+100 elo breakthrough in new rewritten Lco engine ( Ceres)

[JJJ]

Maybe Lczero will be the number 1 again !!!!!

[corres]

...
The 100+ elo gain and 2.3 x speed gain are mentioned on post 2, about
1. Hardwares
2. Testing conditions, time control and
3. Opening book
...

It seems to me an average ~ 50% enhancement in speed and a +30, +40 enhancement in Elo is a real supposing.
It is enough to win against Stockfish+NNUE but it is not a great breakthru for NN engines.

[Werner]

For all users of a GTX Card there is no improvement in speed. I made an example with my GTX 1650 running Ceres with LS15 and LC0 0.26.3 with LS15 under Arena 3.5.1 using a position from old WM test:

with Ceres,LS15, GTX 1650 nps 365
FEN: rnbr2k1/pp3p2/1qp3pp/4Q1N1/4P3/2P3P1/PP3PBP/R3R1K1 w - - 0 1
8/15 00:07 2k 331 +3,04 Sg5-f3 Lc8-g4 De5-f4 Lg4xf3 Lg2xf3 Sb8-d7 e4-e5 Sd7-c5 Te1-e2 h6-h5 Df4-f6 Td8-e8 Ta1-d1 Ta8-d8 Td1-d6
8/1 00:07 3k 337 +3,11 Lg2-h3
8/1 00:08 3k 340 +3,07 De5-d4
8/1 00:08 3k 346 +3,04 Te1-b1
8/16 00:09 3k 350 +2,97 Sg5xf7 Kg8xf7 Te1-e3 Kf7-g8 Lg2-f1 Sb8-d7 Lf1-c4+ Kg8-h7 De5-e7+ Kh7-h8 Te3-f3 Td8-g8 Lc4xg8 Kh8xg8 Ta1-d1 Db6-c5
8/16 00:10 4k 357 +2,95 Sg5xf7 Kg8xf7 Te1-e3 Kf7-g8 Lg2-f1 Sb8-d7 Lf1-c4+ Kg8-h7 De5-e7+ Kh7-h8 Te3-f3 Td8-g8 Lc4xg8 Kh8xg8 Ta1-d1 Db6-c5

with LC0 0.26.3 Cuda, LS15, GTX 1650, nps 1426
FEN: rnbr2k1/pp3p2/1qp3pp/4Q1N1/4P3/2P3P1/PP3PBP/R3R1K1 w - - 0 1
9/18 00:03 4k 1k +3,42 Sg5-f3 Lc8-g4 De5-f4 Lg4xf3 Lg2xf3 Sb8-d7 e4-e5 Sd7-c5 Te1-e2 h6-h5 Df4-h6 Td8-e8 Lf3xh5 g6xh5
9/18 00:03 5k 1k +3,47 Sg5xf7 Kg8xf7 Te1-e3 Kf7-g8 Lg2-f1 Td8-d7 De5-e8+ Kg8-g7 De8xc8 Td7-d8 Dc8-e6 Sb8-d7 Te3-f3 Td8-f8
9/18 00:03 5k 1k +3,37 Sg5-f3 Lc8-g4 De5-f4 Lg4xf3 Lg2xf3 Sb8-d7 e4-e5 Sd7-c5 Te1-e2 h6-h5 Df4-h6 Td8-e8 Lf3xh5 g6xh5
9/18 00:04 6k 1k +3,48 Sg5xf7 Kg8xf7 Te1-e3 Kf7-g8 Lg2-f1 Td8-d7 De5-e8+ Kg8-g7 De8xc8 Td7-d8 Dc8-e6 Sb8-d7 Te3-f3 Td8-f8
9/19 00:04 6k 1k +3,78 Sg5xf7 Kg8xf7 Te1-e3 Kf7-g8 Lg2-f1 Sb8-d7 Lf1-c4+ Kg8-h7 De5-e7+ Kh7-h8 Te3-f3 Td8-g8 Lc4xg8 Kh8xg8 Ta1-d1 Db6xb2 De7-e8+

So maybe there is an improvement for RTX cards. Here even LC0 0.26.3 is running faster with the CUDNN package using CUDNN-fp16 backend.

[Werewolf]

How many cpu threads does it use? Same as Lc0?

[AdminX]

How many cpu threads does it use? Same as Lc0?

Here is a view from Resource Monitor. MY system is a i7-4790, 32 GB Ram and a RTX 2070 Super

[Werner]

I used 2 CPU threads for LC0 as set in lc0.config. I hope Ceres is using the same configuration. The taskmanager is showing the same CPU usage here 12% from 8 threads (I have 4CPU)

[AdminX]

How does that compare with Lc0 v26?

Good Morning Albert,

I normally use the Dev Versions of LC0 and not Version 26. I will say that it appears to give me slight speed increases on my RTX card, even more so as games near the endgame phase on Big Nets when compared the versions I had been using. It does look like there may be some hope for improvement being as this is an early release. +100 Elo ??? I don't think so.

[AdminX]

Here is a 30 minute time control game with Komodo Dragon

My database has this (Nbc6) as the novelty

[d]rn1qkb1r/1p2nppp/p3p1b1/3pP3/3N1PP1/4B3/PPP1B2P/RN1Q1RK1 b kq - 0 10

[pgn]
[Event "Banksia game"]
[Date "2021.01.02"]
[White "Ceres"]
[Black "Remote KDragon"]
[Result "1-0"]
[TimeControl "40/1800+1"]
[Time "07:13:41"]
[Termination "mate"]
[ECO "B12"]
[Opening "Caro-Kann"]

1. e4 {+0.10/14 90074 1079920} c6 {+0.19/29 62743 293743677} 2. d4 {+0.14/13 74212 1064526} d5 {+0.16/30 4558 367569019}
3. e5 {+0.16/6 34061 1258566; B12: Caro-Kann, advance variation} Bf5 {+0.27/31 193636 908042969} 4. Nf3 {+0.17/5 35834 1495183} e6 {+0.31/27 55175 262308914}
5. Be2 {+0.17/4 24151 1575780} a6 {+0.27/28 70235 445647842} 6. O-O {+0.18/7 45083 1221015} c5
7. Be3 {+0.23/17 57917 564156} cxd4 {+0.37/30 47555 239021185} 8. Nxd4 {+0.23/9 37023 843911} Ne7
9. g4 {+0.22/5 30173 1133128} Bg6 {+0.81/29 75350 386966657} 10. f4 {+0.26/10 43519 864543} Nbc6 {+0.75/30 14626 298686572}
11. Nxc6 {+0.26/4 22863 1020966} bxc6 {+0.75/29 36159 304770110} 12. Nd2 Rb8 {+0.78/28 110474 577678926}
13. Ba7 {+0.28/6 37364 1150219} Rb4 {+0.72/29 137689 697377815} 14. Nb3 {+0.44/19 70889 592429} h5
15. gxh5 {+0.46/7 33214 829550} Be4 {+0.92/30 162110 1036531360} 16. c3 {+0.50/12 41672 843699} Rb7
17. Bf2 Qa8 {+1.07/32 103412 502606385} 18. Nc5 {+0.54/12 35849 915059} Rxb2 {+0.98/32 6704 238519278}
19. Nxe4 dxe4 {+1.06/29 30624 174131391} 20. Rb1 Rxb1 {+1.00/30 68766 378962018}
21. Qxb1 {+0.55/6 22729 1075034} Nd5 {+1.05/30 40033 347301065} 22. Qc1 f5 {+0.77/30 76043 409232620}
23. a3 {+0.61/17 75879 953116} e3 {+1.03/29 53444 306162623} 24. Bh4 {+0.55/10 44427 1214311} Be7 {+0.73/30 34433 204126050}
25. Bxe7 {+0.54/4 23205 1328798} Nxe7 {+0.74/29 14892 231865395} 26. Qxe3 Kf7 {+0.71/27 36780 223474985}
27. h3 {+0.70/13 54489 992850} Rb8 {+0.75/26 32797 200132515} 28. Rf2 {+0.72/15 46799 815919} Rb1+ {+0.65/30 51994 261294759}
29. Kh2 {+0.97/14 46042 569291} Qh8 30. Bf3 {+0.92/9 39282 1133597} Rb7 {+1.30/30 19263 325892312}
31. Qd4 {+1.05/12 44820 796309} Qh6 {+1.13/32 175082 1002160072} 32. Rg2 {+1.04/11 37166 966065} Rb3 {+1.14/30 9691 298210359}
33. a4 {+1.21/12 37932 647561} a5 34. Rg5 {+1.67/8 35095 1042659} Qh8
35. h6 {+2.28/9 37203 1149547} Qxh6 {+3.83/25 16162 31448978} 36. Rh5 {+3.69/9 24691 736714} c5
37. Qd2 {+7.02/11 29099 394768} Rb2 38. Qxb2 {+7.26/8 24056 581147} Qxf4+
39. Kg2 {+7.82/4 18483 771286} Ng6 {+5.05/27 7469 22074150} 40. Qb7+ {+9.70/13 370239 6964856} Kg8
41. Qc8+ Nf8 {+5.65/26 2669 12153184} 42. Qxc5 Qd2+ {+5.74/28 89840 77098756}
43. Qf2 {+11.64/7 108051 2546432} Qxc3 44. Rg5 {+12.16/6 121614 4021277} Qxe5 {+6.06/29 68445 146668592}
45. Rg3 {+12.91/6 117756 3465270} Kh7 {+5.52/28 261735 553478243} 46. Kh1 {+12.78/6 109591 2561925} Qf4 {+5.33/30 131958 466470659}
47. Rg2 {+14.28/7 101997 1822830} Qxa4 {+4.25/23 37954 13635217} 48. Qb2 {+12.93/7 94905 2178444} Qd7
49. Qb8 {+12.41/6 88316 2614702} Qe7 {+3.95/27 149750 113610534} 50. Qb7 {+12.80/7 82220 1867155} Qf6 {+4.01/22 28584 16996870}
51. Qb2 {+11.72/7 70309 1602425} e5 {+3.59/26 23089 9370470} 52. Qd2 {+11.37/6 71725 1616332} a4 {+4.17/26 110031 52504134}
53. Qg5 {+14.54/8 61793 930298} Qxg5 54. Rxg5 {+21.15/7 46706 1194211} g6
55. Rg3 {+20.95/7 59257 3114909} e4 {+4.85/26 55054 23887661} 56. Bd1 {+20.96/7 55199 1512228} Nd7
57. Bxa4 {+25.96/6 51398 1696032} Ne5 {+5.45/24 48664 26331266} 58. Rc3 {+23.02/6 47867 1554928} Kg7 {+4.69/24 50372 21037532}
59. Kg2 {+27.55/6 44580 1299755} g5 {+5.49/25 10538 24999167} 60. Rc5 {+30.99/6 41532 1009475} Kf6 {+5.09/21 27971 7237964}
61. Rc8 {+26.24/6 38714 797301} Nf3 {+5.90/23 48270 22870012} 62. Bd1 {+30.42/6 36065 936098} Nd4 {+6.09/27 60860 36032520}
63. Kf2 {+25.99/6 33610 831732} Ke5 {+6.70/27 17088 30079268} 64. Be2 {+21.87/6 31328 693502} Kf4 {+4.48/23 20388 6751588}
65. Bd1 {+20.95/7 29223 644092} Nb5 {+5.39/26 52186 19658562} 66. Rc4 {+22.80/7 27224 597098} Nd6 {+6.13/28 21214 10605200}
67. Rc5 {+22.95/7 25381 583826} Nf7 {+5.67/28 32372 17271325} 68. Bh5 {+22.60/6 23694 535787} Ne5 {+6.94/25 28151 22708884}
69. Kf1 {+23.59/7 22121 445601} Nd3 {+6.13/29 102386 47034642} 70. Ra5 {+23.72/7 20624 412218} Ne5 {+6.88/28 19869 19879105}
71. Be2 {+19.82/6 19252 490131} Ng6 {+6.45/30 63491 39600252} 72. Bc4 {+19.20/6 17972 422988} Nh4 {+5.29/29 36378 25607183}
73. Be6 {+18.89/7 16088 303534} e3 {+250.00/25 11892 6730516} 74. Ke2 {+21.10/6 14182 396580} Ke4
75. Bd7 {+44.21/6 14458 490400} Kf4 {+250.00/42 8618 2622443} 76. Kd3 {+99.99/6 11381 424066} Kg3 {+250.00/41 5450 1650048}
77. Kxe3 {+99.99/5 11902 587499} Ng2+ {+250.00/37 6837 1701247} 78. Ke2 {+99.99/5 12361 769165} Nf4+ {+250.00/47 12587 3904564}
79. Kf1 {+99.99/4 11560 646937} Kh4 {+250.00/65 555 9696656} 80. Bxf5 {+99.99/5 124288 16579304} Nh5
81. Kf2 {+99.99/2 99421 20736335} Ng7 {+250.00/82 11599 100442582} 82. Bg6 {+99.99/2 121298 17881772} Ne6 {+250.00/99 9258 26563491}
83. Ra8 {+99.99/2 112876 17995937} Nf4 84. Rh8+ Nh5
85. Rxh5# 1-0
[/pgn]

[MMarco]

How does that compare with Lc0 v26?

I did a quick bench on my 3080 of CERES and LC0-26.3-cuda. Default settings for CERES, defaults and NNcache=10M, smart-pruning-factor=0 for Lc0. The measured "speedup" (if we can say so, we are comparing nps from two different engines): 30x384 = 15%, 20x256 = 28%, 10x128 = 68%.

Startposition, go movetime 60000

30x384
*****************
LC0-66740
info depth 12 seldepth 42 time 59800 nodes 1383904 score cp 10 nps 25720
CERES-66740
info depth 14 seldepth 36 time 60006 nodes 1773283 score cp 8 nps 29552

20x256
******************
LC0-LS-15
info depth 23 seldepth 68 time 59800 nodes 4135219 score cp 15 nps 77013
CERES-LS-15
info depth 25 seldepth 53 time 56979 nodes 5618322 score cp 11 nps 98603

10x128
*******************
LC0-703810
info depth 14 seldepth 50 time 59800 nodes 8140461 score cp 12 nps 151242
CERES-703810
info depth 17 seldepth 51 time 59991 nodes 15210428 score cp 11 nps 253544

[Albert Silver]

How does that compare with Lc0 v26?

I did a quick bench on my 3080 of CERES and LC0-26.3-cuda. Default settings for CERES, defaults and NNcache=10M, smart-pruning-factor=0 for Lc0. The measured "speedup" (if we can say so, we are comparing nps from two different engines): 30x384 = 15%, 20x256 = 28%, 10x128 = 68%.

Startposition, go movetime 60000

30x384
*****************
LC0-66740
info depth 12 seldepth 42 time 59800 nodes 1383904 score cp 10 nps 25720
CERES-66740
info depth 14 seldepth 36 time 60006 nodes 1773283 score cp 8 nps 29552

20x256
******************
LC0-LS-15
info depth 23 seldepth 68 time 59800 nodes 4135219 score cp 15 nps 77013
CERES-LS-15
info depth 25 seldepth 53 time 56979 nodes 5618322 score cp 11 nps 98603

10x128
*******************
LC0-703810
info depth 14 seldepth 50 time 59800 nodes 8140461 score cp 12 nps 151242
CERES-703810
info depth 17 seldepth 51 time 59991 nodes 15210428 score cp 11 nps 253544

Ok, thanks. This is very interesting and actually seems to match what I suspected is happening. While there may be a normal speedup, if modest, over Lc0 v26, its biggest strength (for play, not for development) is removing the scaling shackles. Meaning smaller nets will gain the most from powerful cards, and above all will suffer less from the scaling issues on multi-gpu setups. Time will tell how accurate this read is, but it is what I am gathering from the results.