Live Brainfish 200420 vs Lc0 25.1 (63483) TC=15m +15s

[mwyoung]

Live Brainfish 200420 vs Lc0 25.1 (63483) TC=15m +15s

Hardware 2950x, RTX 2080 ti

TC=15m +15s

Brainfish 200420
32 Threads
6 man Tablebase
32 GB of hash with large page on.

Lc0 25.1 (63483)
Book Fritz 17 to 6 moves.
Default settings

Live Stream:

[mwyoung]

Lc0 takes the lead.

DESKTOP-CORSAIR, Rapid 15.0min+15.0sec 0

1 Lc0 v0.25.1+git.69105b4 _ +89 ½½1½ 2.5/4
2 Brainfish 200420 64 POPCNT -89 ½½0½ 1.5/4



[pgn][Event "DESKTOP-CORSAIR, Rapid 15.0min+15.0sec"]
[Site "DESKTOP-CORSAIR"]
[Date "2020.05.06"]
[Round "3"]
[White "Brainfish 200420 64 POPCNT"]
[Black "Lc0 v0.25.1+git.69105b4"]
[Result "0-1"]
[ECO "C54"]
[Annotator "0.38;0.09"]
[PlyCount "280"]
[TimeControl "900+15"]

{AMD Ryzen Threadripper 2950X 16-Core Processor 4200 MHz W=57.1 plies; 51,
887kN/s; 474,632,543 TBAs B=10.2 plies; 36kN/s; 259,728 TBAs; Fritz 17.ctg} 1.
e4 {0.01/0 2 White last book move} e5 {B 0} 2. Nf3 {0.01/0 0} Nc6 {B 0} 3. Bc4
{0.01/0 0} Bc5 {B 0} 4. c3 {0.01/0 0} Nf6 {B 0} 5. d3 {0.01/0 0} a6 {B 0} 6.
O-O {0.01/0 0} d6 {B 0 Black last book move} 7. a4 {0.01/0 0} h6 {0.09/10 32}
8. Re1 {0.01/0 0} Ba7 {0.09/11 39} 9. Nbd2 {0.01/0 0 (b4)} Be6 {0.09/11 35} 10.
Bxe6 {0.01/0 0 (b4)} fxe6 {0.08/11 15} 11. Nf1 {0.01/0 0 (b4)} a5 {0.07/11 51}
12. Qb3 {0.38/34 59} Qd7 {0.07/13 13} 13. Rb1 {0.51/34 30 (Be3)} O-O {0.04/12
60} 14. Ng3 {0.51/35 14} Ne7 {0.04/12 45 (Ng4)} 15. h3 {0.38/35 92} c5 {
0.02/11 32 (Ng6)} 16. Be3 {0.55/31 26} Nc6 {0.02/10 12} 17. Qb5 {0.59/36 31
(Rbc1)} Rac8 {0.01/10 70 (Ne8)} 18. Rbd1 {0.51/37 68} Kh7 {0.00/9 15 (Bb8)} 19.
Nh2 {0.67/32 33 (Nf1)} Rf7 {0.00/10 35 (Qf7)} 20. Nhf1 {0.59/55 100 (Nf3)} Ne8
{-0.07/9 50 (Bb8)} 21. Nh2 {0.33/38 104} d5 {-0.08/13 15} 22. Rf1 {0.27/33 13
(Qb3)} Nd6 {-0.09/12 69} 23. Qb3 {0.40/36 17} d4 {-0.09/12 3 (Ne8)} 24. Bd2 {
0.35/38 19 (Bc1)} Rcf8 {-0.09/10 41} 25. Ng4 {0.25/36 38 (c4)} Bb8 {-0.10/11 29
} 26. Rde1 {0.26/37 21 (Nh5)} Nc8 {-0.13/12 37 (Ne8)} 27. Nh5 {0.69/41 64 (Rd1)
} N8e7 {-0.14/16 31 (g6)} 28. Nxg7 {0.26/42 97} Rxg7 {-0.14/16 25} 29. Bxh6 {
0.22/44 29} Rxg4 {-0.15/15 20} 30. Bxf8 {0.16/42 58} Rg8 {-0.15/13 25} 31. Bxe7
{0.16/48 19} Qxe7 {-0.15/12 16} 32. g3 {0.16/44 17 (Kh2)} Bd6 {-0.15/11 126
(Qf7)} 33. Kg2 {0.26/49 20 (Kh2)} Qf7 {-0.18/11 21} 34. h4 {0.16/47 62 (Rg1)}
Bf8 {-0.18/10 32} 35. Rg1 {0.00/47 111} Bh6 {-0.17/11 28} 36. Qd1 {0.00/42 23
(Ref1)} Ne7 {-0.18/10 69 (Nb8)} 37. Qf3 {0.34/39 16} Qe8 {-0.18/11 26 (Qg6)}
38. Ra1 {0.31/38 38 (Qd1)} Rf8 {-0.20/10 34 (b6)} 39. Qe2 {0.39/34 10} Qf7 {
-0.19/10 31 (Qd8)} 40. Raf1 {0.16/44 64} Kh8 {-0.19/10 49 (Rd8)} 41. Qd1 {
0.15/43 12} Kh7 {-0.17/10 38} 42. Kh3 {0.15/45 16 (Qe2)} Rd8 {-0.21/10 25} 43.
Qe2 {0.15/49 27} b6 {-0.19/9 50 (Rf8)} 44. Kh2 {0.15/41 35 (Rg2)} Ng8 {-0.17/9
26 (Nc6)} 45. Kh1 {0.15/40 14 (Kh3)} Nf6 {-0.30/15 17 (Ne7)} 46. g4 {1.19/36 13
} Bf4 {-0.26/15 18} 47. g5 {1.04/40 24} Qh5 {-0.25/14 16} 48. Qxh5+ {0.85/36 12
} Nxh5 {-0.23/13 12} 49. Rd1 {0.65/40 23} Kg6 {-0.22/12 48 (c4)} 50. Kg2 {
0.00/57 42} Ng7 {-0.21/12 30 (Rb8)} 51. Kf3 {0.52/36 12} Kh5 {-0.21/12 6 (c4)}
52. Ke2 {0.00/46 36 (Rb1)} Rd7 {-0.18/11 60} 53. Ra1 {0.00/48 15 (Rb1)} Bh2 {
-0.15/13 42} 54. Rh1 {0.00/45 11} Bf4 {-0.11/20 2} 55. Rhg1 {0.00/52 12 (Ra3)}
Ne8 {-0.09/13 39 (Bh2)} 56. Ra3 {0.00/50 12} Bh2 {-0.06/24 15 (dxc3)} 57. Rh1 {
0.00/45 16} Bf4 {-0.04/26 6} 58. Rg1 {0.00/52 14 (Rb3)} Bh2 {-0.01/9 32 (dxc3)}
59. Rd1 {0.00/50 21 (Rg2)} Nd6 {-0.34/14 18 (Bf4)} 60. Rb3 {0.00/44 16} Nc8 {
-0.29/15 13} 61. Rh1 {0.00/50 15 (Rb5)} Bf4 {-0.22/17 11} 62. Rg1 {0.00/53 13}
Rg7 {-0.17/18 23 (Kg6)} 63. Rb5 {0.00/50 15} Rb7 {-0.17/22 0} 64. Rb3 {0.00/53
10 (Kd1)} Ne7 {-0.20/16 40 (Rg7)} 65. Rh1 {0.00/55 17 (Rb5)} Nc6 {-0.23/15 42
(Nc8)} 66. Kd1 {0.00/48 18 (Rg1)} Nb4 {-0.80/14 8 (Ne7)} 67. cxb4 {0.00/60 10}
cxb4 {-0.82/15 9} 68. Ke2 {-1.18/58 93 (Kc2)} Rh7 {-1.30/11 17 (Rb8)} 69. Rg1 {
-2.61/49 29} Rh8 {-1.44/13 12 (Rc7)} 70. Kf3 {-2.27/44 26} Rf8 {-1.48/14 13}
71. Ke2 {-3.18/40 16} Kxh4 {-1.51/14 12 (Rb8)} 72. Rh1+ {-3.46/41 17} Kxg5 {
-1.47/13 13} 73. Rh7 {-3.52/40 14} Rc8 {-1.48/11 10 (Rb8)} 74. Rg7+ {-3.48/39
34 (Kf3)} Kh6 {-1.51/10 38 (Kh5)} 75. Re7 {-0.96/48 19 (Rb7)} Rc6 {-1.57/10 12
(Kg5)} 76. Kf3 {-3.46/39 15} Bg5 {-1.58/9 19} 77. Rd7 {-3.64/41 9 (Rb7)} Bh4 {
-1.57/10 20 (Bf6)} 78. Kg2 {-4.22/41 20 (Rb7)} Rc2 {-1.81/10 15 (Bf6)} 79. Rf7
{-4.23/39 5} Kg5 {-1.76/11 25 (Kg6)} 80. Rf8 {-4.56/45 26} Kg6 {-1.79/11 5} 81.
Rg8+ {-4.89/34 15 (Kf1)} Kf7 {-2.06/11 12} 82. Rb8 {-4.91/38 15} Rc6 {-2.12/11
8} 83. Kf3 {-5.30/37 12 (Rb7+)} Ke7 {-2.19/10 17 (Be7)} 84. Kg4 {-4.58/43 18}
Bf6 {-2.36/11 22 (Bxf2)} 85. f4 {-2.26/35 6 (Rb7+)} exf4 {-2.49/11 10} 86. Kxf4
{-3.80/41 21} Kd6 {-2.56/11 9} 87. e5+ {-4.08/40 19 (Kf3)} Bxe5+ {-2.69/12 16}
88. Ke4 {-4.18/37 15} Bf6 {-3.07/12 18 (Kc7)} 89. Rb7 {-7.58/32 15 (Rf8)} e5 {
-3.96/11 14} 90. Rb8 {-8.78/34 15} Be7 {-4.30/11 12} 91. Rg8 {-9.19/36 15 (Rb7)
} Rc1 {-5.30/10 21 (Rc2)} 92. Rb8 {-9.38/38 7 (Rg6+)} Re1+ {-6.43/9 14} 93. Kf3
{-9.97/36 20 (Kf5)} Kc6 {-6.81/8 26 (Kc5)} 94. Rc8+ {-10.84/32 18 (Kf2)} Kd5 {
-6.37/8 21} 95. Rc7 {-11.68/32 15 (Rb8)} Bd6 {-7.03/7 16} 96. Rb7 {-12.38/32
15 (Rc8)} Kc6 {-8.32/7 13} 97. Rh7 {-12.96/31 15} b5 {-7.79/7 17 (Ra1)} 98.
axb5+ {-14.85/27 16 (Kf2)} Kxb5 {-9.30/7 8} 99. Rb7+ {-16.86/28 15 (Rd7)} Kc6 {
-9.32/7 18} 100. Ra7 {-18.30/28 15} Kb6 {-9.27/7 19} 101. Rh7 {-19.52/29 15
(Ra8)} Bc5 {-9.54/7 16 (Bc7)} 102. Rh8 {-18.28/28 15} Kb5 {-8.55/7 15 (Re3+)}
103. Rb8+ {-11.11/27 6} Bb6 {-7.98/7 13} 104. Kg4 {-23.95/30 24} Ka4 {-10.35/7
14 (Re3)} 105. Rxb4+ {-148.96/39 15 (Kf3)} axb4 {-12.25/7 8} 106. Kf3 {-149.00/
46 10 (Rxb6)} Re3+ {-13.11/6 25 (e4+)} 107. Kg4 {-149.01/50 21 (Kf2)} Rxd3 {
-16.65/6 14 (e4)} 108. Rxb6 {-149.00/29 9 (Rg8)} Kb3 {-27.37/6 15 (Re3)} 109.
Kf5 {-149.02/44 21} Kxb2 {-59.15/5 14 (Re3)} 110. Ke6 {-149.06/24 8 (Ke4)} Rh3
{-72.80/5 17 (Kc2)} 111. Kd7 {-149.06/23 22 (Kd5)} Rh8 {-73.80/5 15 (Rc3)} 112.
Rg6 {-149.04/27 15 (Kc7)} d3 {-71.76/5 15} 113. Rg2+ {-#16/73 15 (Kc7)} Kc3 {
-84.27/5 15} 114. Rg3 {-#14/64 6 (Kc6)} Rg8 {-107.07/5 15 (Kc2)} 115. Re3 {
-#40/33 23 (Rh3)} Kd2 {-102.71/5 15 (Rg6)} 116. Rxe5 {-#40/33 7 (Rh3)} b3 {
-#101/1 7 (Kc2)} 117. Rb5 {-#15/72 11} b2 {-#101/1 10 (Kc2)} 118. Rxb2+ {
-#21/66 6} Kc3 {-#102/2 27} 119. Rb6 {-#20/78 13 (Rb1)} d2 {-#101/2 7 (Rg4)}
120. Rc6+ {-#22/85 7} Kd3 {-111.54/3 23} 121. Rd6+ {-#21/81 8} Ke2 {-108.98/3
15 (Kc2)} 122. Re6+ {-#20/84 8} Kd1 {-81.43/5 7} 123. Rh6 {-#19/86 8 (Rb6)} Rg3
{-86.66/5 11} 124. Rh1+ {-#18/91 9} Kc2 {-#103/2 12} 125. Ke6 {-#16/99 9 (Rh2)}
d1=Q {-#101/2 29} 126. Rxd1 {-#15/105 10} Kxd1 {-#101/1 7} 127. Kd5 {-#14/95 9}
Ke2 {-90.24/6 24 (Kd2)} 128. Ke5 {-#13/97 10 (Ke4)} Kd3 {-69.85/6 7} 129. Kf6 {
-#12/85 10 (Kf5)} Re3 {-69.22/7 21} 130. Kf5 {-#11/119 9} Re4 {-55.85/7 14
(Kd4)} 131. Kf6 {-#10/119 10 (Kg5)} Ke3 {-46.99/6 8} 132. Kg5 {-#9/92 10 (Kf5)}
Rf4 {-39.40/7 11} 133. Kh5 {-#8/200 10} Rf5+ {-30.82/7 21 (Kf3)} 134. Kh6 {
-#7/245 1 (Kg4)} Kf4 {-24.04/7 10} 135. Kg6 {-#6/245 0} Kg4 {-23.62/7 12} 136.
Kh7 {-#5/245 1 (Kh6)} Kh5 {-#5/3 17 (Kg5)} 137. Kg7 {-#4/245 0} Kg5 {-#4/2 12}
138. Kh8 {-#3/245 0 (Kh7)} Rf7 {-#3/2 23 (Kh6)} 139. Kg8 {-#2/1 0} Kg6 {
-#2/2 10} 140. Kh8 {-#1/1 0} Rf8# {-#1/2 12} 0-1

[/pgn]

[Milos]

Live Brainfish 200420 vs Lc0 25.1 (63483) TC=15m +15s

Hardware 2950x, RTX 2080 ti

TC=15m +15s

Brainfish 200420
32 Threads
6 man Tablebase
32 GB of hash with large page on.

Lc0 25.1 (63483)
Book Fritz 17 to 6 moves.
Default settings

Live Stream:

Lc0 is using exactly 2x more expensive hardware.
For equal cost one could provide SF with 50% more cores and 30% more nps per core, i.e. almost 2x more nps.
That plus short TC means SF has 0 chance.

[mwyoung]

Live Brainfish 200420 vs Lc0 25.1 (63483) TC=15m +15s

Hardware 2950x, RTX 2080 ti

TC=15m +15s

Brainfish 200420
32 Threads
6 man Tablebase
32 GB of hash with large page on.

Lc0 25.1 (63483)
Book Fritz 17 to 6 moves.
Default settings

Live Stream:

Lc0 is using exactly 2x more expensive hardware.
For equal cost one could provide SF with 50% more cores and 30% more nps per core, i.e. almost 2x more nps.
That plus short TC means SF has 0 chance.

This is all news to me. I did not set up my testing rig on hardware price as prices change. But the Lc0 ratio on equal hardware. The Lc0 Ratio is 1.1 with the standard Stockfish with no large page support.. And prices change on hardware, but not the Lc0 ratio. So what is your point.

I have been testing Lc0 for a long time. And this is the first time. I have ever seen someone say that Lc0 gets an advantage over Stockfish at shorter time controls. I have tested Lc0 over much longer time controls. And this is where Lc0 has its best strength over Stockfish. If Stockfish is going to be better then Lc0. The testing shows it will be at shorter time controls. When Lc0 has a chance to make a tactical blunder to Stockfish.

And Stockfish has in this match an unlimited full book of Brainfish plus large page making Stockfish about 10% faster then the standard Stockfish, but Lc0 is limited to only 6 moves in its Fritz 17 book.

[Milos]

This is all news to me. I did not set up my testing rig on hardware price as prices change. But the Lc0 ratio on equal hardware. The Lc0 Ratio is 1.1. And prices change on hardware, but not the Lc0 ratio. So what is your point.

Leela ratio is just some arbitrary number. Imaginary number that comes from some totally unrelated testing on unrelated hardware.
It has nothing to do with any measurable parameter coming from real world, such as power or price of hardware.

I have been testing Lc0 for a long time. And this is the first time. I have ever seen someone say that Lc0 gets an advantage over Stockfish at shorter time controls. I have tested Lc0 over much longer time controls. And this is where Lc0 has its best strength over Stockfish. If Stockfish is going to be better then Lc0. The testing shows it will be at shorter time controls. When Lc0 has a chance to make a tactical blunder to Stockfish.

That is simply false. MCTS scales worse than A/B and this is know since forever. Once you have strong enough hardware or long enough TC A/B is gonna start scaling better. A/B also scales better on short TCs and weak hardware. Where MCTS performs better is rapid.

And Stockfish has in this match an unlimited full book of Brainfish plus large page making Stockfish about 10% faster then the standard Stockfish, but Lc0 is limited to only 6 move in its Fritz 17 book.

Brainfish book at this hardware only gives small time advantage from move 6 until SF is out of book, assuming you limit BF book depth to something like 15-20. If you don't, then using the book actually weakens SF.

[mwyoung]

This is all news to me. I did not set up my testing rig on hardware price as prices change. But the Lc0 ratio on equal hardware. The Lc0 Ratio is 1.1. And prices change on hardware, but not the Lc0 ratio. So what is your point.

Leela ratio is just some arbitrary number. Imaginary number that comes from some totally unrelated testing on unrelated hardware.
It has nothing to do with any measurable parameter coming from real world, such as power or price of hardware.

I have been testing Lc0 for a long time. And this is the first time. I have ever seen someone say that Lc0 gets an advantage over Stockfish at shorter time controls. I have tested Lc0 over much longer time controls. And this is where Lc0 has its best strength over Stockfish. If Stockfish is going to be better then Lc0. The testing shows it will be at shorter time controls. When Lc0 has a chance to make a tactical blunder to Stockfish.

That is simply false. MCTS scales worse than A/B and this is know since forever. Once you have strong enough hardware or long enough TC A/B is gonna start scaling better. A/B also scales better on short TCs and weak hardware. Where MCTS performs better is rapid.

And Stockfish has in this match an unlimited full book of Brainfish plus large page making Stockfish about 10% faster then the standard Stockfish, but Lc0 is limited to only 6 move in its Fritz 17 book.

Brainfish book at this hardware only gives small time advantage from move 6 until SF is out of book, assuming you limit BF book depth to something like 15-20. If you don't, then using the book actually weakens SF.

Yes, it is all bull shit and just imaginary numbers. A standard means nothing.

Where where you saying this over the many test of Stockfish and Lc0. I have posted before at long and short time controls.
What conditions do you want to test that will show a Stockfish win on my hardware given what you have said. Post the conditions for the test given your comments! How much time does Stockfish need to beat Lc0 on my hardware. And it will be tested. Unless Stockfish needs many hours a game to beat Lc0.

The reason for this test is that many have said Brainfish, the best Stockfish version with the best rating will beat Lc0.

Some have said the only reason that Stockfish was losing in other test to Lc0 like TCEC was a shorter book.

If you think that Stockfish can beat Lc0. Give the conditions of your test. And I will test them.

[Milos]

Where where you saying this over the many test of Stockfish and Lc0. I have posted before at long and short time controls.
What conditions do you want to test that will show a Stockfish win on my hardware given what you have said. Post the conditions for the test given your comments! How much time does Stockfish need to beat Lc0 on my hardware. And it will be tested. Unless Stockfish needs many hours a game to beat Lc0.

The reason for this test is that many have said Brainfish, the best Stockfish version with the best rating will beat Lc0.

Some have said the only reason that Stockfish was losing in other test to Lc0 like TCEC was a shorter book.

If you think that Stockfish can beat Lc0. Give the conditions of your test. And I will test them.

TC for SF 90min+30sec. TC for Lc0 45min+15sec.
HT on. Contempt 0.
Limit BF book to depth 12.
Everything else exactly the same as you have now.

[mwyoung]

Where where you saying this over the many test of Stockfish and Lc0. I have posted before at long and short time controls.
What conditions do you want to test that will show a Stockfish win on my hardware given what you have said. Post the conditions for the test given your comments! How much time does Stockfish need to beat Lc0 on my hardware. And it will be tested. Unless Stockfish needs many hours a game to beat Lc0.

The reason for this test is that many have said Brainfish, the best Stockfish version with the best rating will beat Lc0.

Some have said the only reason that Stockfish was losing in other test to Lc0 like TCEC was a shorter book.

If you think that Stockfish can beat Lc0. Give the conditions of your test. And I will test them.

TC for SF 90min+30sec. TC for Lc0 45min+15sec.
HT on. Contempt 0.
Limit BF book to depth 12.
Everything else exactly the same as you have now.

Game on! I can cut the speed in half for Lc0 by turning off FP16. This is the best I can do in the Fritz GUI. Time control will be 90m + 30s for both engines. Lc0 will run at the staritng position at 8 Kns with FP-16 turned off vs 16 Kns with FP-16 on.

This is the end of this match. New match starting match under Milo's conditions of 1/2 speed for Lc0 on my hardware with longer TC.