Развернуть для просмотра

Владимир001 · 05.11.2020, 03:30

Развернуть для просмотра

Hardware: i5-8400 CPU @ 2.80GHz, 6 cores, 32GB, gpu RTX 2060, 6GB
Gauntlet: SF-dev vs. LC0 24x320 nets
Network IDs: T60B.7-105, T60B.7-135, 64255, 65162, 65297, 65373, 65374, 65726, 65762
LC0-version: v0.26.0
SF-version: SF-dev 20092708 (NNUE hybrid)
Software: Cutechess-CLI
Time control: LC0: 1,500 nodes/move, SF-dev: 300,000 nodes/move @ 4 Threads, 512MB Hash
Book: Chad's 8-ply book with 10000 positions, order=sequential
Tablebases: 6-men
Adjudication: -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Ordo: -a 0 -A "Stockfish.dev" -W -D -U "0,1,2,3,4,5,6,7,8,9,10" -s 2000 -n 1
Comment:

Код:

   # PLAYER                :  RATING  ERROR   POINTS  PLAYED   (%)  CFS(%)      W      D      L  D(%)
+  1 lc0.net.65762         :    34.4    3.2  10952.0   20000    55      92   5494  10916   3590    55
+  2 lc0.net.65726         :    31.1    3.2  10862.0   20000    54      66   5317  11090   3593    55
   3 lc0.net.65297         :    30.2    3.2  10836.0   20000    54      59   5374  10924   3702    55
   4 lc0.net.65373         :    29.7    3.2  10822.0   20000    54      85   5360  10924   3716    55
   5 lc0.net.65162         :    27.2    3.3  10754.0   20000    54      94   5255  10998   3747    55
   6 lc0.net.65374         :    23.7    3.2  10657.5   20000    53      90   5127  11061   3812    55
   7 lc0.net.T60B.7-135    :    20.8    3.1  10576.5   20000    53      63   4994  11165   3841    56
   8 lc0.net.T60B.7-105    :    20.0    3.3  10556.0   20000    53     100   4984  11144   3872    56
   9 lc0.net.64255         :     3.6    3.3  10100.0   20000    51      98   4463  11274   4263    56
- 10 Stockfish.dev         :     0.0   ----  83884.0  180000    47     ---  34136  99496  46368    55

White advantage = 57.39 +/- 0.54
Draw rate (equal opponents) = 57.36 % +/- 0.12

PGN:
match-166

match-167

[свернуть]

Владимир001 · 05.11.2020, 11:21

вот забавная книга открытия гандикапа для матчей двигателя. Его всего 20 отверстий, они необычны, но забавны для просмотра и могут помочь увеличить масштаб elo для матчей процессоров. Каждое открытие содержит оценку SF13 D33 от первого выхода из книги (в тегах аннотаций). Каждое открытие неуклонно увеличивает eval для белых, колеблясь от +.70 до +2.65.

handicaps20.rar

Владимир001 · 05.11.2020, 11:24

MSE пешки для 10000 позиций, вычисленных на depth1 v depth2. Может быть полезно установить шкалу:

Код:

MSE (pawns): depth 4 vs 5 = 3.07373 ;  depth 4 vs 21 = 5.02954 
MSE (pawns): depth 5 vs 6 = 2.73058 ;  depth 5 vs 21 = 3.95479 
MSE (pawns): depth 6 vs 7 = 1.96972 ;  depth 6 vs 21 = 2.77346 
MSE (pawns): depth 7 vs 8 = 1.6978 ;  depth 7 vs 21 = 2.25397 
MSE (pawns): depth 8 vs 9 = 1.32313 ;  depth 8 vs 21 = 1.75539 
MSE (pawns): depth 9 vs 10 = 1.20607 ;  depth 9 vs 21 = 1.45507 
MSE (pawns): depth 10 vs 11 = 0.747268 ;  depth 10 vs 21 = 0.858147 
MSE (pawns): depth 11 vs 12 = 0.381017 ;  depth 11 vs 21 = 0.428419 
MSE (pawns): depth 12 vs 13 = 0.270915 ;  depth 12 vs 21 = 0.33081 
MSE (pawns): depth 13 vs 14 = 0.203332 ;  depth 13 vs 21 = 0.261033 
MSE (pawns): depth 14 vs 15 = 0.166416 ;  depth 14 vs 21 = 0.213916 
MSE (pawns): depth 15 vs 16 = 0.143403 ;  depth 15 vs 21 = 0.181549 
MSE (pawns): depth 16 vs 17 = 0.124597 ;  depth 16 vs 21 = 0.161026 
MSE (pawns): depth 17 vs 18 = 0.112789 ;  depth 17 vs 21 = 0.138084 
MSE (pawns): depth 18 vs 19 = 0.1048 ;  depth 18 vs 21 = 0.122028 
MSE (pawns): depth 19 vs 20 = 0.0967496 ;  depth 19 vs 21 = 0.101625 
MSE (pawns): depth 20 vs 21 = 0.0939 ;  depth 20 vs 21 = 0.0939

Владимир001 · 05.11.2020, 15:08

"CCWC CPU\GPU Candidates League' 1min+1secs 4cpu и 1gpu (38%мощности) со случайными отверстиями перевернуты, из новых 650 отверстий SuFi\2ply book compilation

Владимир001 · 05.11.2020, 20:11

Развернуть для просмотра

Match: J92-330 vs stockfish 12 in fixed nodes
LC0 version: 0.26.2, built with cuda11 + cudnn7
LC0 options: cudnn-fp16, 1 thread, default for the rest
SF options: 1 thread, 1GB Hash.
Time control: Leela: 1 kn/move, SF 12: 400Kn/move
Hardware: RTX 2070, i7-7700
Book: Balsa 2724 openings
Tablebases: 6-man TB for both.
Adjudication: 6-man TB, -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Software:cutechess-cli
Comments: J92-330 got -0.5 elo vs SF 12, so close to win!!.

Код:

Leela: 1 thread, 1Kn/move
SF 12: 1 thread, 400Kn/move

Balsa 2724

   # PLAYER                :  RATING  ERROR   POINTS  PLAYED   (%)  CFS(%)      W      D      L
-  1 Stockfish_12.400k     :       0   ----  33136.0   65376  50.7      56  12935  40402  12039
+  2 lc0.net.J92-330.1k    :       0      6   2720.5    5448  49.9      51   1023   3395   1030
   3 lc0.net.J92-270.1k    :      -1      6   2720.0    5448  49.9      68   1063   3314   1071
   4 lc0.net.J92-290.1k    :      -2      6   2705.5    5448  49.7      53   1012   3387   1049
   5 lc0.net.J92-260.1k    :      -3      6   2703.0    5448  49.6      54   1004   3398   1046
   6 lc0.net.J92-280.1k    :      -3      6   2699.5    5448  49.6      54   1045   3309   1094
   7 lc0.net.J92-190.1k    :      -4      6   2696.0    5448  49.5      54    996   3400   1052
   8 lc0.net.J92-250.1k    :      -4      6   2693.0    5448  49.4      61   1015   3356   1077
   9 lc0.net.J92-230.1k    :      -5      6   2684.0    5448  49.3      51    980   3408   1060
  10 lc0.net.J92-300.1k    :      -5      6   2683.5    5448  49.3      55   1009   3349   1090
  11 lc0.net.J92-320.1k    :      -6      6   2679.5    5448  49.2      77   1004   3351   1093
  12 lc0.net.J92-220.1k    :      -9      6   2656.0    5448  48.8      96    978   3356   1114
  13 lc0.net.J92-100.1k    :     -16      6   2599.5    5448  47.7     ---    910   3379   1159

https://pixeldrain.com/u/RdGzgt1i

[свернуть]

Владимир001 · 06.11.2020, 00:29

Развернуть для просмотра

Match: LJ2 Lender net vs representative T60 net
Lc0-version: 26.1
Lc0 options: MovesLeftMaxEffect=0.1 option.MovesLeftThreshold=0.0 option.MovesLeftSlope=0.007 option.MovesLeftQuadraticFactor=1.0 option.MovesLeftScaledFactor=0.0 option.MovesLeftConstantFactor=0.0 MovesLeftConstantFactor=0.0 MoveOverheadMs=0
Time/Node control: 800 fixed nodes
Hardware: RTX 2080/RTX 2070, AMD 2950X
Book: endgame-16-piece-book_Q-0.0-0.6.pgn, random with reverse.
Tablebases: 6-man Syzygy
Adjudication: 6-man Syzygy and -draw movenumber=50 movecount=6 score=10 -resign movecount=5 score=500 twosided=true
Chess Interface: cutechess-cli
Comments: Ongoing tests of LJ2 (320x24b) endgame training progress.
Results:

Код:

   # PLAYER             :  RATING  ERROR  POINTS  PLAYED   (%)    W    L     D  D(%)  CFS(%)
   1 lc0.net.65560      :     0.0   ----  2185.0    4000    55  780  410  2810    70     100
   2 lc0.net.LJ2-140    :   -21.2    8.0   939.5    2000    47  214  335  1451    73     100
   3 lc0.net.LJ2-80     :   -43.9    8.7   875.5    2000    44  196  445  1359    68     ---

https://pixeldrain.com/u/PtoavbFM

[свернуть]

Владимир001 · 06.11.2020, 11:18

Развернуть для просмотра

Match Gauntlet
LC0-version: lc0-v0.26.3 cudnn-version
LC0 options: Standard option except Threads=1, NNCacheSize=0, MinibatchSize=1.
Time control: 1 Node/move
Hardware: i7-4790 with 16 GByte RAM + GTX1650
Speed: Not important for this type of testing.
Book: Chad-opening book with 100 000 different openings with the most popular human openings.
Book options: Opening order: Sequential. Play each opening twice. Play all opening in the book both for White and Black.
Tablebases: No Tablebases.
Adjudication: Only 5-men syzygy
Software: CuteChess 1.2.0 https://github.com/cutechess/cutechess/releases / Graphic tools is programmed by glbchess64, in Python programming tools.
Comment: This is result out of Ordo 1.2.6 https://github.com/michiguel/Ordo/releases/tag/v1.2.6
Ordo string: ordo.exe -Q -N 0 -D -a 0 -A "J92-145" -W -n4 -s500 -G -V -U "0,1,2,3,4,5,6,7,8,9,10" -j stats.txt -o rating.txt -p games.pgn
Results:
pgn-file: https://ufile.io/0170kecb
Explanation: Trying to find the best net in clean policy Elo.

Код:

   # PLAYER     :  RATING  ERROR    POINTS  PLAYED   (%)  CFS(%)      W      D      L  D(%)
   1 J92-320    :      17      1  104871.0  200000  52.4     100  67558  74626  57816  37.3
   2 J92-145    :       0      1   95129.0  200000  47.6     ---  57816  74626  67558  37.3

White advantage = 24.54 +/- 0.56
Draw rate (equal opponents) = 37.53 % +/- 0.10

Код:

My list until now (Elo against J92-145) 1 Node/move:
White advantage = 23.98 +/- 0.57
Draw rate (equal opponents) = 36.51 % +/- 0.10
J92-310:             19     Elo    (Test with 200k games)
J92-320:             17     Elo    (Test with 200k games)
J92-270:             17     Elo    (Test with 200k games)
J92-280:             11     Elo    (Test with 200k games)
J92-260:             11     Elo    (Test with 200k games)
J92-290:             10     Elo    (Test with 200k games)
J92-240:              8     Elo    (Test with 200k games)
J92-210:              5     Elo    (Test with 200k games)
J92-220:              3     Elo    (Test with 200k games)
J92-205:              3     Elo    (Test with 200k games)
J92-230:              2     Elo    (Test with 200k games)
J92-190:              1     Elo    (Test with 200k games)
J92-145:              0     Elo    (Reference)
J92-100:             -2     Elo    (Test with 200k games)

[свернуть]

Alvir · 06.11.2020, 22:15

Тест с дискорда:

Код:

ClementinСегодня, в 18:02
Nodes match: J92 5000 npm vs SF 12 2T 2500000 npm
Openings: TCEC SuFi 10-19 (500 pos.)
Hardware: Ryzen 7 3750H/GTX 1660 Ti, Ryzen 9 4900H/RTX 2060
Lc0 26.3: nncache=100000, threads=1
MLH: moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-constant-factor=0, moves-left-scaled-factor=1, moves-left-quadratic-factor=0
SF 12: threads=2, hash=64
TBs and adj.: syzygy 5-men, draw 5 moves 5cp move 30, resign (two-sided) 5 moves 600cp
Comments: J92-330, J92-320, J92-260 added. Very good performance for J92-330 here.
Games: https://gofile.io/d/UKxPJj
# PLAYER                 :  RATING  ERROR  PLAYED   (%)  CFS    W     D     L   D(%)
1 lc0-263-j92-330-5k     :    22.9   15.4   1000  53.05   80   263   535   202  53.50
2 lc0-263-j92-310-5k     :    13.5   15.1   1000  51.80   57   247   542   211  54.20
3 lc0-263-j92-320-5k     :    11.6   13.4   1000  51.55   59   248   535   217  53.50
4 lc0-263-j92-270-5k     :     9.4   12.4   1000  51.25   70   230   565   205  56.50
5 lc0-263-j92-280-5k     :     3.7   16.5   1000  50.50   54   239   532   229  53.20
6 lc0-263-J92-190-5k     :     2.6   14.9   1000  50.35   52   227   553   220  55.30
7 lc0-263-j92-260-5k     :     2.2   12.9   1000  50.30   63   226   554   220  55.40
8 stockfish-12-2T-2500k  :     0.0   ----   7000  48.74  ---  1504  3816  1680  54.51

White advantage = 90.47 +/- 2.93
Draw rate (equal opponents) = 59.10 % +/- 0.59

Alvir · 06.11.2020, 22:22

Мини матч:(BanksiaGui)

Код:

Result:
-----------------------------------------------------------------------------
  #  name             games    wins   draws  losses  score%    los%  elo+/-
  1. lc0-v0.26.3-330    100      26      57      17    54.5    91.5    31.4
  2. lc0-v0.26.3-190    100      17      57      26    45.5     8.5   -31.4

Cross table:
-----------------------------------------------------------------------------
  #  name               score%   games         1         2
  1. lc0-v0.26.3-330      54.5     100         x      54.5
  2. lc0-v0.26.3-190      45.5     100      45.5         x

Tech:
-----------------------------------------------------------------------------

Tech (average nodes, depths, time/m per move, others per game), counted for computing moves only, ignored moves with zero nodes:
  #  name               nodes/m         NPS  depth/m   time/m    moves     time
  1. lc0-v0.26.3-330         3K       10692      8.5      0.3     45.4     11.4
  2. lc0-v0.26.3-190         3K       10556      8.5      0.3     45.4     11.5
     all ---                 2K       10624      8.5      0.3     45.4     11.4

RTX 2060super, 3k nodes per move, book-SuFi19-TCEC19_NoomenSelect

Владимир001 · 06.11.2020, 23:20

Цитата: Сообщение от Alvir

Тест с дискорда:

Код:

ClementinСегодня, в 18:02
Nodes match: J92 5000 npm vs SF 12 2T 2500000 npm
Openings: TCEC SuFi 10-19 (500 pos.)
Hardware: Ryzen 7 3750H/GTX 1660 Ti, Ryzen 9 4900H/RTX 2060
Lc0 26.3: nncache=100000, threads=1
MLH: moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-constant-factor=0, moves-left-scaled-factor=1, moves-left-quadratic-factor=0
SF 12: threads=2, hash=64
TBs and adj.: syzygy 5-men, draw 5 moves 5cp move 30, resign (two-sided) 5 moves 600cp
Comments: J92-330, J92-320, J92-260 added. Very good performance for J92-330 here.
Games: https://gofile.io/d/UKxPJj
# PLAYER                 :  RATING  ERROR  PLAYED   (%)  CFS    W     D     L   D(%)
1 lc0-263-j92-330-5k     :    22.9   15.4   1000  53.05   80   263   535   202  53.50
2 lc0-263-j92-310-5k     :    13.5   15.1   1000  51.80   57   247   542   211  54.20
3 lc0-263-j92-320-5k     :    11.6   13.4   1000  51.55   59   248   535   217  53.50
4 lc0-263-j92-270-5k     :     9.4   12.4   1000  51.25   70   230   565   205  56.50
5 lc0-263-j92-280-5k     :     3.7   16.5   1000  50.50   54   239   532   229  53.20
6 lc0-263-J92-190-5k     :     2.6   14.9   1000  50.35   52   227   553   220  55.30
7 lc0-263-j92-260-5k     :     2.2   12.9   1000  50.30   63   226   554   220  55.40
8 stockfish-12-2T-2500k  :     0.0   ----   7000  48.74  ---  1504  3816  1680  54.51

White advantage = 90.47 +/- 2.93
Draw rate (equal opponents) = 59.10 % +/- 0.59

Партии :
https://gofile.io/d/UKxPJj

Владимир001 · 07.11.2020, 13:57

Развернуть для просмотра

Match: Match of of SF Dev 20092822 4CPU vs Lc0 J92-330
Lc0-version: v0.26.2 + v0.26.3 (since J92-260)
Lc0-options : -Backend=cudnn -NNCacheSize=2000000 -MaxOutOfOrderEvalsFactor=2.4 -MaxCollisionEvents=256 -TwoFoldDraws=true
MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0
SF-version: Stockfish 20092822 x64 bmi2 (date: 28.09.2020)
SF/Cfish-options: threads=4 hash=2048 MB, contempt=24, NNUE=true, net=nn-baeb9ef2d183.nnue (default)
Time control: 5min + 5s
Hardware: notebook with mobile i7-7700HQ 4x2.8GHz (8 threads with HT on), mobile Geforce GTX 1050 TI 4GB
Book: Balsa_Special.pgn, 10 plies, sequentially, colors reversed
Tablebases: Syzygy 5-men on SSD
Speed: J92: 961 nps, SF Dev NNUE 4CPU: 3680k nps, Cfish NNUE 4CPU: 4198k nps
Adjudication: Draw at 50 moves rule, 3-fold repetition, after 200 moves, at 50 moves with eval = 0.00 ; 5-men Gaviota tablebase adjudication
Software: Arena 3.5.1, ponder = off
Elo rating: Ordo 1.2.6
Comments: J92-330 added. ERET test results are included in the results table (60s time limit per position).

Код:

   # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)  ERET test
   1 Cfish-NNUE 071020_4CPU  :     7.1   22.6    51.0     100  51.0      73    2   98    0  98.0  96/111
   2 Stockfish_20092822_4CPU :     0.0   ----   637.0    1100  57.9     100  190  894   16  81.3  95/111
   3 Lc0.net.JH.J92-280      :   -46.2   25.9    43.5     100  43.5      58    3   81   16  81.0  80/111
   4 Lc0.net.JH.J92-330      :   -49.8   26.3    43.0     100  43.0      50    4   78   18  78.0  79/111    <= NEW
   5 Lc0.net.JH.J92-260      :   -49.8   26.9    43.0     100  43.0      58    1   84   15  84.0  81/111
   6 Lc0.net.JH.J92-210      :   -53.4   26.0    42.5     100  42.5      50    1   83   16  83.0  81/111
   7 Lc0.net.JH.J92-270      :   -53.4   26.0    42.5     100  42.5      77    1   83   16  83.0  81/111
   8 Lc0.net.JH.J92-220      :   -67.9   28.7    40.5     100  40.5      57    2   77   21  77.0  79/111
   9 Lc0.net.JH.J92-190      :   -71.6   28.3    40.0     100  40.0      50    1   78   21  78.0  79/111
  10 Lc0.net.JH.J92-310      :   -71.6   29.3    40.0     100  40.0      57    1   78   21  78.0  77/111
  11 Lc0.net.JH.J92-240      :   -75.3   30.0    39.5     100  39.5      75    0   79   21  79.0  80/111
  12 Stockfish_11_4CPU       :   -90.2   31.3    37.5     100  37.5     ---    0   75   25  75.0  91/111

https://pixeldrain.com/u/9fbKd2at

[свернуть]

Владимир001 · 07.11.2020, 14:02

Развернуть для просмотра

Hardware: i7-8770K CPU @ 3.70GHz, 6 cores (12 Threads), 32GB, Gpu - RTX 2080 | Lc0 vs SF Dev Scaling Continuation
Gauntlet: SF-dev vs. LC0 30x384 net
Network IDs: Lc0 J92-270
LC0-version: v0.26.3
SF-version: SF-dev 5th Oct (NNUE hybrid)
Software: Cutechess-CLI
Time control: Lc0 80k vs SF NNUE 28000k per move. | Lc0 with default settings and 1 thread, SF NNUE with default settings and 4 threads.
Book: 8moves_v3.pgn (SF Regression book at LTC)
Tablebases: 6-men Synergy WDL with TB Adjudication
Adjudication: -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Ordo: ordo-win64.exe -Q -N "1,1" -D -a 0 -A "Stockfish_Dev4T" -W -n4 -s5000 -p games.pgn -U "0,1,2,3,4,5,6,7,8,9,10" -o rating.txt
Comment: This is the continuation of the my scaling test at 80000 nodes of Lc0 Jh 270 vs 28000k nodes SF NNUE (5th Oct).

Код:

 # PLAYER              :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 Stockfish_Dev_4T    :     0.0   ----  1011.0    2000  50.5      93  130  1762  108  88.1
   2 Lc0 Jh 270          :    -3.9    5.2   989.0    2000  49.5     ---  108  1762  130  88.1

White advantage = 24.58 +/- 2.72
Draw rate (equal opponents) = 90.19 % +/- 0.74

https://pixeldrain.com/l/scuMCCBF#item=0
2 файла

[свернуть]

Владимир001 · 08.11.2020, 18:35

Развернуть для просмотра

Hardware: i5-8400 CPU @ 2.80GHz, 6 cores, 32GB, gpu RTX 2060, 6GB
Gauntlet: SF-dev vs. LC0 24x320 nets
Network IDs: T60B.7-105, T60B.7-135, 64255, 65162, 65297, 65373, 65374, 65726, 65762, 65830, 65889
LC0-version: v0.26.0
SF-version: SF-dev 20092708 (NNUE hybrid)
Software: Cutechess-CLI
Time control: LC0: 1,500 nodes/move, SF-dev: 300,000 nodes/move @ 4 Threads, 512MB Hash
Book: Chad's 8-ply book with 10000 positions, order=sequential
Tablebases: 6-men
Adjudication: -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Ordo: -a 0 -A "Stockfish.dev" -W -D -U "0,1,2,3,4,5,6,7,8,9,10" -s 2000 -n 1
Comment

Код:

   # PLAYER                :  RATING  ERROR    POINTS  PLAYED   (%)  CFS(%)      W       D      L  D(%)
+  1 lc0.net.65889         :    35.5    3.2   10982.0   20000    55      68   5529   10906   3565    55
   2 lc0.net.65762         :    34.4    3.2   10952.0   20000    55      65   5494   10916   3590    55
+  3 lc0.net.65830         :    33.5    3.2   10927.5   20000    55      85   5457   10941   3602    55
   4 lc0.net.65726         :    31.1    3.2   10862.0   20000    54      66   5317   11090   3593    55
   5 lc0.net.65297         :    30.2    3.2   10836.0   20000    54      59   5374   10924   3702    55
   6 lc0.net.65373         :    29.7    3.2   10822.0   20000    54      85   5360   10924   3716    55
   7 lc0.net.65162         :    27.2    3.2   10754.0   20000    54      93   5255   10998   3747    55
   8 lc0.net.65374         :    23.7    3.2   10657.5   20000    53      90   5127   11061   3812    55
   9 lc0.net.T60B.7-135    :    20.8    3.3   10576.5   20000    53      63   4994   11165   3841    56
  10 lc0.net.T60B.7-105    :    20.0    3.1   10556.0   20000    53     100   4984   11144   3872    56
  11 lc0.net.64255         :     3.6    3.2   10100.0   20000    51      99   4463   11274   4263    56
- 12 Stockfish.dev         :     0.0   ----  101974.5  220000    46     ---  41303  121343  57354    55

White advantage = 57.57 +/- 0.50
Draw rate (equal opponents) = 57.29 % +/- 0.11
PGN
2 файла 
https://pixeldrain.com/l/z5SxUbyK#item=0

[свернуть]

Владимир001 · 10.11.2020, 00:53

Развернуть для просмотра

Match: Gauntlet SF201028 vs J92-270 & J92-190
Hardware: Geforce RTX 2070S & AMD Threadripper 3970X
Lc0 version: v0.26.3
Lc0 options: Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 option.MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0
SF options: Threads=23 Hash=4096
Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks)
Time control: 5m+5s
Book: openings-dd-4ply-100.pgn, sequential, color reversed
Tablebases: 6-man syzygy
Adjudication: 6-man TB, -draw movenumber=35 movecount=5 score=15 -resign movecount=5 score=1000
Software: cutechess-cli
Comments: J92-190 was used to establish a baseline with my new CPU. I have limited the number of threads used by stockfish to match it to the RTX2070S with the ratio set to 875 with lc0 running a 256x20 net. The time control was chosen so that lc0 got more than 100,000 nodes per move. J92-270 was the first comparison test after performing an SPRT test of J92-270 vs J92-190 (see next post).

Код:

   # PLAYER              :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L
   1 stockfish_201028    :     0.0   ----   208.5     400  52.1      97   23  371    6
   2 lc0.net.J92-270     :   -12.3   12.4    96.5     200  48.3      72    3  187   10
   3 lc0.net.J92-190     :   -17.6   12.8    95.0     200  47.5     ---    3  184   13

White advantage = 20.21 +/- 4.64
Draw rate (equal opponents) = 96.40 % +/- 1.3

2 файла :
https://pixeldrain.com/l/qRe4YPW6#item=0

[свернуть]

Владимир001 · 10.11.2020, 01:01

Развернуть для просмотра

Match: SPRT test of J92-270 vs J92-190
Hardware: Geforce RTX 2070S & AMD Threadripper 3970X
Lc0 version: v0.26.3
Lc0 options: Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 option.MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0
Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks)
Time control: 1.6s+0.03s
Book: book-ply6-unifen-Q-0.0-0.25.pgn, random, color reversed
Tablebases: 6-man syzygy
Adjudication: 6-man TB, -draw movenumber=20 movecount=5 score=10 -resign movecount=3 score=500
Software: cutechess-cli
Comments: My new process is to use an SPRT test to determine which nets to then test against SFdev at 5m+5s.

Код:

   # PLAYER             :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)     W     D     L
   1 lc0.net.J92-270    :     5.3    3.9  4912.5    9678  50.8     100  1556  6713  1409
   2 lc0.net.J92-190    :     0.0   ----  4765.5    9678  49.2     ---  1409  6713  1556

White advantage = 20.49 +/- 1.99
Draw rate (equal opponents) = 69.78 % +/- 0.46

Код:

python sprt.py --elo0=0 --elo1=3 --results 1409 6713 1556:
[Elo0,Elo1]                 :  [0.00,3.00]
Confidence level            :  95.00%
Elo                         :  4.78
Confidence interval         :  [0.77,8.72] (95.00%)
LOS                         :  99.00%
LLR [u,l]                   :  2.94  [-2.94,2.94

https://pixeldrain.com/u/5f8FW94C

[свернуть]

Владимир001 · 10.11.2020, 17:14

Развернуть для просмотра

Match: SPRT test of J92-280 vs J92-270
Hardware: Geforce RTX 2070S & AMD Threadripper 3970X
Lc0 version: v0.26.3
Lc0 options: Threads=1 Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0
Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks)
Time control: 1.6s+0.03s
Book: book-ply6-unifen-Q-0.0-0.25.pgn, random, color reversed
Tablebases: 6-man syzygy
Adjudication: 6-man TB, -draw movenumber=20 movecount=5 score=10 -resign movecount=3 score=500
Software: cutechess-cli

Код:

   # PLAYER             :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)     W     D     L
   1 lc0.net.J92-280    :     5.1    3.8  5148.0   10149  50.7     100  1608  7080  1461
   2 lc0.net.J92-270    :     0.0   ----  5001.0   10149  49.3     ---  1461  7080  1608

White advantage = 26.89 +/- 1.96
Draw rate (equal opponents) = 70.48 % +/- 0.46

python sprt.py --elo0=0 --elo1=3 --results 1461 7080 1608:

Код:

[Elo0,Elo1]                 :  [0.00,3.00]
Confidence level            :  95.00%
Elo                         :  4.53
Confidence interval         :  [0.64,8.35] (95.00%)
LOS                         :  98.85%
LLR [u,l]                   :  2.94  [-2.94,2.94

https://pixeldrain.com/u/PC39bmVM

[свернуть]

svoitsl · 10.11.2020, 20:01

3000 самоЭло!!!!

Владимир001 · 11.11.2020, 00:36

Развернуть для просмотра

Network IDs: T60B.7-105, T60B.7-135, 64255, 65162, 65297, 65373, 65374, 65726, 65762, 65830, 65889, 65939
LC0-version: v0.26.0
SF-version: SF-dev 20092708 (NNUE hybrid)
Software: Cutechess-CLI
Time control: LC0: 1,500 nodes/move, SF-dev: 300,000 nodes/move @ 4 Threads, 512MB Hash
Book: Chad's 8-ply book with 10000 positions, order=sequential
Tablebases: 6-men
Adjudication: -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Ordo: -a 0 -A "Stockfish.dev" -W -D -U "0,1,2,3,4,5,6,7,8,9,10" -s 2000 -n 1
Comment:

Код:

   # PLAYER                :  RATING  ERROR    POINTS  PLAYED   (%)  CFS(%)      W       D      L  D(%)
+  1 lc0.net.65939         :    38.2    3.3   11058.0   20000    55      88   5592   10932   3476    55
   2 lc0.net.65889         :    35.5    3.3   10982.0   20000    55      67   5529   10906   3565    55
   3 lc0.net.65762         :    34.4    3.2   10952.0   20000    55      65   5494   10916   3590    55
   4 lc0.net.65830         :    33.5    3.2   10927.5   20000    55      85   5457   10941   3602    55
   5 lc0.net.65726         :    31.1    3.3   10862.0   20000    54      66   5317   11090   3593    55
   6 lc0.net.65297         :    30.2    3.2   10836.0   20000    54      59   5374   10924   3702    55
   7 lc0.net.65373         :    29.7    3.2   10822.0   20000    54      85   5360   10924   3716    55
   8 lc0.net.65162         :    27.2    3.2   10754.0   20000    54      93   5255   10998   3747    55
   9 lc0.net.65374         :    23.7    3.2   10657.5   20000    53      90   5127   11061   3812    55
  10 lc0.net.T60B.7-135    :    20.8    3.3   10576.5   20000    53      62   4994   11165   3841    56
  11 lc0.net.T60B.7-105    :    20.0    3.3   10556.0   20000    53     100   4984   11144   3872    56
  12 lc0.net.64255         :     3.6    3.3   10100.0   20000    51      98   4463   11274   4263    56
- 13 Stockfish.dev         :     0.0   ----  110916.5  240000    46     ---  44779  132275  62946    55

White advantage = 57.59 +/- 0.47
Draw rate (equal opponents) = 57.27 % +/- 0.11

PGN:
https://pixeldrain.com/u/EUdjbKrt

[свернуть]

Владимир001 · 11.11.2020, 00:41

Развернуть для просмотра

Match: 30b nets J92-190 vs J92-330 in time control
LC0 version:0.26.3
LC0 options: 1 thread, move-overhead=50
Time control: 5s+0.5s
Hardware: RTX 2070
Book: Sufi book (293 openings)
Tablebases: 6-man TB for both.
Speed: 9.1 knps (benchmark with 1 thread)
Adjudication: 6-man TB, -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=800
Software:cutechess-cli
Comments: J92-330 beats J92-190 quite hard (+88 -45 =453, +26 elo). I don't know why J92-330 overperforms here. In low node test (about 1 knpm), J92-330 is good, but is not as strong as in this test (about 10 knpm). Does it scale much better in high nodes (TCEC bonus 3 result shows J92-330 is very strong too)?

Код:

tc=5s+0.5s, 1 thread, move-overhead=50, RTX 2070

J92-190 vs others

   # PLAYER             :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L
+  1 lc0.net.J92-330    :      26     14   314.5     586  53.7      99   88   453   45
   2 lc0.net.J92-270    :       2     13   295.0     586  50.3      64   60   470   56
-  3 lc0.net.J92-190    :       0   ----  1754.0    3516  49.9      57  399  2710  407
   4 lc0.net.J92-220    :      -1     13   292.0     586  49.8      58   70   444   72
   5 lc0.net.J92-180    :      -3     12   290.5     586  49.6      50   67   447   72
   6 lc0.net.J92-210    :      -3     13   290.5     586  49.6      93   64   453   69
   7 lc0.net.J92-100    :     -17     13   279.5     586  47.7     ---   58   443   85

https://pixeldrain.com/u/bt3ZhF5h

[свернуть]

Владимир001 · 11.11.2020, 21:38

Развернуть для просмотра

Match: T73 and 703810 vs Stockfish 12
Time control: 2s + 0.2s
Hardware: RTX 3080, 10900k
Openings: Balsa_v500.pgn + TCEC SuFi 10-19 = 1000 pos.
Stockfish: threads=1, hash=32
Lc0 26.3 cuda 703810: threads=2, minibatch-size = 256, MLH tcec-19
Lc0 26.3 cuda 730382: threads=2, minibatch-size = 204, MLH tcec-19
Bench (200ms, startpos.): 730382 = 107 knps, 708010 = 130 knps, SF-12 = 1 600 knps
TBs and adj.: 5-men syzygy, draw 5 moves 5cp move 30th, resign (two-sided) 5 moves 600cp
Comments: T73 less than 20 elo behind 703810. I'm not sure if T73 has MLH, but I included it in the config file.
Games: https://gofile.io/d/JyI9I2

Код:

# PLAYER          : RATING  ERROR  PLAYED   (%)   CFS   W     D    L   D(%)
1 lc0-263-703810  :   38.6   6.8    2000   55.38  100  513  1189  298  59.45
2 lc0-263-730382  :   20.4   6.7    2000   52.85  100  452  1210  338  60.50
3 stockfish-12    :    0.0   4.2    4000   45.89  ---  636  2399  965  59.98

Engine            Depth  MIDG   EARLY    ENDG    LATE
lc0-263-703810    10.29  10.85 | 11.45 | 9.72 | 8.90
lc0-263-730382    10.03  10.53 | 11.27 | 9.50 | 8.74
stockfish-12      21.15  18.18 | 18.14 | 21.88 | 32.63

[свернуть]

Владимир001 · 11.11.2020, 21:40

Развернуть для просмотра

Match: Gauntlet of CorChess NNUE 4CPU against Lc0 J92-330 + J92-280
Lc0-version: v0.26.3
Lc0-options : -Backend=cudnn -NNCacheSize=2000000 -MaxOutOfOrderEvalsFactor=2.4 -MaxCollisionEvents=256 -TwoFoldDraws=true
MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0
CorChess-version: Corchess NNUE 1.3 081120 x64 bmi2 (date: 08.11.2020), compiled by I. Ivec
CorChess-options: threads=4 hash=4096 MB, contempt=12, NNUE=true, net=nn-c3ca321c51c9.nnue (default)
Time control: 5min + 5s
Hardware: notebook with mobile i7-7700HQ 4x2.8GHz (8 threads with HT on), mobile Geforce GTX 1050 TI 4GB
Book: Balsa_Special.pgn, 10 plies, sequentially, colors reversed
Tablebases: Syzygy 5-men on SSD
Speed: around 20k nodes per move for J92 // J92: 961 nps, CorChess NNUE 4CPU: 4112k nps
Adjudication: Draw at 50 moves rule, 3-fold repetition, after 200 moves, at 50 moves with eval = 0.00 ; 5-men Gaviota tablebase adjudication
Software: Arena 3.5.1, ponder = off
Elo rating: Ordo 1.2.6
Comments: Match of the best SF clone (so far better than Cfish and SF Dev in my tests) against the 2 best Lc0 nets.
Against SF Dev 4CPU (S19 SuFi version) at the same TC J92-280 was slightly stronger than J92-330.

Код:

   # PLAYER                        :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 Corchess NNUE 1.3_081120_4CPU :     0.0   ----   116.0     200  58.0     100   33  166    1  83.0
   2 Lc0.net.JH.J92-280            :   -46.1   25.0    43.5     100  43.5      87    1   85   14  85.0
   3 Lc0.net.JH.J92-330            :   -67.8   28.9    40.5     100  40.5     ---    0   81   19  81.0

https://pixeldrain.com/u/vQCbAv9o

[свернуть]

Владимир001 · 11.11.2020, 21:45

Развернуть для просмотра

Match: 30b nets J92-190 vs J92-330 in fixed nodes (1K, 10K)
LC0 version:0.26.3
LC0 options: 1 thread, move-overhead=50
Time control: 5s+0.5s
Hardware: RTX 2070
Book: Hert book (500 openings)
Tablebases: 6-man TB for both.
Adjudication: 6-man TB, -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000
Software:cutechess-cli
Comments: This test aims to check if J92-330 "scales" better than J92-190 (in 1K and 10K nodes/move). J92-330 beats J92-190 at about +7 elo (1K/move), +6 elo (10K/move). While J92-330 wins, it does not prove J92-330 scales better from 1Kn/move to 10Kn/move.

Код:

1k / move

   # PLAYER                :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L
   1 lc0.net.J92-330.1k    :       7     10   509.5    1000  51.0      91  132  755  113
   2 lc0.net.J92-190.1k    :       0   ----   490.5    1000  49.0     ---  113  755  132

White advantage = 35.53 +/- 5.61
Draw rate (equal opponents) = 77.24 % +/- 1.38


10 k / move

   # PLAYER                 :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L
   1 lc0.net.J92-330.10k    :       6      7   508.5    1000  50.9      95   69  879   52
   2 lc0.net.J92-190.10k    :       0   ----   491.5    1000  49.1     ---   52  879   69

White advantage = 27.05 +/- 3.82
Draw rate (equal opponents) = 90.50 % +/- 1.06

https://pixeldrain.com/u/Cktpciqu

[свернуть]

Владимир001 · 12.11.2020, 23:14

Матч: ERET test (Eigenmann Rapid Engine Test) с 111 тестовыми позициями
контроль времени: анализ с ограничением времени 60 секунд на ход/позицию

Развернуть для просмотра

Match: ERET test (Eigenmann Rapid Engine Test) with 111 test positions
time control: analysis with time limit of 60s per move/position
LC0-version: v0.21.2 up to v0.26.3
LC0-options: -Backend=cudnn -NNcache=3000000 -MaxCcollisionEvents=256
MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0
AB engine options: threads=4, hash=4096 MB
hardware: mobile Geforce GTX 1050 TI 4GB, mobile i7-7700HQ 4x2.8GHz (8 threads with HT)
software: Arena GUI, "Automatic Analysis" feature with EPD file.
Tablebases: Syzygy 5-men on SSD
Comment: EPD file: https://glarean-magazin.ch/wp-conten...gine-Chess.zip
Results (solutions, percent):

Код:

A) CPU engines (mobile i7-7700HQ 4x2.8GHz):
CorChess NNUE 1.3_081120 4CPU   :  97 = 87.4%   <== NEW, *net: nn-c3ca321c51c9.nnue*
Cfish 081120 NNUE 4CPU          :  96 = 86.5%   <== NEW   *net: nn-c3ca321c51c9.nnue*
Stockfish 20111018 NNUE 4CPU    :  95 = 85.6%   <== NEW, *net: nn-c3ca321c51c9.nnue*
Stockfish 20092822 NNUE 4CPU    :  95 = 85.6%   ** TCEC S19 Champion ** (net: nn-baeb9ef2d183.nnue)
Stockfish 12 4CPU (NNUE)        :  94 = 84.7%   *net: nn-82215d0fd0df.nnue*
BrainFish_200724_4CPU           :  92 = 82.9%
Stockfish 20073110 4CPU         :  92 = 82.9%   *last version before NNUE support was added*
LizardFish0.1+Cfish 081120_4CPU :  91 = 82.9%   <== NEW, *LizardFish NNUE net by @dkappe based on Komodo 14 evals*
Stockfish 11 4CPU               :  91 = 82.9%
Stockfish 10 4CPU               :  90 = 81.1%
Komodo-13.2.5_4CPU              :  88 = 79.3%
Komodo 12.1.1 4CPU (freeware)   :  84 = 75.7%   <== NEW
SlowChess Blitz Cl. v2.4 4CPU   :  83 = 74.8%   <== NEW
Nemorino 6.00 NNUE 4CPU         :  77 = 69.4%   *net: nemorino_net_6.nnue*
SlowChess Blitz Cl. v2.2 4CPU   :  77 = 69.4%
Xiphos 0.6 4CPU                 :  76 = 68.4%
Ethereal 12.75 pext-avx2 4CPU   :  73 = 65.8%
Igel 2.8.0 NNUE 4CPU            :  72 = 64.9%   *default net (night nurse v0.2.5b)*
Laser 1.7 4CPU                  :  72 = 64.9%
Komodo-MCTS-13.2.5_4CPU         :  62 = 55.9%
Lc0.net.703350-blas_4CPU        :  60 = 54.1%
Lc0.net.591003-blas_kiudee_4CPU :  60 = 54.1%
Fire 7.1 4CPU                   :  60 = 54.1%
Lc0.net.11258-112x9se_blas_4CPU :  59 = 53.2%
Lc0.net.DK.Bad_Gyal_7_blas_4CPU :  57 = 51.4%
Lc0.net.LD2_blas_4CPU  (10x128) :  56 = 50.4%

Код:

B) GPU engines (mobile Geforce GTX 1050 TI 4GB):
Lc0-0.22.0-dev.Leelenstein 14.3 :  83 = 74.8%
Lc0.net.JH.J92-270  (30x384)    :  81 = 73.0%   ** TCEC Cup 7 net **
Lc0.net.JH.J92-260  (30x384)    :  81 = 73.0%
Lc0.net.JH.J92-280  (30x384)    :  80 = 72.1%
Lc0.net.66000 (24x320)          :  80 = 72.1%   <== NEW
Lc0.net.65500 (24x320)          :  80 = 72.1%
Lc0.net.JH.J92-330  (30x384)    :  79 = 71.2%   <== NEW
Lc0.net.JH.J92-190  (30x384)    :  79 = 71.2%   ** TCEC S19 SuFi net **
Lc0.net.65701 (24x320)          :  79 = 71.2%
Lc0.net.JH.J92-320  (30x384)    :  78 = 70.3%   <== NEW
Lc0.net.JH.J92-310  (30x384)    :  77 = 69.4%
Lc0.net.JH.J92-300  (30x384)    :  77 = 69.4%
Allie-v0.7.Leelenstein 14.3     :  77 = 69.4%
Lc0.net.JH.J92-290  (30x384)    :  76 = 68.4%
Lc0.net.65411                   :  76 = 68.4%
Lc0.net.42810                   :  76 = 68.4%
Lc0.net.JH.J92-100  (30x384)    :  75 = 67.6%   ** TCEC S19 DivP net **
Lc0.net.SV.384x30-T60-4619      :  75 = 67.6%
Lc0.net.SV.384x30-T60-4229      :  75 = 67.6%   ** TCEC Cup 6 net **
Lc0.net.JH.T40B.4-160           :  75 = 67.6%   ** S16 DivP net **
Lc0.net.JH.T40B.4-260           :  75 = 67.6%
Lc0.net.703810  (10x128)        :  75 = 67.6%
Lc0.net.JH.J13B.2-136  (24x320) :  74 = 66.7%   *Terminator 2 net*
Lc0.net.SV.256x20-T40-1541      :  74 = 66.7%
Lc0.net.591003_kiudee  (10x128) :  74 = 66.7%
Lc0.net.42850                   :  73 = 65.8%
Lc0.net.11248                   :  69 = 62.2%
Lc0.net.LD2          (10x128)   :  67 = 60.4%
Lc0.net.32930                   :  66 = 59.4%
Lc0.net.DK.Bad_Gyal_9  (10x128) :  65 = 58.6%

[свернуть]

Владимир001 · 12.11.2020, 23:24

Матч
LC0-версия: lc0-v0.26.3 cudnn-версия

Развернуть для просмотра

Match Gauntlet
LC0-version: lc0-v0.26.3 cudnn-version
LC0 options: Standard option except Threads=1, NNCacheSize=0, MinibatchSize=1.
Time control: 1 Node/move
Hardware: i7-4790 with 16 GByte RAM + GTX1650
Speed: Not important for this type of testing.
Book: Chad-opening book with 100 000 different openings with the most popular human openings.
Book options: Opening order: Sequential. Play each opening twice. Play all opening in the book both for White and Black.
Tablebases: No Tablebases.
Adjudication: Only 5-men syzygy
Software: CuteChess 1.2.0
https://github.com/cutechess/cutechess/releases
Graphic tools is programmed by glbchess64, in Python programming tools.
Comment: This is result out of Ordo 1.2.6 https://github.com/michiguel/Ordo/releases/tag/v1.2.6
Ordo string: ordo.exe -Q -N 0 -D -a 0 -A "J92-145" -W -n4 -s500 -G -V -U "0,1,2,3,4,5,6,7,8,9,10" -j stats.txt -o rating.txt -p games.pgn
Results: pgn-file: https://ufile.io/by9zvwfg
Explanation: Trying to find the best net in clean policy Elo.

Код:

   # PLAYER     :  RATING  ERROR    POINTS  PLAYED   (%)  CFS(%)      W      D      L  D(%)
   1 J92-330    :      14      1  104076.0  200000  52.0     100  67090  73972  58938  37.0
   2 J92-145    :       0      1   95924.0  200000  48.0     ---  58938  73972  67090  37.0

White advantage = 29.00 +/- 0.56
Draw rate (equal opponents) = 37.24 % +/- 0.10

My list until now (Elo against J92-145) 1 Node/move:

Код:

White advantage = 23.98 +/- 0.57
Draw rate (equal opponents) = 36.51 % +/- 0.10
J92-310:             19     Elo    (Test with 200k games)
J92-320:             17     Elo    (Test with 200k games)
J92-270:             17     Elo    (Test with 200k games)
J92-330:             14     Elo    (Test with 200k games)
J92-280:             11     Elo    (Test with 200k games)
J92-260:             11     Elo    (Test with 200k games)
J92-290:             10     Elo    (Test with 200k games)
J92-240:              8     Elo    (Test with 200k games)
J92-210:              5     Elo    (Test with 200k games)
J92-220:              3     Elo    (Test with 200k games)
J92-205:              3     Elo    (Test with 200k games)
J92-230:              2     Elo    (Test with 200k games)
J92-190:              1     Elo    (Test with 200k games)
J92-145:              0     Elo    (Reference)
J92-100:             -2     Elo    (Test with 200k games)

[свернуть]

Владимир001 · 13.11.2020, 20:28

https://www.sp-cc.de/nn-vs-sf-testing.htm

Владимир001 · 13.11.2020, 20:32

SF201028 vs J92-280

Развернуть для просмотра

Match: Gauntlet SF201028 vs J92-280
Hardware: Geforce RTX 2070S & AMD Threadripper 3970X
Lc0 version: v0.26.3
Lc0 options: Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 option.MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0
SF options: Threads=23 Hash=4096
Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks)
Time control: 5m+5s
Book: openings-dd-4ply-100.pgn, sequential, color reversed
Tablebases: 6-man syzygy
Adjudication: 6-man TB, -draw movenumber=35 movecount=5 score=15 -resign movecount=5 score=1000
Software: cutechess-cli
Comments: J92-280 beat J92-270 in a very short TC SPRT test that I posted earlier. This test indidates that STC tests aren't necessarily a good indicator of strength against SFdev, although the difference is within the error bar. Also it did get more wins that the other nets.

Код:

   # PLAYER              :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L
   1 stockfish_201028    :     0.0   ----   313.5     600  52.3      97   38  551   11
   2 lc0.net.J92-270     :   -12.3   13.3    96.5     200  48.3      71    3  187   10
   3 lc0.net.J92-190     :   -17.6   13.7    95.0     200  47.5      50    3  184   13
   4 lc0.net.J92-280     :   -17.6   13.3    95.0     200  47.5     ---    5  180   15

White advantage = 24.04 +/- 3.95
Draw rate (equal opponents) = 96.20 % +/- 1.20

https://pixeldrain.com/u/XF7v9jTx

[свернуть]

svoitsl · 13.11.2020, 23:11

Если дело пойдет так и дальше,то TCEC 20 будет довольно интересным

Хорошо бы ещё для простых любителей сеть 256х20 натренировать, а то последняя новогодняя только есть 1541

Переяслав · 14.11.2020, 00:13

Цитата: Сообщение от svoitsl

а то последняя новогодняя только есть 1541

LS 14.3 есть в свободном доступе для LCZero 0.24.1

По немного странныи армагедонным тестам Штефана Поля она неплохо котируется, на уровне с коммерческой 15-ой

Владимир001 · 14.11.2020, 02:07

ICCF Position Test

Развернуть для просмотра

LC0 version: 0.26.3
LC0 options (all): threads=1, minibatch-size=192, moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-quadratic-factor=0, moves-left-scaled-factor=1, move-overhead-ms=0
LC0 options (zz_tune): cpuct=1.35, cpuct-factor=3.468, cpuct-base=20253, fpu-value=0.228, policy-softmax-temp=1.497, root-has-own-cpuct-params=false
LC0 options (trre_projected_tune): cpuct=2.3, cpuct-at-root=4.3, fpu-value=0.9, policy-softmax-temp=2.2
Time control: 50s per move (~500knpm)
Net: 703810
Hardware: gtx970 + i5-4570
Software: position-tester
Comment: testing different tunes at high nodes with random high level iccf positions. should be more representative of real games than ERET which is mostly hand selected tactical problems. seems like a good idea to make the @zz4032 tuned values into the new lc0 defaults

Код:

   # PLAYER                         :      SCORE
   1 lc0-zz_tune                    :     116/200
   2 lc0-trre_projected_tune        :     114/200
   3 lc0-defaults                   :     112/200

[свернуть]

Омегазеро · 14.11.2020, 05:16

Цитата: Сообщение от Омегазеро

Опубликованы тесты скоростей Лилы на новых RTX-картах в сравнении со старыми 2080

lczero 0.26.3-rc2 benchmark --cuda-fp16 384x30-t60-4619.pb.gz --num-positions=1 (начальная позиция)

RTX 2080 Ti - 384x30 = 21278 nps
RTX 2080 Super - 384x30 = 14915 nps

RTX 3080 FYI - 384x30 = 23222 (+75%)

RTX 3090 - 320x24 = 47636 nps (+126%)
RTX 3090 - 384x30 = 30591 nps (+130%)

Опубликованы тесты rtx 3070 в сравнении с другими картами (сеть 384x30, режим cuda-fp16)

2070 -----> 11.5 knps
2070 Super -----> 13.2 knps
2080 Super -----> 15.5 knps
2080 Ti -----> 21.3 knps
3070 -----> 16.8 knps
3080 -----> 25.2 knps
3090 -----> 30.6 knps

При этом 3080 можно купить за 87 тысяч https://www.nix.ru/autocatalog/palit...80_484079.html
3070 можно купить за 54 тысячи https://pc4you.ru/catalog/komplektuy...eokarty/37898/
2080Ti можно купить за 79 тысяч https://becompact.ru/tovary/videokar...0-ti-aero-11g/

В пересчете на nps цена одинаковая

05.11.2020, 03:30	#7231
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Hardware: i5-8400 CPU @ 2.80GHz, 6 cores, 32GB, gpu RTX 2060, 6GB Gauntlet: SF-dev vs. LC0 24x320 nets Network IDs: T60B.7-105, T60B.7-135, 64255, 65162, 65297, 65373, 65374, 65726, 65762 LC0-version: v0.26.0 SF-version: SF-dev 20092708 (NNUE hybrid) Software: Cutechess-CLI Time control: LC0: 1,500 nodes/move, SF-dev: 300,000 nodes/move @ 4 Threads, 512MB Hash Book: Chad's 8-ply book with 10000 positions, order=sequential Tablebases: 6-men Adjudication: -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000 Ordo: -a 0 -A "Stockfish.dev" -W -D -U "0,1,2,3,4,5,6,7,8,9,10" -s 2000 -n 1 Comment: Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) + 1 lc0.net.65762 : 34.4 3.2 10952.0 20000 55 92 5494 10916 3590 55 + 2 lc0.net.65726 : 31.1 3.2 10862.0 20000 54 66 5317 11090 3593 55 3 lc0.net.65297 : 30.2 3.2 10836.0 20000 54 59 5374 10924 3702 55 4 lc0.net.65373 : 29.7 3.2 10822.0 20000 54 85 5360 10924 3716 55 5 lc0.net.65162 : 27.2 3.3 10754.0 20000 54 94 5255 10998 3747 55 6 lc0.net.65374 : 23.7 3.2 10657.5 20000 53 90 5127 11061 3812 55 7 lc0.net.T60B.7-135 : 20.8 3.1 10576.5 20000 53 63 4994 11165 3841 56 8 lc0.net.T60B.7-105 : 20.0 3.3 10556.0 20000 53 100 4984 11144 3872 56 9 lc0.net.64255 : 3.6 3.3 10100.0 20000 51 98 4463 11274 4263 56 - 10 Stockfish.dev : 0.0 ---- 83884.0 180000 47 --- 34136 99496 46368 55 White advantage = 57.39 +/- 0.54 Draw rate (equal opponents) = 57.36 % +/- 0.12 PGN: match-166 match-167 [свернуть]

05.11.2020, 11:21	#7232
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero вот забавная книга открытия гандикапа для матчей двигателя. Его всего 20 отверстий, они необычны, но забавны для просмотра и могут помочь увеличить масштаб elo для матчей процессоров. Каждое открытие содержит оценку SF13 D33 от первого выхода из книги (в тегах аннотаций). Каждое открытие неуклонно увеличивает eval для белых, колеблясь от +.70 до +2.65. handicaps20.rar

05.11.2020, 15:08	#7234
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero "CCWC CPU\GPU Candidates League' 1min+1secs 4cpu и 1gpu (38%мощности) со случайными отверстиями перевернуты, из новых 650 отверстий SuFi\2ply book compilation

06.11.2020, 00:29	#7236
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: LJ2 Lender net vs representative T60 net Lc0-version: 26.1 Lc0 options: MovesLeftMaxEffect=0.1 option.MovesLeftThreshold=0.0 option.MovesLeftSlope=0.007 option.MovesLeftQuadraticFactor=1.0 option.MovesLeftScaledFactor=0.0 option.MovesLeftConstantFactor=0.0 MovesLeftConstantFactor=0.0 MoveOverheadMs=0 Time/Node control: 800 fixed nodes Hardware: RTX 2080/RTX 2070, AMD 2950X Book: endgame-16-piece-book_Q-0.0-0.6.pgn, random with reverse. Tablebases: 6-man Syzygy Adjudication: 6-man Syzygy and -draw movenumber=50 movecount=6 score=10 -resign movecount=5 score=500 twosided=true Chess Interface: cutechess-cli Comments: Ongoing tests of LJ2 (320x24b) endgame training progress. Results: Код: # PLAYER : RATING ERROR POINTS PLAYED (%) W L D D(%) CFS(%) 1 lc0.net.65560 : 0.0 ---- 2185.0 4000 55 780 410 2810 70 100 2 lc0.net.LJ2-140 : -21.2 8.0 939.5 2000 47 214 335 1451 73 100 3 lc0.net.LJ2-80 : -43.9 8.7 875.5 2000 44 196 445 1359 68 --- https://pixeldrain.com/u/PtoavbFM [свернуть]

06.11.2020, 11:18	#7237
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match Gauntlet LC0-version: lc0-v0.26.3 cudnn-version LC0 options: Standard option except Threads=1, NNCacheSize=0, MinibatchSize=1. Time control: 1 Node/move Hardware: i7-4790 with 16 GByte RAM + GTX1650 Speed: Not important for this type of testing. Book: Chad-opening book with 100 000 different openings with the most popular human openings. Book options: Opening order: Sequential. Play each opening twice. Play all opening in the book both for White and Black. Tablebases: No Tablebases. Adjudication: Only 5-men syzygy Software: CuteChess 1.2.0 https://github.com/cutechess/cutechess/releases / Graphic tools is programmed by glbchess64, in Python programming tools. Comment: This is result out of Ordo 1.2.6 https://github.com/michiguel/Ordo/releases/tag/v1.2.6 Ordo string: ordo.exe -Q -N 0 -D -a 0 -A "J92-145" -W -n4 -s500 -G -V -U "0,1,2,3,4,5,6,7,8,9,10" -j stats.txt -o rating.txt -p games.pgn Results: pgn-file: https://ufile.io/0170kecb Explanation: Trying to find the best net in clean policy Elo. Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) 1 J92-320 : 17 1 104871.0 200000 52.4 100 67558 74626 57816 37.3 2 J92-145 : 0 1 95129.0 200000 47.6 --- 57816 74626 67558 37.3 White advantage = 24.54 +/- 0.56 Draw rate (equal opponents) = 37.53 % +/- 0.10 Код: My list until now (Elo against J92-145) 1 Node/move: White advantage = 23.98 +/- 0.57 Draw rate (equal opponents) = 36.51 % +/- 0.10 J92-310: 19 Elo (Test with 200k games) J92-320: 17 Elo (Test with 200k games) J92-270: 17 Elo (Test with 200k games) J92-280: 11 Elo (Test with 200k games) J92-260: 11 Elo (Test with 200k games) J92-290: 10 Elo (Test with 200k games) J92-240: 8 Elo (Test with 200k games) J92-210: 5 Elo (Test with 200k games) J92-220: 3 Elo (Test with 200k games) J92-205: 3 Elo (Test with 200k games) J92-230: 2 Elo (Test with 200k games) J92-190: 1 Elo (Test with 200k games) J92-145: 0 Elo (Reference) J92-100: -2 Elo (Test with 200k games) [свернуть]

07.11.2020, 13:57	#7241
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: Match of of SF Dev 20092822 4CPU vs Lc0 J92-330 Lc0-version: v0.26.2 + v0.26.3 (since J92-260) Lc0-options : -Backend=cudnn -NNCacheSize=2000000 -MaxOutOfOrderEvalsFactor=2.4 -MaxCollisionEvents=256 -TwoFoldDraws=true MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0 SF-version: Stockfish 20092822 x64 bmi2 (date: 28.09.2020) SF/Cfish-options: threads=4 hash=2048 MB, contempt=24, NNUE=true, net=nn-baeb9ef2d183.nnue (default) Time control: 5min + 5s Hardware: notebook with mobile i7-7700HQ 4x2.8GHz (8 threads with HT on), mobile Geforce GTX 1050 TI 4GB Book: Balsa_Special.pgn, 10 plies, sequentially, colors reversed Tablebases: Syzygy 5-men on SSD Speed: J92: 961 nps, SF Dev NNUE 4CPU: 3680k nps, Cfish NNUE 4CPU: 4198k nps Adjudication: Draw at 50 moves rule, 3-fold repetition, after 200 moves, at 50 moves with eval = 0.00 ; 5-men Gaviota tablebase adjudication Software: Arena 3.5.1, ponder = off Elo rating: Ordo 1.2.6 Comments: J92-330 added. ERET test results are included in the results table (60s time limit per position). Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) ERET test 1 Cfish-NNUE 071020_4CPU : 7.1 22.6 51.0 100 51.0 73 2 98 0 98.0 96/111 2 Stockfish_20092822_4CPU : 0.0 ---- 637.0 1100 57.9 100 190 894 16 81.3 95/111 3 Lc0.net.JH.J92-280 : -46.2 25.9 43.5 100 43.5 58 3 81 16 81.0 80/111 4 Lc0.net.JH.J92-330 : -49.8 26.3 43.0 100 43.0 50 4 78 18 78.0 79/111 <= NEW 5 Lc0.net.JH.J92-260 : -49.8 26.9 43.0 100 43.0 58 1 84 15 84.0 81/111 6 Lc0.net.JH.J92-210 : -53.4 26.0 42.5 100 42.5 50 1 83 16 83.0 81/111 7 Lc0.net.JH.J92-270 : -53.4 26.0 42.5 100 42.5 77 1 83 16 83.0 81/111 8 Lc0.net.JH.J92-220 : -67.9 28.7 40.5 100 40.5 57 2 77 21 77.0 79/111 9 Lc0.net.JH.J92-190 : -71.6 28.3 40.0 100 40.0 50 1 78 21 78.0 79/111 10 Lc0.net.JH.J92-310 : -71.6 29.3 40.0 100 40.0 57 1 78 21 78.0 77/111 11 Lc0.net.JH.J92-240 : -75.3 30.0 39.5 100 39.5 75 0 79 21 79.0 80/111 12 Stockfish_11_4CPU : -90.2 31.3 37.5 100 37.5 --- 0 75 25 75.0 91/111 https://pixeldrain.com/u/9fbKd2at [свернуть]

07.11.2020, 14:02	#7242
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Hardware: i7-8770K CPU @ 3.70GHz, 6 cores (12 Threads), 32GB, Gpu - RTX 2080 \| Lc0 vs SF Dev Scaling Continuation Gauntlet: SF-dev vs. LC0 30x384 net Network IDs: Lc0 J92-270 LC0-version: v0.26.3 SF-version: SF-dev 5th Oct (NNUE hybrid) Software: Cutechess-CLI Time control: Lc0 80k vs SF NNUE 28000k per move. \| Lc0 with default settings and 1 thread, SF NNUE with default settings and 4 threads. Book: 8moves_v3.pgn (SF Regression book at LTC) Tablebases: 6-men Synergy WDL with TB Adjudication Adjudication: -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000 Ordo: ordo-win64.exe -Q -N "1,1" -D -a 0 -A "Stockfish_Dev4T" -W -n4 -s5000 -p games.pgn -U "0,1,2,3,4,5,6,7,8,9,10" -o rating.txt Comment: This is the continuation of the my scaling test at 80000 nodes of Lc0 Jh 270 vs 28000k nodes SF NNUE (5th Oct). Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) 1 Stockfish_Dev_4T : 0.0 ---- 1011.0 2000 50.5 93 130 1762 108 88.1 2 Lc0 Jh 270 : -3.9 5.2 989.0 2000 49.5 --- 108 1762 130 88.1 White advantage = 24.58 +/- 2.72 Draw rate (equal opponents) = 90.19 % +/- 0.74 https://pixeldrain.com/l/scuMCCBF#item=0 2 файла [свернуть]

10.11.2020, 00:53	#7244
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: Gauntlet SF201028 vs J92-270 & J92-190 Hardware: Geforce RTX 2070S & AMD Threadripper 3970X Lc0 version: v0.26.3 Lc0 options: Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 option.MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0 SF options: Threads=23 Hash=4096 Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks) Time control: 5m+5s Book: openings-dd-4ply-100.pgn, sequential, color reversed Tablebases: 6-man syzygy Adjudication: 6-man TB, -draw movenumber=35 movecount=5 score=15 -resign movecount=5 score=1000 Software: cutechess-cli Comments: J92-190 was used to establish a baseline with my new CPU. I have limited the number of threads used by stockfish to match it to the RTX2070S with the ratio set to 875 with lc0 running a 256x20 net. The time control was chosen so that lc0 got more than 100,000 nodes per move. J92-270 was the first comparison test after performing an SPRT test of J92-270 vs J92-190 (see next post). Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L 1 stockfish_201028 : 0.0 ---- 208.5 400 52.1 97 23 371 6 2 lc0.net.J92-270 : -12.3 12.4 96.5 200 48.3 72 3 187 10 3 lc0.net.J92-190 : -17.6 12.8 95.0 200 47.5 --- 3 184 13 White advantage = 20.21 +/- 4.64 Draw rate (equal opponents) = 96.40 % +/- 1.3 2 файла : https://pixeldrain.com/l/qRe4YPW6#item=0 [свернуть]

10.11.2020, 01:01	#7245
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: SPRT test of J92-270 vs J92-190 Hardware: Geforce RTX 2070S & AMD Threadripper 3970X Lc0 version: v0.26.3 Lc0 options: Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 option.MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0 Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks) Time control: 1.6s+0.03s Book: book-ply6-unifen-Q-0.0-0.25.pgn, random, color reversed Tablebases: 6-man syzygy Adjudication: 6-man TB, -draw movenumber=20 movecount=5 score=10 -resign movecount=3 score=500 Software: cutechess-cli Comments: My new process is to use an SPRT test to determine which nets to then test against SFdev at 5m+5s. Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L 1 lc0.net.J92-270 : 5.3 3.9 4912.5 9678 50.8 100 1556 6713 1409 2 lc0.net.J92-190 : 0.0 ---- 4765.5 9678 49.2 --- 1409 6713 1556 White advantage = 20.49 +/- 1.99 Draw rate (equal opponents) = 69.78 % +/- 0.46 Код: python sprt.py --elo0=0 --elo1=3 --results 1409 6713 1556: [Elo0,Elo1] : [0.00,3.00] Confidence level : 95.00% Elo : 4.78 Confidence interval : [0.77,8.72] (95.00%) LOS : 99.00% LLR [u,l] : 2.94 [-2.94,2.94 https://pixeldrain.com/u/5f8FW94C [свернуть]

10.11.2020, 17:14	#7246
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: SPRT test of J92-280 vs J92-270 Hardware: Geforce RTX 2070S & AMD Threadripper 3970X Lc0 version: v0.26.3 Lc0 options: Threads=1 Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0 Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks) Time control: 1.6s+0.03s Book: book-ply6-unifen-Q-0.0-0.25.pgn, random, color reversed Tablebases: 6-man syzygy Adjudication: 6-man TB, -draw movenumber=20 movecount=5 score=10 -resign movecount=3 score=500 Software: cutechess-cli Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L 1 lc0.net.J92-280 : 5.1 3.8 5148.0 10149 50.7 100 1608 7080 1461 2 lc0.net.J92-270 : 0.0 ---- 5001.0 10149 49.3 --- 1461 7080 1608 White advantage = 26.89 +/- 1.96 Draw rate (equal opponents) = 70.48 % +/- 0.46 python sprt.py --elo0=0 --elo1=3 --results 1461 7080 1608: Код: [Elo0,Elo1] : [0.00,3.00] Confidence level : 95.00% Elo : 4.53 Confidence interval : [0.64,8.35] (95.00%) LOS : 98.85% LLR [u,l] : 2.94 [-2.94,2.94 https://pixeldrain.com/u/PC39bmVM [свернуть]

10.11.2020, 20:01	#7247
svoitsl Гроссмейстер Регистрация: 19.04.2017 Сообщений: 4,734 Сказал(а) спасибо: 1,204 Поблагодарили 2,853 раз(а) в 1,885 сообщениях Репутация: 159	Re: LCZero 3000 самоЭло!!!!

11.11.2020, 00:41	#7249
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: 30b nets J92-190 vs J92-330 in time control LC0 version:0.26.3 LC0 options: 1 thread, move-overhead=50 Time control: 5s+0.5s Hardware: RTX 2070 Book: Sufi book (293 openings) Tablebases: 6-man TB for both. Speed: 9.1 knps (benchmark with 1 thread) Adjudication: 6-man TB, -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=800 Software:cutechess-cli Comments: J92-330 beats J92-190 quite hard (+88 -45 =453, +26 elo). I don't know why J92-330 overperforms here. In low node test (about 1 knpm), J92-330 is good, but is not as strong as in this test (about 10 knpm). Does it scale much better in high nodes (TCEC bonus 3 result shows J92-330 is very strong too)? Код: tc=5s+0.5s, 1 thread, move-overhead=50, RTX 2070 J92-190 vs others # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L + 1 lc0.net.J92-330 : 26 14 314.5 586 53.7 99 88 453 45 2 lc0.net.J92-270 : 2 13 295.0 586 50.3 64 60 470 56 - 3 lc0.net.J92-190 : 0 ---- 1754.0 3516 49.9 57 399 2710 407 4 lc0.net.J92-220 : -1 13 292.0 586 49.8 58 70 444 72 5 lc0.net.J92-180 : -3 12 290.5 586 49.6 50 67 447 72 6 lc0.net.J92-210 : -3 13 290.5 586 49.6 93 64 453 69 7 lc0.net.J92-100 : -17 13 279.5 586 47.7 --- 58 443 85 https://pixeldrain.com/u/bt3ZhF5h [свернуть]

11.11.2020, 21:38	#7250
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: T73 and 703810 vs Stockfish 12 Time control: 2s + 0.2s Hardware: RTX 3080, 10900k Openings: Balsa_v500.pgn + TCEC SuFi 10-19 = 1000 pos. Stockfish: threads=1, hash=32 Lc0 26.3 cuda 703810: threads=2, minibatch-size = 256, MLH tcec-19 Lc0 26.3 cuda 730382: threads=2, minibatch-size = 204, MLH tcec-19 Bench (200ms, startpos.): 730382 = 107 knps, 708010 = 130 knps, SF-12 = 1 600 knps TBs and adj.: 5-men syzygy, draw 5 moves 5cp move 30th, resign (two-sided) 5 moves 600cp Comments: T73 less than 20 elo behind 703810. I'm not sure if T73 has MLH, but I included it in the config file. Games: https://gofile.io/d/JyI9I2 Код: # PLAYER : RATING ERROR PLAYED (%) CFS W D L D(%) 1 lc0-263-703810 : 38.6 6.8 2000 55.38 100 513 1189 298 59.45 2 lc0-263-730382 : 20.4 6.7 2000 52.85 100 452 1210 338 60.50 3 stockfish-12 : 0.0 4.2 4000 45.89 --- 636 2399 965 59.98 Engine Depth MIDG EARLY ENDG LATE lc0-263-703810 10.29 10.85 \| 11.45 \| 9.72 \| 8.90 lc0-263-730382 10.03 10.53 \| 11.27 \| 9.50 \| 8.74 stockfish-12 21.15 18.18 \| 18.14 \| 21.88 \| 32.63 [свернуть]

11.11.2020, 21:40	#7251
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: Gauntlet of CorChess NNUE 4CPU against Lc0 J92-330 + J92-280 Lc0-version: v0.26.3 Lc0-options : -Backend=cudnn -NNCacheSize=2000000 -MaxOutOfOrderEvalsFactor=2.4 -MaxCollisionEvents=256 -TwoFoldDraws=true MLH-settings S19 SuFi: -MovesLeftMaxEffect=0.2 -MovesLeftSlope=0.004 -MovesLeftThreshold=0.0 -MovesQuadraticFactor=0.0 -MovesScaledFactor=1.0 CorChess-version: Corchess NNUE 1.3 081120 x64 bmi2 (date: 08.11.2020), compiled by I. Ivec CorChess-options: threads=4 hash=4096 MB, contempt=12, NNUE=true, net=nn-c3ca321c51c9.nnue (default) Time control: 5min + 5s Hardware: notebook with mobile i7-7700HQ 4x2.8GHz (8 threads with HT on), mobile Geforce GTX 1050 TI 4GB Book: Balsa_Special.pgn, 10 plies, sequentially, colors reversed Tablebases: Syzygy 5-men on SSD Speed: around 20k nodes per move for J92 // J92: 961 nps, CorChess NNUE 4CPU: 4112k nps Adjudication: Draw at 50 moves rule, 3-fold repetition, after 200 moves, at 50 moves with eval = 0.00 ; 5-men Gaviota tablebase adjudication Software: Arena 3.5.1, ponder = off Elo rating: Ordo 1.2.6 Comments: Match of the best SF clone (so far better than Cfish and SF Dev in my tests) against the 2 best Lc0 nets. Against SF Dev 4CPU (S19 SuFi version) at the same TC J92-280 was slightly stronger than J92-330. Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) 1 Corchess NNUE 1.3_081120_4CPU : 0.0 ---- 116.0 200 58.0 100 33 166 1 83.0 2 Lc0.net.JH.J92-280 : -46.1 25.0 43.5 100 43.5 87 1 85 14 85.0 3 Lc0.net.JH.J92-330 : -67.8 28.9 40.5 100 40.5 --- 0 81 19 81.0 https://pixeldrain.com/u/vQCbAv9o [свернуть]

11.11.2020, 21:45	#7252
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Развернуть для просмотра Match: 30b nets J92-190 vs J92-330 in fixed nodes (1K, 10K) LC0 version:0.26.3 LC0 options: 1 thread, move-overhead=50 Time control: 5s+0.5s Hardware: RTX 2070 Book: Hert book (500 openings) Tablebases: 6-man TB for both. Adjudication: 6-man TB, -draw movenumber=50 movecount=5 score=8 -resign movecount=5 score=1000 Software:cutechess-cli Comments: This test aims to check if J92-330 "scales" better than J92-190 (in 1K and 10K nodes/move). J92-330 beats J92-190 at about +7 elo (1K/move), +6 elo (10K/move). While J92-330 wins, it does not prove J92-330 scales better from 1Kn/move to 10Kn/move. Код: 1k / move # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L 1 lc0.net.J92-330.1k : 7 10 509.5 1000 51.0 91 132 755 113 2 lc0.net.J92-190.1k : 0 ---- 490.5 1000 49.0 --- 113 755 132 White advantage = 35.53 +/- 5.61 Draw rate (equal opponents) = 77.24 % +/- 1.38 10 k / move # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L 1 lc0.net.J92-330.10k : 6 7 508.5 1000 50.9 95 69 879 52 2 lc0.net.J92-190.10k : 0 ---- 491.5 1000 49.1 --- 52 879 69 White advantage = 27.05 +/- 3.82 Draw rate (equal opponents) = 90.50 % +/- 1.06 https://pixeldrain.com/u/Cktpciqu [свернуть]

12.11.2020, 23:24	#7254
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero Матч LC0-версия: lc0-v0.26.3 cudnn-версия Развернуть для просмотра Match Gauntlet LC0-version: lc0-v0.26.3 cudnn-version LC0 options: Standard option except Threads=1, NNCacheSize=0, MinibatchSize=1. Time control: 1 Node/move Hardware: i7-4790 with 16 GByte RAM + GTX1650 Speed: Not important for this type of testing. Book: Chad-opening book with 100 000 different openings with the most popular human openings. Book options: Opening order: Sequential. Play each opening twice. Play all opening in the book both for White and Black. Tablebases: No Tablebases. Adjudication: Only 5-men syzygy Software: CuteChess 1.2.0 https://github.com/cutechess/cutechess/releases Graphic tools is programmed by glbchess64, in Python programming tools. Comment: This is result out of Ordo 1.2.6 https://github.com/michiguel/Ordo/releases/tag/v1.2.6 Ordo string: ordo.exe -Q -N 0 -D -a 0 -A "J92-145" -W -n4 -s500 -G -V -U "0,1,2,3,4,5,6,7,8,9,10" -j stats.txt -o rating.txt -p games.pgn Results: pgn-file: https://ufile.io/by9zvwfg Explanation: Trying to find the best net in clean policy Elo. Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) 1 J92-330 : 14 1 104076.0 200000 52.0 100 67090 73972 58938 37.0 2 J92-145 : 0 1 95924.0 200000 48.0 --- 58938 73972 67090 37.0 White advantage = 29.00 +/- 0.56 Draw rate (equal opponents) = 37.24 % +/- 0.10 My list until now (Elo against J92-145) 1 Node/move: Код: White advantage = 23.98 +/- 0.57 Draw rate (equal opponents) = 36.51 % +/- 0.10 J92-310: 19 Elo (Test with 200k games) J92-320: 17 Elo (Test with 200k games) J92-270: 17 Elo (Test with 200k games) J92-330: 14 Elo (Test with 200k games) J92-280: 11 Elo (Test with 200k games) J92-260: 11 Elo (Test with 200k games) J92-290: 10 Elo (Test with 200k games) J92-240: 8 Elo (Test with 200k games) J92-210: 5 Elo (Test with 200k games) J92-220: 3 Elo (Test with 200k games) J92-205: 3 Elo (Test with 200k games) J92-230: 2 Elo (Test with 200k games) J92-190: 1 Elo (Test with 200k games) J92-145: 0 Elo (Reference) J92-100: -2 Elo (Test with 200k games) [свернуть]

13.11.2020, 20:28	#7255
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero https://www.sp-cc.de/nn-vs-sf-testing.htm

13.11.2020, 20:32	#7256
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero SF201028 vs J92-280 Развернуть для просмотра Match: Gauntlet SF201028 vs J92-280 Hardware: Geforce RTX 2070S & AMD Threadripper 3970X Lc0 version: v0.26.3 Lc0 options: Backend=cuda-fp16 NNCacheSize=4000000 MinibatchSize=80 option.MaxOutOfOrderEvalsFactor=2.4 MovesLeftMaxEffect=0.2 MovesLeftThreshold=0.0 MovesLeftSlope=0.004 MovesLeftQuadraticFactor=0.0 MovesLeftScaledFactor=1.0 MovesLeftConstantFactor=0.0 SF options: Threads=23 Hash=4096 Speed: Lc0:12.7 kN/s, SF201029: 34,266 kN/s (from benchmarks) Time control: 5m+5s Book: openings-dd-4ply-100.pgn, sequential, color reversed Tablebases: 6-man syzygy Adjudication: 6-man TB, -draw movenumber=35 movecount=5 score=15 -resign movecount=5 score=1000 Software: cutechess-cli Comments: J92-280 beat J92-270 in a very short TC SPRT test that I posted earlier. This test indidates that STC tests aren't necessarily a good indicator of strength against SFdev, although the difference is within the error bar. Also it did get more wins that the other nets. Код: # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L 1 stockfish_201028 : 0.0 ---- 313.5 600 52.3 97 38 551 11 2 lc0.net.J92-270 : -12.3 13.3 96.5 200 48.3 71 3 187 10 3 lc0.net.J92-190 : -17.6 13.7 95.0 200 47.5 50 3 184 13 4 lc0.net.J92-280 : -17.6 13.3 95.0 200 47.5 --- 5 180 15 White advantage = 24.04 +/- 3.95 Draw rate (equal opponents) = 96.20 % +/- 1.20 https://pixeldrain.com/u/XF7v9jTx [свернуть]

13.11.2020, 23:11	#7257
svoitsl Гроссмейстер Регистрация: 19.04.2017 Сообщений: 4,734 Сказал(а) спасибо: 1,204 Поблагодарили 2,853 раз(а) в 1,885 сообщениях Репутация: 159	Re: LCZero Если дело пойдет так и дальше,то TCEC 20 будет довольно интересным Хорошо бы ещё для простых любителей сеть 256х20 натренировать, а то последняя новогодняя только есть 1541

14.11.2020, 00:13	#7258
Переяслав Гроссмейстер Регистрация: 12.07.2020 Сообщений: 1,784 Сказал(а) спасибо: 650 Поблагодарили 2,306 раз(а) в 1,004 сообщениях Записей в дневнике: 20 Репутация: 136	Re: LCZero Цитата: Сообщение от svoitsl а то последняя новогодняя только есть 1541 LS 14.3 есть в свободном доступе для LCZero 0.24.1 По немного странныи армагедонным тестам Штефана Поля она неплохо котируется, на уровне с коммерческой 15-ой __________________ PGN Spy - top tournaments

14.11.2020, 02:07	#7259
Владимир001 Эксперт Регистрация: 19.12.2019 Сообщений: 7,484 Сказал(а) спасибо: 4,655 Поблагодарили 12,178 раз(а) в 5,475 сообщениях Репутация: 259	Re: LCZero ICCF Position Test Развернуть для просмотра LC0 version: 0.26.3 LC0 options (all): threads=1, minibatch-size=192, moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-quadratic-factor=0, moves-left-scaled-factor=1, move-overhead-ms=0 LC0 options (zz_tune): cpuct=1.35, cpuct-factor=3.468, cpuct-base=20253, fpu-value=0.228, policy-softmax-temp=1.497, root-has-own-cpuct-params=false LC0 options (trre_projected_tune): cpuct=2.3, cpuct-at-root=4.3, fpu-value=0.9, policy-softmax-temp=2.2 Time control: 50s per move (~500knpm) Net: 703810 Hardware: gtx970 + i5-4570 Software: position-tester Comment: testing different tunes at high nodes with random high level iccf positions. should be more representative of real games than ERET which is mostly hand selected tactical problems. seems like a good idea to make the @zz4032 tuned values into the new lc0 defaults Код: # PLAYER : SCORE 1 lc0-zz_tune : 116/200 2 lc0-trre_projected_tune : 114/200 3 lc0-defaults : 112/200 [свернуть]

14.11.2020, 05:16	#7260
Омегазеро Гроссмейстер Регистрация: 20.01.2019 Сообщений: 1,067 Сказал(а) спасибо: 2 Поблагодарили 701 раз(а) в 343 сообщениях Репутация: 92	Re: LCZero Цитата: Сообщение от Омегазеро Опубликованы тесты скоростей Лилы на новых RTX-картах в сравнении со старыми 2080 lczero 0.26.3-rc2 benchmark --cuda-fp16 384x30-t60-4619.pb.gz --num-positions=1 (начальная позиция) RTX 2080 Ti - 384x30 = 21278 nps RTX 2080 Super - 384x30 = 14915 nps RTX 3080 FYI - 384x30 = 23222 (+75%) RTX 3090 - 320x24 = 47636 nps (+126%) RTX 3090 - 384x30 = 30591 nps (+130%) Опубликованы тесты rtx 3070 в сравнении с другими картами (сеть 384x30, режим cuda-fp16) 2070 -----> 11.5 knps 2070 Super -----> 13.2 knps 2080 Super -----> 15.5 knps 2080 Ti -----> 21.3 knps 3070 -----> 16.8 knps 3080 -----> 25.2 knps 3090 -----> 30.6 knps При этом 3080 можно купить за 87 тысяч https://www.nix.ru/autocatalog/palit...80_484079.html 3070 можно купить за 54 тысячи https://pc4you.ru/catalog/komplektuy...eokarty/37898/ 2080Ti можно купить за 79 тысяч https://becompact.ru/tovary/videokar...0-ti-aero-11g/ В пересчете на nps цена одинаковая Последний раз редактировалось Омегазеро; 14.11.2020 в 05:47.

Здесь присутствуют: 8 (пользователей: 0 , гостей: 8)