Best Nets for Lc0 Page

[Hugo]

Hi there

intresting threads test:
Lc0 standalone, Leelenstein 12.1 (after go nodes 5000000) = ~ 26650 Ns
Stockfish 031219 T13; 4069Mb hash after 2 min analyse: ~13225 kNs

same test when both engines are working at the same time:
Lc0 , Leelenstein 12.1 (after go nodes 5000000) = ~ 25100 Ns
Stockfish 031219 T13; 4069Mb hash after 2 min analyse: ~12200 kNs

so it looks like both engines reduce Ns when working at the same time.
In case of Stockfish it is a minus of 6 - 8 %
In case of Lc0 it is a minus of 6 - 7 %

The impact on the results might be verry low

I will continue my tests with the same setup, have already invested 20 days.

next test candidate is NN 61653. Is it realy eye to eye with SF ?

C.K.

Hi

when the latest test is done, I will make a comparison with several constellations. For example let SF analyse endless with 13 T and compare Lc0 speed with 2 threads. same with Lc0 standalone.
Right now I could not see any difference in Lc0 with two cpu or two threads. But I did not look at it when other threads are bussy.

C.K.

[Laskos]

Hi there

intresting threads test:
Lc0 standalone, Leelenstein 12.1 (after go nodes 5000000) = ~ 26650 Ns
Stockfish 031219 T13; 4069Mb hash after 2 min analyse: ~13225 kNs

same test when both engines are working at the same time:
Lc0 , Leelenstein 12.1 (after go nodes 5000000) = ~ 25100 Ns
Stockfish 031219 T13; 4069Mb hash after 2 min analyse: ~12200 kNs

so it looks like both engines reduce Ns when working at the same time.
In case of Stockfish it is a minus of 6 - 8 %
In case of Lc0 it is a minus of 6 - 7 %

The impact on the results might be verry low

I will continue my tests with the same setup, have already invested 20 days.

next test candidate is NN 61653. Is it realy eye to eye with SF ?

C.K.

Hi

when the latest test is done, I will make a comparison with several constellations. For example let SF analyse endless with 13 T and compare Lc0 speed with 2 threads. same with Lc0 standalone.
Right now I could not see any difference in Lc0 with two cpu or two threads. But I did not look at it when other threads are bussy.

C.K.

Should be fine, as speeds here is what matters, and they are fair. There are no severe issues with heavy contention on same hyperthread, bogging down the results (say in Windows 8 some results, in Windows 10 other).

[Nordlandia]

In case of SF vs Leela.

For TC like bullet (1+0 and 1+1. Do it make sense to enable ponder at the cost as less threads available for stockfish. Time is vital in bullet and the result of ponder is somewhat better time management. 60% ponder hit mean 1.6x base time in theory.

[shrapnel]

There is a sergio1572

Yes, sergio 1572 is the latest and it certainly appears to be very strong. In the following Game, Brainfish didn't know it was beaten and kept returning 0.00 Scores until it was too late..
[pgn][Event "Rated game, 7m+3s"]
[Site "Engine Room"]
[Date "2019.12.25"]
[Round "?"]
[White "Shrapnel, Lc0 v0.23.1+git.683"]
[Black "Auryn, Brainfish 151219 64"]
[Result "1-0"]
[ECO "C42"]
[WhiteElo "2474"]
[BlackElo "2449"]
[Annotator "0.39;0.31"]
[PlyCount "133"]
[EventDate "2019.12.25"]
[SourceTitle "playchess.com"]
[TimeControl "420+3"]

{Brainfish 151219 64 POPCNT (5 cores): 30.0 plies; 5,892kN/s AMD Ryzen 7 1700X
Eight-Core Processor 3393MHz, (8 cores, 16 threads), Perfect2019.ctg,
32768 MB} 1. e4 {B 0} e5 {B 0} 2. Nf3 {B 0} Nf6 {B 0} 3. Nxe5 {B 0} d6 {B 0} 4.
Nf3 {B 0} Nxe4 {B 0} 5. d4 {B 0} d5 {B 0} 6. Bd3 {B 0} Bd6 {B 0} 7. O-O {B 0}
O-O {B 0} 8. c4 {B 0} c6 {B 0} 9. Re1 {B 0} Bf5 {B 0} 10. Qb3 {B 0} Na6 {B 0}
11. cxd5 {B 0} cxd5 {B 0} 12. Nc3 {B 0} Be6 {B 0} 13. a3 {B 0} Nc7 {B 0} 14.
Be3 {B 0} f5 {B 0} 15. Rac1 {B 0} Kh8 {B 0} 16. Nb5 {0.39/11 9} Nxb5 {B 0} 17.
Qxb5 {0.37/11 6} b6 {B 0 (a5)} 18. Bd2 {0.53/10 16} Re8 {0.31/28 18} 19. Rc2 {
0.50/11 14 (Da4)} a5 {0.00/26 22} 20. Bf1 {0.43/11 20 (Ae3)} Rf8 {0.28/27 30
(Rb8)} 21. Bc1 {0.42/11 29} f4 {0.00/31 0 (h6)} 22. Bd3 {0.70/10 9} Rf5 {
0.00/32 0} 23. Rce2 {0.69/12 10 (Da4)} Bb4 {0.00/31 27 (Qc7)} 24. axb4 {
0.87/16 13} Bd7 {0.00/33 0} 25. Qxd5 {0.97/19 7} Rxd5 {0.00/34 6} 26. Bxe4 {
1.02/16 1} Qg8 {0.00/39 82 (Bc6)} 27. Bxf4 {1.14/15 28 (Axd5)} Rf8 {0.00/34 9}
28. Bg3 {1.15/16 31 (Axd5)} Rb5 {0.00/28 9 (Bg4)} 29. h3 {1.27/14 16 (bxa5)}
axb4 {0.15/30 70 (Rxb4)} 30. Kh2 {1.56/13 25 (Te3)} Rh5 {0.20/27 22} 31. d5 {
1.55/19 0 (Ad6)} Bb5 {1.22/33 80} 32. Re3 {2.10/22 20} Rxd5 {1.30/32 10} 33.
Bxd5 {2.14/23 0} Qxd5 {1.96/34 10} 34. Re5 {2.21/25 1} Qd7 {1.96/36 5} 35. Re7
{2.24/21 1} Qd8 {2.30/37 15} 36. Bc7 {2.35/20 0} Qc8 {2.32/37 5} 37. Bxb6 {
2.53/20 22 (Ae5)} Qf5 {2.44/30 17} 38. R1e5 {2.63/22 3} Qf6 {2.57/30 2} 39. Bd4
{2.87/22 16} Qd6 {3.61/30 17} 40. Kh1 {3.07/21 2 (Rg1)} Ra8 {3.84/31 35 (Rf6)}
41. Re1 {5.91/19 41} Qh6 {3.93/33 0 (Ra1)} 42. Rxg7 {11.95/15 68 (Axg7+)} Qxg7
{3.99/29 5} 43. Bxg7+ {13.54/10 0} Kxg7 {4.11/30 5} 44. Re4 {15.21/14 15 (Te7+)
} Bc6 {3.29/26 4 (b3)} 45. Rg4+ {18.44/10 25 (Txb4)} Kf6 {3.75/29 3 (Kh6)} 46.
Rxb4 {23.70/9 5} Bd5 {4.18/35 23 (Rh8)} 47. Kh2 {25.83/8 22 (Tb6+)} Rg8 {
4.15/26 4 (Rh8)} 48. Rd4 {29.46/7 9 (Tb5)} Ke6 {4.06/24 8 (Rb8)} 49. Rd3 {
24.08/9 9 (Ce1)} h5 {3.47/24 8 (Ba8)} 50. b4 {26.10/7 10 (h4)} Kd6 {4.04/24 13
(Rb8)} 51. b5 {26.37/7 12 (Ce1)} Kc5 {3.54/24 6 (Rf8)} 52. Re3 {23.51/7 10
(Ch4)} Kxb5 {3.99/25 4 (Kd6)} 53. Nh4 {26.41/7 6 (Te5)} Kc4 {3.75/26 3 (Kc6)}
54. Re5 {28.07/7 7} Rh8 {4.46/30 0 (Rf8)} 55. Rg5 {28.17/7 9 (f4)} Kd4 {
4.46/28 6 (Rf8)} 56. Nf5+ {27.29/7 8 (Rg3)} Kd3 {4.08/22 2 (Kc3)} 57. h4 {
27.39/7 7 (Ch4)} Bf7 {4.13/25 2} 58. Ng3 {28.26/8 3} Rh6 {4.09/31 0 (Be8)} 59.
Rf5 {28.70/8 9} Rh7 {4.71/35 0} 60. Re5 {29.92/7 5 (f3)} Rh8 {4.85/30 2 (Kc3)}
61. f3 {30.24/7 6} Kd4 {4.86/37 0 (Kc2)} 62. Ra5 {30.34/7 4 (Tf5)} Kd3 {
5.13/31 3 (Kc3)} 63. Ra7 {30.53/6 5 (Te5)} Bg6 {5.27/32 8} 64. Ra6 {29.99/7 2
(Tg7)} Rg8 {5.40/33 5 (Bf5)} 65. Rb6 {31.26/7 6 (Te6)} Kc4 {5.44/28 2} 66. Re6
{30.34/8 4 (Ta6)} Kd5 {5.46/29 3 (Rg7)} 67. Ra6 {31.69/7 3 (Tb6) Auryn,
Brainfish 151219 64 abbandona (Lag: Av=0.96s, max=3.2s)} 1-0
[/pgn]

[corres]

"There is no single "best net" for Leela but there a few worth recommending for various purposes....

As CCC proves if you have 3 x RTX 2080 Ti you can use any good net (J13B.2-188, T40B.4-160, 42850, 40.T8.610,...) against Stockfish.dev using even 90 threads and your system is in pair with Stockfish.dev or even better than Stockfish.dev.
There are problems if you have only one NVIDIA GPU or especially if you have an AMD GPU or you can use only your CPU.
If you have one RTX card you should use one of the best T40 net: T40B.4-160 or -260, 42850 or 42872. Because this nets are derived from nets of Leela there is no big difference among them in Elo.
Without any RTX card you should use one of the destilled net from dkappe.
If for somebody is important every Elo that one should make a lot of test on his own system to choose the best net and parameters what is the most optimal for his own system.

[Laskos]

"There is no single "best net" for Leela but there a few worth recommending for various purposes....

As CCC proves if you have 3 x RTX 2080 Ti you can use any good net (J13B.2-188, T40B.4-160, 42850, 40.T8.610,...) against Stockfish.dev using even 90 threads and your system is in pair with Stockfish.dev or even better than Stockfish.dev.
There are problems if you have only one NVIDIA GPU or especially if you have an AMD GPU or you can use only your CPU.
If you have one RTX card you should use one of the best T40 net: T40B.4-160 or -260, 42850 or 42872. Because this nets are derived from nets of Leela there is no big difference among them in Elo.
Without any RTX card you should use one of the destilled net from dkappe.
If for somebody is important every Elo that one should make a lot of test on his own system to choose the best net and parameters what is the most optimal for his own system.

It's dependent on time control too. To real LTC, 24b nets seem to be the best for RTX GPU or a bunch of them. T40 are still very close at LTC and better at blitz. I hope the main T60 run will make 24b nets to be the strongest even in blitz, and significantly stronger at LTC, on RTX GPUs.

[corres]

There is a sergio1572

Yes, sergio 1572 is the latest and it certainly appears to be very strong.
...

It seems to me Sergio's 384x30-t40-1573 net became stronger mainly in the endgame.

[shrapnel]

It seems to me Sergio's 384x30-t40-1573 net became stronger mainly in the endgame.

At the 27th move, Lc0 said score is 1+ while Brainfish insisted it was 0.00. 27th move is End Game ? I didn't know that.
Anyway, Brainfish thought it was a fluke and came back for the re-match....and got beaten with Black....
Sergio 1573 NN really is the best when more than 1 GPU is used.
Against same level players as Auryn, Lc0 using 2 GPUs gets mostly Draws with jhorthos, Leelenfish and T 60 'nets.
More wins ONLY when using Sergio nets.
So, no matter what the Testers say, I am 100 % certain that Sergio NNs are best for people with more than 1 2080 or 2080 Ti.
[pgn][Event "Rated game, 7m+3s"]
[Site "Engine Room"]
[Date "2019.12.25"]
[Round "?"]
[White "Auryn, Brainfish 151219 64"]
[Black "Shrapnel, Lc0 v0.23.1+git.683"]
[Result "0-1"]
[ECO "B26"]
[WhiteElo "2430"]
[BlackElo "2489"]
[Annotator "-0.23;-0.11"]
[PlyCount "158"]
[EventDate "2019.12.25"]
[SourceTitle "playchess.com"]
[TimeControl "420+3"]

{Brainfish 151219 64 POPCNT (5 cores): 29.5 plies; 5,667kN/s AMD Ryzen 7 1700X
Eight-Core Processor 3393MHz, (8 cores, 16 threads), Perfect2019.ctg,
32768 MB} 1. e4 {B 0} c5 {B 0} 2. Nc3 {B 0} Nc6 {B 0} 3. g3 {B 0} g6 {B 0} 4.
Bg2 {B 0} Bg7 {B 0} 5. d3 {B 0} Rb8 {B 0} 6. Be3 {B 0} d6 {-0.11/11 12} 7. Qd2
{B 0 (a4)} b5 {-0.24/10 7} 8. Nge2 {B 0 (a3)} Nd4 {-0.26/11 15} 9. O-O {
B 0 (Nd1)} h5 {-0.24/12 12} 10. h4 {B 0 (Nd1)} b4 {-0.49/11 9} 11. Nd1 {B 0} a5
{-0.49/11 7} 12. a3 {B 0 (c3)} b3 {-0.65/13 12} 13. c3 {B 0} Nc2 {-0.68/12 6}
14. Rc1 {B 0} e5 {-0.72/12 11} 15. c4 {B 0} Nh6 {-0.77/13 14} 16. Nec3 {
-0.23/29 48 (f3)} Ng4 {-0.91/13 13 (Ae6)} 17. Nd5 {-0.28/27 10} O-O {-0.91/13
13 (Ae6)} 18. N1c3 {0.00/30 28} Kh7 {-0.89/14 1} 19. Qd1 {0.00/29 7} Rb7 {
-0.93/15 11 (Ae6)} 20. Bd2 {-0.48/28 34} Bh6 {-0.98/16 0} 21. Bxh6 {-0.59/27 12
} Nxh6 {-0.98/16 0} 22. f4 {-0.45/27 17} f6 {-1.03/14 11} 23. Qd2 {-0.57/28 18}
Rbf7 {-1.04/17 12} 24. Ne2 {-0.78/28 20 (Ne3)} exf4 {-1.23/15 31 (Ag4)} 25.
Nexf4 {-0.74/26 10 (Rxf4)} Qe8 {-1.45/13 20 (Cg4)} 26. Ne2 {-0.84/27 37 (Qxa5)}
Qe5 {-1.56/13 17} 27. Qf4 {-0.90/30 0} Bd7 {-1.58/13 16 (Dxf4)} 28. Qxe5 {
-0.59/27 5} dxe5 {-1.56/14 6} 29. Nec3 {-0.78/31 17} Nd4 {-1.56/11 1} 30. Rce1
{-0.91/34 62} Kg7 {-1.56/13 0} 31. Nb6 {-0.98/30 26 (a4)} Nc2 {-1.68/11 41} 32.
Re2 {-1.13/35 3} Ng4 {-1.67/13 28} 33. Bh3 {-1.05/35 0 (Nbd5)} Nd4 {-1.71/11 13
} 34. Ree1 {-0.82/35 1} Bc6 {-1.67/12 19} 35. Nbd5 {-1.21/36 51} a4 {-1.70/16 0
} 36. Nd1 {-1.18/33 49 (Bg2)} Re8 {-1.74/14 1} 37. N1c3 {-1.28/32 23 (Bg2)} f5
{-1.84/15 7} 38. exf5 {-1.46/32 0} gxf5 {-1.87/13 9} 39. Bxg4 {-1.45/31 4 (Bg2)
} hxg4 {-2.06/10 18} 40. Kg2 {-1.45/38 0} Nf3 {-2.09/10 16 (Td7)} 41. Re2 {
-1.45/34 5} Kg6 {-2.10/9 17} 42. Re3 {-1.45/34 4 (Rff2)} f4 {-2.10/12 14 (Cd4)}
43. h5+ {-1.88/27 16} Kxh5 {-2.04/16 0} 44. Rh1+ {-1.84/28 12} Kg5 {-2.03/17 1}
45. Ne4+ {-1.98/25 7 (gxf4+)} Kg6 {-2.04/20 17} 46. gxf4 {-2.80/29 2} exf4 {
-2.00/21 7 (Txf4)} 47. Rxf3 {-1.81/26 2} gxf3+ {-1.97/24 33 (Axd5)} 48. Kxf3 {
-2.86/29 4} Bxd5 {-1.97/24 0} 49. cxd5 {-2.86/27 2} Re5 {-1.97/22 0} 50. Rg1+ {
-2.73/26 2} Kh6 {-1.94/22 13} 51. Rh1+ {-3.22/27 3} Kg7 {-1.92/23 0} 52. Nxc5 {
-3.31/29 20} Rxd5 {-2.14/18 23} 53. Ne6+ {-3.32/30 0} Kf6 {-2.13/18 0} 54. Nxf4
{-3.28/27 5} Rc5 {-2.20/16 0} 55. Ke3 {-3.24/26 10 (d4)} Rc2 {-2.34/15 7} 56.
Nd5+ {-3.42/27 0} Kf5 {-2.37/15 3} 57. Rb1 {-3.50/26 2} Rh7 {-2.43/15 4} 58.
Nb4 {-3.93/27 0 (Nb6)} Re7+ {-2.70/13 10} 59. Kd4 {-3.85/30 0} Rd7+ {-2.80/13 2
} 60. Ke3 {-4.10/29 2} Rh2 {-2.96/12 3} 61. Rf1+ {-3.91/25 0} Ke6 {-3.01/11 5}
62. Na6 {-4.21/28 2} Kd6 {-3.18/10 5} 63. Rf6+ {-4.30/28 8} Ke5 {-3.27/11 5}
64. Rf3 {-4.32/30 0} Rxb2 {-3.30/11 7 (Td4)} 65. Nc5 {-4.39/29 5} Rd4 {-3.52/
11 3} 66. Rh3 {-4.39/27 0} Rg4 {-3.60/11 3} 67. Rh5+ {-4.46/28 13 (d4+)} Kf6 {
-4.07/9 13} 68. Kf3 {-4.53/32 0 (Rh6+)} Rbg2 {-4.30/9 3 (Tgg2)} 69. Ne4+ {
-4.67/24 10} Kg6 {-4.00/9 7 (Re7)} 70. Rb5 {-4.53/20 2 (Ra5)} Rg1 {-4.48/8 2}
71. Rb7 {-5.16/24 6 (Rb6+)} Rh4 {-7.95/7 4} 72. Rb6+ {-5.59/23 2 (Ke3)} Kg7 {
-7.25/7 7 (Rf7)} 73. Nd6 {-4.53/22 1 (Nc5)} Rh3+ {-9.24/6 3 (Th2)} 74. Ke4 {
-6.62/22 6} Rd1 {-8.15/7 3 (Tg4+)} 75. Nc4 {-5.92/24 2 (d4)} Rdxd3 {-8.29/7 2}
76. Nb2 {-6.44/26 4} Rdg3 {-7.64/7 4 (Tc3)} 77. Kf4 {-7.25/26 4 (Nxa4)} Rc3 {
-8.11/7 3 (Tf3+)} 78. Rb7+ {-7.56/27 5 (Nxa4)} Kf6 {-18.20/6 3} 79. Kg4 {
-7.90/27 2 (Rb6+)} Ke5 {-14.79/6 4 (Tcg3+) Auryn,Brainfish 151219 64 abbandona}
0-1[/pgn]

[Hai]

It seems to me Sergio's 384x30-t40-1573 net became stronger mainly in the endgame.

Sergio 1573 NN really is the best when more than 1 GPU is used.
Against same level players as Auryn, Lc0 using 2 GPUs gets mostly Draws with jhorthos, Leelenfish and T 60 'nets.
More wins ONLY when using Sergio nets.
So, no matter what the Testers say, I am 100 % certain that Sergio NNs are best for people with more than 1 2080 or 2080 Ti.

I can confirm this too.
I think that Sergio 1573 NN or the next one will be also better for people with only 1 RTX GPU.
We need to create and test even bigger Nets.

[corres]

It seems to me Sergio's 384x30-t40-1573 net became stronger mainly in the endgame.

At the 27th move, Lc0 said score is 1+ while Brainfish insisted it was 0.00. 27th move is End Game ? I didn't know that.

Why do you know my statement is based on your only one game?