Leela Zero Cpu rating?

[h1a8]

1. Are there any rating lists that shows how strong Leela Zero is (cpu version) on various hardware? I have an i7 10510u laptop with 16gb ram.

2. Is the formula about doubling the nps to give about 60 elo still valid with the cpu version of Leela?


3. Which net files can I use for Leela to get 1700-2000 elo play? I want to play against a Leela about that strength.

[ThatsIt]

Just have a look @

CEGT 40/20
http://www.cegt.net/40_40%20Rating%20Li ... s/287.html

CEGT 40/4
http://www.cegt.net/40_4_Ratinglist/40_ ... n/594.html

CEGT 5'+3" ponder=on
http://www.cegt.net/5Plus3Rating/BestVe ... EW/32.html

Best wishes,
G.S.
(CEGT team)

[h1a8]

Thanks I didnt know Blas meant cpu.

[fastgm]

http://www.fastgm.de

60 sec + 0.6 sec
10 min + 6 sec
60 min + 15 sec

are three cpu rating lists with Leela Zero cpu versions.

[dkappe]

3. Which net files can I use for Leela to get 1700-2000 elo play? I want to play against a Leela about that strength.

Ah, now the hard bit. All of the nets mentioned above are way stronger than what you require. There are basically two ways of weakening lc0 (the engine that uses the networks):

1. Use a small or minimally trained network
2. Reduce the size of the search tree
3. Both of the above

Unlike crippled ab engines which basically play like a drunk GM, some of these small nets on limited nodes can play a very entertaining brand of chess.

You can try out two small nets on limited nodes playing at lichess here:

• https://lichess.org/@/MiniHuman

• https://lichess.org/@/leela2200

The first is a 32x4 net (typically cpu nets are 128x10, gpu 256x20 or larger) running on a raspberry pi 3. It’s pretty strong for a human and may be stronger than what you require. The second is a 16x2 net distilled from the old t10 nets. It may be around 2200 lichess? They were trained from different data and should have a very different style from one another.

Anyhow, try those out and see where you end up.

[dkappe]

A recent game that may give you a taste. I think it’s looking at 4000 nodes, which even on cpu is instant.

[pgn]

[Event "Casual Classical game"]
[Site "https://lichess.org/16wxEGOl"]
[Date "2020.03.06"]
[Round "-"]
[White "Art-Vega"]
[Black "leela2200"]
[Result "0-1"]
[UTCDate "2020.03.06"]
[UTCTime "19:58:35"]
[WhiteElo "2286"]
[BlackElo "1500"]
[BlackTitle "BOT"]
[Variant "Standard"]
[TimeControl "1800+15"]
[ECO "E20"]
[Opening "Nimzo-Indian Defense: Kmoch Variation"]
[Termination "Normal"]
[Annotator "lichess.org"]

1. d4 Nf6 2. c4 e6 3. Nc3 Bb4 4. f3 { E20 Nimzo-Indian Defense: Kmoch Variation } d5 5. a3 Bxc3+ 6. bxc3 c5 7. e3 O-O 8. cxd5 exd5 9. Bd3 b6 10. Ne2 Re8 11. O-O Ba6 12. Ng3 Bxd3 13. Qxd3 Nc6 14. Ra2 Rc8 15. Re2 Qd7 16. Bb2 Na5? { (-0.05 → 1.18) Mistake. Rcd8 was best. } (16... Rcd8 17. Rd1) 17. e4 dxe4?! { (1.11 → 2.09) Inaccuracy. Kh8 was best. } (17... Kh8 18. e5 Ng8 19. f4 f6 20. Bc1 Nc4 21. Rfe1 a5 22. h3 Qc6 23. Qf3 b5 24. Kh2) 18. fxe4 cxd4 19. cxd4 Nc4 20. Bc1?! { (1.75 → 0.75) Inaccuracy. Ba1 was best. } (20. Ba1) 20... Ne5 21. Qd1 Nc4?? { (0.95 → 2.80) Blunder. Rxc1 was best. } (21... Rxc1 22. Qxc1) 22. Bg5 Ng4 23. h3 h6 24. Bc1 Nf6?! { (2.59 → 3.47) Inaccuracy. Rcd8 was best. } (24... Rcd8 25. hxg4) 25. Bxh6 Qc6 26. Bxg7 Kxg7 27. e5 Nd5 28. Nh5+?? { (3.60 → 0.95) Blunder. Nf5+ was best. } (28. Nf5+ Kf8) 28... Kg8 29. Qe1 Qg6?? { (1.11 → 3.65) Blunder. Re6 was best. } (29... Re6 30. Qf2 Qc7 31. Qf5 Kh8 32. Nf4 Nxf4 33. Rxf4 Rh6 34. Qg5 Rg6 35. Qh5+ Kg7 36. Ref2) 30. Qh4?? { (3.65 → 0.00) Blunder. Nf4 was best. } (30. Nf4) 30... Nce3? { (0.00 → 1.27) Mistake. Re6 was best. } (30... Re6 31. Rf3 Qh7 32. Qg5+ Qg6) 31. Rf6?? { (1.27 → -6.07) Blunder. Nf6+ was best. } (31. Nf6+) 31... Rc1+ 32. Kh2 Nf1+ 33. Rxf1 Rxf1 34. Re4 f5 35. Re2?! { (-5.23 → -8.10) Inaccuracy. Nf4 was best. } (35. Nf4 Nxf4 36. Rxf4 Rxf4 37. Qxf4 Rd8 38. Kg1 Rd7 39. a4 Rg7 40. Qf3 f4 41. Qd5+ Kh7) 35... Kf7 36. g4 f4 37. Rf2 Rxf2+ 38. Qxf2 Qe4 { White resigns. } 0-1

[/pgn]

[Zenmastur]

3. Which net files can I use for Leela to get 1700-2000 elo play? I want to play against a Leela about that strength.

Ah, now the hard bit. All of the nets mentioned above are way stronger than what you require. There are basically two ways of weakening lc0 (the engine that uses the networks):

1. Use a small or minimally trained network
2. Reduce the size of the search tree
3. Both of the above

Unlike crippled ab engines which basically play like a drunk GM, some of these small nets on limited nodes can play a very entertaining brand of chess.

You can try out two small nets on limited nodes playing at lichess here:

• https://lichess.org/@/MiniHuman

• https://lichess.org/@/leela2200

The first is a 32x4 net (typically cpu nets are 128x10, gpu 256x20 or larger) running on a raspberry pi 3. It’s pretty strong for a human and may be stronger than what you require. The second is a 16x2 net distilled from the old t10 nets. It may be around 2200 lichess? They were trained from different data and should have a very different style from one another.

Anyhow, try those out and see where you end up.

Can these nets be downloaded?

[Graham Banks]

CCRL 40/15

Code: Select all

CCRL 40/15 Rating List - Custom engine selection
1127982 games played by 2628 programs, run by 23 testers
Ponder off, General books (up to 12 moves), 3-4-5 piece EGTB
Time control: Equivalent to 40 moves in 15 minutes on an Intel i7-4770k.
Computed on March 7, 2020 with Bayeselo based on 1'127'982 games
Tested by CCRL team, 2005-2020, http://ccrl.chessdom.com/ccrl/4040/

Rank                 Engine                   Elo   +    -   Score  AvOp  Games
1 Lc0 0.21.3 wLD2 64-bit                  3095  +33  -33  51.2%  -15.7   284
  Lc0 0.22.0 wLD2 64-bit                  3095  +50  -49  57.9%  -54.6   126
  Lc0 0.23.2 wLD2 64-bit                  3070  +56  -56  47.2%  +11.0    88
  Lc0 0.20.1 w36089 64-bit                3054  +21  -21  57.5%  -57.0   829
  Lc0 0.21.1 w11258 64-bit                2963  +51  -50  56.0%  -38.2   126
  Lc0 0.20.1 w11258 64-bit                2953  +34  -34  50.2%   -4.2   274
  Lc0 0.20.2 w32930 64-bit                2907  +33  -33  52.8%  -21.1   282
  Lc0 0.19.1 w11248 64-bit                2863  +35  -35  50.6%   -5.1   254
  Leela Chess 0.10 w323 64-bit            2651  +18  -18  52.2%  -17.6  1102

CCRL 40/2

Code: Select all

CCRL Blitz Rating List - Custom engine selection
2205595 games played by 2595 programs, run by 22 testers
Ponder off, General books (up to 12 moves), 3-4-5 piece EGTB
Time control: From Jan 2020, equivalent to 2'+1" on an Intel i7-4770K. See the "about" page.
Computed on March 7, 2020 with Bayeselo based on 2'205'595 games
Tested by CCRL team, 2005-2020, http://ccrl.chessdom.com/ccrl/404/

Rank               Engine                Elo   +    -   Score  AvOp  Games
1 Lc0 0.20.1 w36089 64-bit           2798  +19  -19  55.1%  -41.6   982
  Lc0 0.21.0 w32930 64-bit           2757  +80  -94  12.8% +319.9    82
  Lc0 0.20.2 w32930 64-bit           2679  +80  -90  18.8% +239.0    64
  Leela Chess 0.10 w483 64-bit       2607  +22  -22  48.6%  +11.1   725
  Leela Chess 0.10 w323 64-bit       2538  +23  -22  57.2%  -55.7   754
  Leela Chess 0.10 w272 64-bit       2309  +33  -32  56.1%  -46.9   350
  Leela Chess 0.9 w272 64-bit        2272  +33  -33  51.3%   -9.8   352
  Leela Chess 0.7 w187 64-bit        2236  +22  -22  63.3% -104.8   814
  Leela Chess 0.5 w123 64-bit        1964  +32  -32  46.9%  +24.7   373

[dkappe]

The distilled networks can be found here: https://github.com/dkappe/leela-chess-w ... d-Networks

The mean girl nets can be found here: https://github.com/dkappe/leela-chess-w ... -style-net

[dkappe]

CCRL 40/15

Code: Select all

CCRL 40/15 Rating List - Custom engine selection
1127982 games played by 2628 programs, run by 23 testers
Ponder off, General books (up to 12 moves), 3-4-5 piece EGTB
Time control: Equivalent to 40 moves in 15 minutes on an Intel i7-4770k.
Computed on March 7, 2020 with Bayeselo based on 1'127'982 games
Tested by CCRL team, 2005-2020, http://ccrl.chessdom.com/ccrl/4040/

Rank                 Engine                   Elo   +    -   Score  AvOp  Games
1 Lc0 0.21.3 wLD2 64-bit                  3095  +33  -33  51.2%  -15.7   284
  Lc0 0.22.0 wLD2 64-bit                  3095  +50  -49  57.9%  -54.6   126
  Lc0 0.23.2 wLD2 64-bit                  3070  +56  -56  47.2%  +11.0    88
  Lc0 0.20.1 w36089 64-bit                3054  +21  -21  57.5%  -57.0   829
  Lc0 0.21.1 w11258 64-bit                2963  +51  -50  56.0%  -38.2   126
  Lc0 0.20.1 w11258 64-bit                2953  +34  -34  50.2%   -4.2   274
  Lc0 0.20.2 w32930 64-bit                2907  +33  -33  52.8%  -21.1   282
  Lc0 0.19.1 w11248 64-bit                2863  +35  -35  50.6%   -5.1   254
  Leela Chess 0.10 w323 64-bit            2651  +18  -18  52.2%  -17.6  1102

CCRL 40/2

Code: Select all

CCRL Blitz Rating List - Custom engine selection
2205595 games played by 2595 programs, run by 22 testers
Ponder off, General books (up to 12 moves), 3-4-5 piece EGTB
Time control: From Jan 2020, equivalent to 2'+1" on an Intel i7-4770K. See the "about" page.
Computed on March 7, 2020 with Bayeselo based on 2'205'595 games
Tested by CCRL team, 2005-2020, http://ccrl.chessdom.com/ccrl/404/

Rank               Engine                Elo   +    -   Score  AvOp  Games
1 Lc0 0.20.1 w36089 64-bit           2798  +19  -19  55.1%  -41.6   982
  Lc0 0.21.0 w32930 64-bit           2757  +80  -94  12.8% +319.9    82
  Lc0 0.20.2 w32930 64-bit           2679  +80  -90  18.8% +239.0    64
  Leela Chess 0.10 w483 64-bit       2607  +22  -22  48.6%  +11.1   725
  Leela Chess 0.10 w323 64-bit       2538  +23  -22  57.2%  -55.7   754
  Leela Chess 0.10 w272 64-bit       2309  +33  -32  56.1%  -46.9   350
  Leela Chess 0.9 w272 64-bit        2272  +33  -33  51.3%   -9.8   352
  Leela Chess 0.7 w187 64-bit        2236  +22  -22  63.3% -104.8   814
  Leela Chess 0.5 w123 64-bit        1964  +32  -32  46.9%  +24.7   373

Looking at these, it’s not clear which networks are used. The 32930 and 11258 nets themselves would be terrible on cpu. Are these distilled networks?