MBS halbiert bzw. geviertelt

By Lothar Jung Date 2024-03-09 08:50 Upvotes 1

**Comment:** Sf 16.1 vs t3 & BT4 with various minibatch-size settings
**Hardware:** R7 2700X (8c/16t) | RTX 3080 10GB (SM=68) | Syzygy-6
**Stockfish:** ver. 16.1, abrok avx2, 14 threads, 2GB hash, Sy50mr=off, MoveOv=0
**Lc0:** https://github.com/Ergodice/lc0 master 5c1051f (13Feb2024)
**Lc0 settings:** defaults except vanilla|winner, Minibatchsize=34 or 68 or 136, MoOvMs=0, StrictTiming
**Lc0 Nets:**   BT4-1024x15x32h-swa-6147500.pb.gz ; t3-512x15x16h-swa-2815000.pb.gz
**Openings:** UHO_Lichess_4852_v1.epd seq. rev. same for all Lc0's
**T/C's   Stockfish:** Game 6s + .1s **Lc0:** Game 15s + .25s
**Cute Chess GUI:** draw 35. 3 .20, resign 2 5.00 twosided, timemargin=200, policy=round
**Gauntlet times:** B 14h 55m, t 14h 30m, total 1d 5h 25m```Lc0 :mbs RATING ERROR POINTS PLAYED (%) CFS(%)   W D L D(%)
1 Lc0-BT4:34    :   5.9   24.6   203.0 400 50.8 68   112 182 106 45.5
2 Stockfish-16.1 :   0.0   ----   617.0   1200 51.4 76   339 556 305 46.3
3 Lc0-BT4:68    : -8.9   25.2   195.5 400 48.9 89   106 179 115 44.8
4 Lc0-BT4:136 : -30.7   24.5   184.5 400 46.1   --- 87 195 118 48.8

Lc0 :mbs RATING ERROR POINTS PLAYED (%) CFS(%)   W D L D(%)
1 Lc0-t3:68 : 14.8   25.3   207.5 400 51.9 76   108 199   93 49.8
2 Lc0-t3:34 :   2.0   24.6   201.0 400 50.3 56   113 176 111 44.0
3 Stockfish-16.1 :   0.0   ----   597.5   1200 49.8 83   320 555 325 46.3
4 Lc0-t3:136    : -11.8   24.4   194.0 400 48.5   ---   104 180 116 45.0```

By Jörg Oster Date 2024-03-09 09:58

Lothar Jung schrieb:

**Comment:** Sf 16.1 vs t3 & BT4 with various minibatch-size settings
**Hardware:** R7 2700X (8c/16t) | RTX 3080 10GB (SM=68) | Syzygy-6
**Stockfish:** ver. 16.1, abrok avx2, 14 threads, 2GB hash, Sy50mr=off, MoveOv=0
**Lc0:** <a class='ura' href='https://github.com/Ergodice/lc0'>https://github.com/Ergodice/lc0</a> master 5c1051f (13Feb2024)
**Lc0 settings:** defaults except vanilla|winner, Minibatchsize=34 or 68 or 136, MoOvMs=0, StrictTiming
**Lc0 Nets:**   BT4-1024x15x32h-swa-6147500.pb.gz ; t3-512x15x16h-swa-2815000.pb.gz
**Openings:** UHO_Lichess_4852_v1.epd seq. rev. same for all Lc0's
**T/C's   Stockfish:** Game 6s + .1s **Lc0:** Game 15s + .25s
**Cute Chess GUI:** draw 35. 3 .20, resign 2 5.00 twosided, timemargin=200, policy=round
**Gauntlet times:** B 14h 55m, t 14h 30m, total 1d 5h 25m```Lc0 :mbs RATING ERROR POINTS PLAYED (%) CFS(%)   W D L D(%)
1 Lc0-BT4:34    :   5.9   24.6   203.0 400 50.8 68   112 182 106 45.5
2 Stockfish-16.1 :   0.0   ----   617.0   1200 51.4 76   339 556 305 46.3
3 Lc0-BT4:68    : -8.9   25.2   195.5 400 48.9 89   106 179 115 44.8
4 Lc0-BT4:136 : -30.7   24.5   184.5 400 46.1   --- 87 195 118 48.8

Lc0 :mbs RATING ERROR POINTS PLAYED (%) CFS(%)   W D L D(%)
1 Lc0-t3:68 : 14.8   25.3   207.5 400 51.9 76   108 199   93 49.8
2 Lc0-t3:34 :   2.0   24.6   201.0 400 50.3 56   113 176 111 44.0
3 Stockfish-16.1 :   0.0   ----   597.5   1200 49.8 83   320 555 325 46.3
4 Lc0-t3:136    : -11.8   24.4   194.0 400 48.5   ---   104 180 116 45.0```

Lc0 :mbs          RATING  ERROR  POINTS  PLAYED  (%)  CFS(%)   W    D    L  D(%)
1 Lc0-BT4:34     :   5.9   24.6   203.0    400  50.8    68   112  182  106  45.5
2 Stockfish-16.1 :   0.0   ----   617.0   1200  51.4    76   339  556  305  46.3
3 Lc0-BT4:68     :  -8.9   25.2   195.5    400  48.9    89   106  179  115  44.8
4 Lc0-BT4:136    : -30.7   24.5   184.5    400  46.1   ---    87  195  118  48.8

Lc0 :mbs          RATING  ERROR  POINTS  PLAYED  (%)  CFS(%)   W    D    L  D(%)
1 Lc0-t3:68      :  14.8   25.3   207.5    400  51.9    76   108  199   93  49.8
2 Lc0-t3:34      :   2.0   24.6   201.0    400  50.3    56   113  176  111  44.0
3 Stockfish-16.1 :   0.0   ----   597.5   1200  49.8    83   320  555  325  46.3
4 Lc0-t3:136     : -11.8   24.4   194.0    400  48.5   ---   104  180  116  45.0

Bei dem großen und damit langsameren BT4-Netz ist die kleinere mbs besser,
und bei dem etwas schnelleren t3 geht schon die doppelte mbs von 68.
Ist aber sehr wahrscheinlich stark abhängig von der verwendeten Hardware.

Insgesamt nicht unerwartet, wenn auch wie so oft noch nicht aus der errorbar ...