great for some things, but not all.

[corres]

If there is a small difference in Elo between two engines like Stockfish and Leele with well selected hardware, opening book and time control any of the two engines can be made winner.
Even without intention too.
And somebody can make big words about it.

[Uri Blass]

If they figure out how to get GPU cards the do alpha beta, or they get GPU cards to do full recursion without wait, then the Alpha beta searcher engines will mop the floor with the NN engines, laughing all the way to the bank.
The compute cycles of the GPU cards are orders of magnitude higher than the CPU approach.
It looks like a miracle.
It's a giant nuclear pile strapped to a 1950 Ford pickup.
Beautiful, brutal, and it goes.

I do not see contradiction between alpha-beta and NN.
from the following link
http://www.chessdom.com/alliestein-the- ... -tcec-s15/

Adam Treat said the following:

"I do think that AlphaBeta is a superior search method and that there is no reason that an AB+(eval method) engine can’t compete favorably with an MCTS+(eval method) regardless the eval method. But this is just my theory and worth very little until proven. Only the future will tell."

[supersharp77]

If they figure out how to get GPU cards the do alpha beta, or they get GPU cards to do full recursion without wait, then the Alpha beta searcher engines will mop the floor with the NN engines, laughing all the way to the bank.
The compute cycles of the GPU cards are orders of magnitude higher than the CPU approach.
It looks like a miracle.
It's a giant nuclear pile strapped to a 1950 Ford pickup.
Beautiful, brutal, and it goes.

Well said Dann!! I 100% agree....."Lets Compare Apples With Apples and Oranges with Oranges..Shall We?"

[corres]

CCC 9: The Gaunlet Final
Stockfish : Leelenstein(!) = 34 : 32 after 66 games.
So where is Leela?
This match also proves on appropriate hardware Stockfish and their developers can keep step with NN engines.
From these competitions we, the user of chess engines gain the most.
Thanks for every each team!

[Nay Lin Tun]

Leela is overhyped cos she plays beautiful chess , like resurrected Paul Morphy or Tal with steroid!!

First she sacrificed one pawn, then two pawns, and then one piece!
[pgn][Event "CCC 9: The Gauntlet Semifinals (5|2)"] [Site "?"] [Date "2019.07.04"] [Round "1"] [White "Stockfish"] [Black "Lc0"] [Result "0-1"] [ECO "A01"] [GameDuration "00:17:32"] [GameEndTime "2019-07-04T19:32:24.961 PDT"] [GameStartTime "2019-07-04T19:14:52.527 PDT"] [Opening "Nimzovich-Larsen attack"] [PlyCount "220"] [TimeControl "300+2"] [Variation "Modern Variation"] 1. b3 e5 2. Bb2 Nc6 3. e3 d5 4. Bb5 Ne7 5. Bxe5 a6 6. Bxc6+ Nxc6 7. Bb2 Qg5 8. Kf1 Bg4 9. Ne2 O-O-O 10. d4 h5 11. h3 Bf5 12. Nd2 Nb4 13. Nf3 Qf6 14. Ne1 g5 15. Qd2 Qg6 16. c4 Nd3 17. cxd5 Nxb2 18. Qxb2 Bd6 19. Rd1 h4 20. f3 Rhe8 21. Qd2 Kb8 22. e4 Bxe4 23. fxe4 Rxe4 24. Nc3 Re7 25. Qd3 Qf6+ 26. Qf3 Qg7 27. Nc2 Rde8 28. Qf5 Qh6 29. Kg1 Bg3 30. Rf1 c6 31. Qd3 Qd6 32. b4 cxd5 33. a4 f5 34. Qxf5 Rc7 35. Nd1 Re2 36. Qf8+ Qxf8 37. Rxf8+ Ka7 38. Nce3 Rc1 39. Rf1 Rd2 40. a5 Kb8 41. Rf8+ Kc7 42. Rf7+ Kd8 43. Rf1 Ke8 44. Rh2 Rxd4 45. Nf2 Bxh2+ 46. Kxh2 Rxf1 47. Nxf1 Rxb4 48. Ne3 Rb5 49. Kg1 Rxa5 50. Kf1 Kd7 51. Nd3 Kd6 52. Nf5+ Ke6 53. Nd4+ Kf6 54. Nc2 Rb5 55. Ke2 a5 56. Kd2 a4 57. Ndb4 Ke5 58. Nd3+ Kd6 59. Ke3 Rb1 60. Kd4 Rg1 61. Nce1 Rf1 62. Kc3 Ke7 63. Nf3 Kf6 64. Kb4 b5 65. Nfe1 Rg1 66. Ka3 Kf5 67. Nc2 Ke4 68. Nc5+ Ke5 69. Ne1 Kd6 70. Ncd3 d4 71. Kb4 Kd5 72. Ka3 Kc4 73. Nb2+ Kc3 74. Nd1+ Kd2 75. Nf3+ Kxd1 76. Nxg1 d3 77. Kb2 b4 78. Nf3 Ke2 79. Kc1 a3 80. Nd4+ Ke1 81. Nf3+ Kf2 82. Kd2 a2 83. Nxg5 a1=Q 84. Ne4+ Kxg2 85. Kxd3 Qa2 86. Kd4 Qa3 87. Nc5 Kh1 88. Kd5 Kg1 89. Ke4 Kh2 90. Kd4 Kxh3 91. Ne4 b3 92. Nf2+ Kg2 93. Nd1 h3 94. Ne3+ Kg1 95. Nd1 h2 96. Nc3 h1=Q 97. Ne2+ Kh2 98. Nc3 b2 99. Nd5 b1=N 100. Ke5 Qxd5+ 101. Kxd5 Qe7 102. Kd4 Qd6+ 103. Kc4 Kg2 104. Kb5 Na3+ 105. Ka4 Qb6 106. Kxa3 Kf1 107. Ka4 Ke1 108. Ka3 Kd1 109. Ka4 Kc2 110. Ka3 Qb3# 0-1 [/pgn]