NNUE at LTC

[stavros]

its me or nnue is weaker at LTC? i see the longer TC the smaller difference than regular sf

[Raphexon]

ELO compression is the biggest factor probably.

Take FastGM.de
Difference between SF11 and Eth 12.25 is 253 elo at 60+0.6.
At 60m+15s it's 170 elo.

That doesn't mean Eth scales better than SF11, it's just less likely to blunder away a draw.
You can replace Ethereal with most engines on that list. That would mean either SF11 has horrible scaling, or elo compression is a thing.


To compare scaling it's better to take the time odds needed to make SFdev equal to SFNNUE at a given short TC.
Then increase TC and check if results have changed with the same time odds.

My 27-6 net also performed better at 10m+6s than at 60s+0.6s. (Fastgm list)

[stavros]

i see from mwyoung tournaments (TC 10min,30mins,32 threads) that nnue vs sf is close to =

[bmp1974]

NNUE plays beautiful positional chess at LTC (10min+20sec). Sample games
[pgn][Event "Stockfinn vs Fire rapid, Rapid 10.0min+"]
[Site "DESKTOP-IRD8JP2"]
[Date "2020.07.31"]
[Round "1"]
[White "SF+NNUE PO 270720 x64 avx2"]
[Black "Komodo 14 64-bit"]
[Result "1-0"]
[ECO "C18"]
[Annotator "0.37;0.35"]
[PlyCount "91"]
[TimeControl "600+20"]

{Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz 3600 MHz W=33.0 plies; 8,341kN/s;
731,556 TBAs; French Defense Powerbook 2012.ctg B=34.6 plies; 15,710kN/s; 549,
744 TBAs; French Defense Powerbook 2012.ctg} 1. e4 {B 0} e6 {B 0} 2. d4 {
0.37/34 25} d5 {B 0} 3. Nc3 {0.43/33 30} Bb4 {B 0} 4. e5 {0.35/31 16} c5 {B 0}
5. a3 {0.32/32 24} Bxc3+ {B 0} 6. bxc3 {0.22/34 25} Qa5 {B 0} 7. Bd2 {0.66/32
17} Qa4 {B 0} 8. Qb1 {0.52/33 35} c4 {B 0} 9. Be2 {0.70/34 22} Nc6 {B 0} 10.
Bh5 {0.79/39 77} Nge7 {0.35/33 36} 11. Ne2 {0.72/34 27} O-O {0.46/33 56 (Bd7)}
12. O-O {0.86/31 18} Nf5 {0.46/35 103 (Ng6)} 13. Bg4 {1.18/33 59} Nce7 {
0.46/34 28} 14. Re1 {0.97/32 31} Bd7 {0.56/35 112 (h6)} 15. Nf4 {1.51/32 26
(g3)} Qa5 {0.58/35 221 (Kh8)} 16. Nh5 {1.70/28 20} Kh8 {0.82/36 73} 17. Qd1 {
1.83/28 19 (Qc1)} Ba4 {0.76/35 90 (Rg8)} 18. Bxf5 {2.43/32 24 (Bh3)} Nxf5 {
1.14/32 17 (exf5)} 19. g4 {2.75/32 19} Ne7 {1.48/31 20} 20. Bg5 {3.01/32 38
(Re3)} Ng8 {1.03/31 19} 21. Qd2 {3.09/33 43 (Re3)} Qb5 {0.94/34 34} 22. Rab1 {
3.17/33 36 (f4)} Qe8 {1.44/32 37 (Qd7)} 23. Re3 {3.66/29 27 (Bh4)} f5 {1.97/34
58 (b5)} 24. exf6 {4.06/35 31} gxf6 {2.04/34 25} 25. Bf4 {4.18/35 27 (Bh4)} Qg6
{1.86/33 16} 26. f3 {4.47/35 55 (h3)} Rf7 {2.28/38 59 (b6)} 27. Rxe6 {4.86/31
21} Qxc2 {2.04/36 15 (Bxc2)} 28. Qxc2 {5.00/33 23} Bxc2 {2.42/37 16} 29. Rbe1 {
5.61/33 24} Bg6 {2.45/37 18 (Raf8)} 30. Nxf6 {6.37/33 22} Nxf6 {2.57/37 19} 31.
Be5 {6.23/32 26} Kg7 {2.73/39 44} 32. g5 {6.39/33 23 (Kf2)} Rd7 {3.19/36 45
(Kg8)} 33. Bxf6+ {7.17/31 25} Kf8 {3.35/34 19} 34. Kf2 {7.54/31 22} Re8 {
3.50/37 51} 35. Rxe8+ {7.45/31 34} Bxe8 {3.54/35 20} 36. Kg3 {7.57/31 35} Bg6 {
3.59/36 42} 37. Kf4 {7.92/32 33} Rc7 {3.76/37 41 (Bh5)} 38. h4 {8.51/31 28} Kf7
{3.59/36 20 (Bh5)} 39. Ke5 {10.23/30 25} Rc8 {4.01/35 20 (Kf8)} 40. Kxd5 {
11.65/31 33} Re8 {4.18/32 19} 41. Rxe8 {12.28/42 24 (Be5)} Kxe8 {5.64/33 21}
42. Kxc4 {12.88/38 22} Bh5 {6.28/34 20 (Bf7+)} 43. f4 {13.18/38 22} Kd7 {
6.88/34 20} 44. d5 {13.55/36 29} a5 {6.98/31 20 (Bf3)} 45. Kd4 {13.83/37 29} b5
{9.97/33 20 (Kd6)} 46. Ke5 {14.11/33 31 (Be5) adjud.} 1-0

[/pgn]
[pgn][Event "Stockfinn vs Fire rapid, Rapid 10.0min+"]
[Site "DESKTOP-IRD8JP2"]
[Date "2020.07.29"]
[Round "1"]
[White "SF+NNUE PO 270720 x64 avx2"]
[Black "Komodo 14 64-bit"]
[Result "1-0"]
[ECO "B07"]
[Annotator "0.37;0.52"]
[PlyCount "115"]
[TimeControl "600+20"]

{Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz 3600 MHz W=33.9 plies; 6,048kN/s; 6,
082,954 TBAs; Noomen.ctg B=36.5 plies; 16,962kN/s; 9,842,736 TBAs; Noomen.ctg}
1. e4 {B 0 White last book move} d6 {B 0} 2. d4 {0.37/34 25} Nf6 {B 0} 3. Nc3 {
0.31/31 18} g6 {B 0} 4. Be3 {0.49/28 27} Bg7 {B 0} 5. Qd2 {0.76/27 14} c6 {B 0}
6. Bh6 {0.58/32 42} Bxh6 {B 0} 7. Qxh6 {0.63/34 24} Qa5 {B 0} 8. Bd3 {0.95/28
19} c5 {B 0} 9. d5 {1.20/29 19} Nbd7 {B 0 Black last book move} 10. O-O-O {
1.16/32 32} b5 {0.52/35 36 (Ng4)} 11. Bxb5 {1.09/38 59} Rb8 {0.67/37 35} 12.
Bxd7+ {1.01/36 47} Bxd7 {0.53/36 27 (Kxd7)} 13. Nge2 {0.90/33 43 (f3)} Bb5 {
0.83/35 37 (Ng4)} 14. Rhe1 {0.81/37 56} Bxe2 {0.70/37 31} 15. Rxe2 {0.76/36 16}
Qb4 {0.73/38 43} 16. Kd2 {0.80/35 20} Qc4 {0.74/37 56} 17. Rb1 {1.11/38 65}
Qd4+ {0.76/37 64} 18. Kc1 {1.11/40 25} g5 {0.73/37 36} 19. Qh3 {1.18/38 36} Nd7
{0.84/37 30} 20. Nd1 {1.00/40 39 (b3)} Qf6 {0.82/35 30 (h5)} 21. Qf5 {1.41/32
22 (b3)} Qh6 {0.95/35 31} 22. Re3 {1.53/34 21} Qxh2 {0.91/39 68} 23. Qxg5 {
1.57/36 24} Nf6 {0.96/38 29 (h6)} 24. Rh3 {1.85/36 22} Qg1 {0.88/38 35} 25. e5
{1.62/34 29} Rg8 {1.07/40 39} 26. Qf5 {1.80/35 28} dxe5 {1.00/39 34} 27. Qxe5 {
2.02/38 36} Nd7 {1.19/37 26} 28. Qe4 {1.99/38 84} Rb4 {1.26/39 58} 29. Qxh7 {
2.13/36 33} Rxg2 {1.00/38 53} 30. Rh1 {2.08/36 46} Rg7 {1.14/38 30} 31. Rxg1 {
2.01/34 28} Rxh7 {1.16/37 25} 32. Rg8+ {2.10/33 22} Nf8 {1.16/5 0} 33. b3 {
2.14/34 28} Rf4 {1.02/37 43 (Rd4)} 34. Rb2 {1.93/32 57 (c4)} Rh1 {1.11/41 80
(f5)} 35. c4 {1.97/31 15} f5 {1.18/39 52} 36. Rg7 {2.28/32 70} Re1 {1.11/37 24}
37. Rg3 {2.18/30 9} Kf7 {1.20/37 26} 38. Re3 {1.90/29 47 (Rd2)} Rxe3 {1.03/35
27} 39. fxe3 {2.14/27 20} Rh4 {1.10/38 34} 40. Rg2 {1.96/29 46 (Rf2)} Nd7 {
1.19/36 35 (Kf6)} 41. Nf2 {2.45/27 10} Ne5 {0.90/37 23} 42. Kb2 {2.73/27 17} e6
{1.57/37 94 (Rh6)} 43. dxe6+ {3.07/25 10} Kxe6 {1.61/38 32} 44. Ka3 {3.89/29 30
} Rh7 {1.81/40 96 (Kd6)} 45. Ka4 {4.40/33 10} Kd6 {2.01/38 77 (Rb7)} 46. Nd1 {
4.85/30 20} Rb7 {1.99/36 13} 47. Nc3 {5.02/35 10} Rb4+ {2.06/36 14} 48. Ka3 {
5.56/36 15} Rb6 {2.22/36 14} 49. Rd2+ {5.79/35 16} Ke6 {2.12/38 25} 50. e4 {
6.02/35 13} Ra6+ {2.09/36 16 (fxe4)} 51. Kb2 {6.22/35 15} fxe4 {2.13/39 39} 52.
Nxe4 {6.51/36 15} Rc6 {2.21/35 14} 53. Rd8 {7.32/37 17 (Rd5)} Ng6 {2.21/36 27
(Kf5)} 54. Rd5 {8.45/36 13} Nf4 {2.34/37 20} 55. Rg5 {8.80/38 14} Nd3+ {
2.45/37 13} 56. Kc3 {9.02/40 15 (Kc2)} Ne5 {2.52/36 15} 57. Rh5 {9.21/40 15}
Rc7 {2.80/35 17} 58. Kb2 {9.30/44 15 (Kc2) adjud.} 1-0

[/pgn]
[pgn][Event "Stockfinn vs Fire rapid, Rapid 10.0min+"]
[Site "DESKTOP-IRD8JP2"]
[Date "2020.07.27"]
[Round "1"]
[White "SF+NNUE AIO 270720 x64 avx2"]
[Black "Deep Shredder 13 x64"]
[Result "1-0"]
[ECO "C96"]
[Annotator "0.11;0.21"]
[PlyCount "105"]
[TimeControl "600+20"]

{Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz 3600 MHz W=63.9 plies; 8,430kN/s;
315,785 TBAs; Noomen.ctg B=26.4 plies; 16,303kN/s; Noomen.ctg} 1. e4 {B 0} e5 {
B 0} 2. Nf3 {B 0} Nc6 {B 0} 3. Bb5 {B 0} a6 {B 0} 4. Ba4 {B 0} Nf6 {B 0 Both
last book move} 5. O-O {0.11/29 31} Be7 {0.21/29 30} 6. Re1 {0.22/31 22 (Bxc6)}
b5 {0.13/29 35} 7. Bb3 {0.22/32 26} d6 {0.10/30 31 (0-0)} 8. c3 {0.21/34 19}
O-O {0.10/30 50} 9. h3 {0.14/36 22} Na5 {0.14/29 47} 10. Bc2 {0.13/35 18} c5 {
0.15/29 23} 11. d4 {0.14/35 33} exd4 {0.23/29 71 (Qc7)} 12. cxd4 {0.25/31 19}
Re8 {0.20/29 37 (cxd4)} 13. Nc3 {0.29/35 65 (Bf4)} Bb7 {0.21/31 47 (cxd4)} 14.
d5 {0.34/29 17 (b3)} Nd7 {0.23/31 50} 15. b3 {0.31/33 27} Bf6 {0.24/30 40} 16.
Bb2 {0.27/35 32} Ne5 {0.28/31 23 (Bc8)} 17. Nxe5 {0.30/68 134} Bxe5 {0.27/31 46
} 18. Rb1 {0.29/36 20} Bc8 {0.32/31 31} 19. Ne2 {0.33/34 23} Bxb2 {0.33/31 73
(Bd7)} 20. Rxb2 {0.38/35 35} Qf6 {0.35/31 88 (Bd7)} 21. Rb1 {0.33/33 20} Bd7 {
0.32/31 23 (Nb7)} 22. f4 {0.46/34 88 (Qd2)} b4 {0.35/29 97 (Nb7)} 23. Qd2 {
0.51/34 39} Nb7 {0.38/28 41} 24. Ng3 {0.58/31 39} Bb5 {0.51/30 65 (Qh6)} 25.
Rbd1 {0.75/31 30} g6 {0.60/30 18} 26. Qe3 {0.71/31 28} Qg7 {0.67/30 40 (Qb2)}
27. e5 {0.86/30 25} f5 {0.67/29 14 (Rf8)} 28. Bxf5 {1.36/32 50 (Nf1)} gxf5 {
0.63/31 31} 29. Nxf5 {1.26/30 15} Qg6 {0.87/31 43} 30. g4 {1.12/30 22} h5 {
0.94/31 35 (a5)} 31. Rd2 {1.89/27 24 (Qe4)} Rab8 {1.94/25 30 (h4)} 32. e6 {
2.49/26 46 (Nh4)} Rbc8 {2.31/28 15 (h4)} 33. Ng3 {5.22/26 20 (Rg2)} Qh6 {
4.72/27 28} 34. g5 {5.91/27 27} Qh8 {6.28/28 26} 35. f5 {6.86/28 33} Rf8 {
6.82/28 24 (Qe5)} 36. f6 {9.68/29 23 (e7)} Qh7 {10.32/29 22} 37. f7+ {10.34/32
28} Kh8 {11.51/30 22} 38. Rf2 {10.79/32 23 (Qf4)} Nd8 {12.24/29 20} 39. Rf6 {
11.29/36 30 (Qf4)} Bd3 {13.20/26 21 (Nxf7)} 40. Rh6 {24.14/36 20} Nxf7 {
15.62/26 15} 41. exf7 {68.72/40 22} Bg6 {14.22/27 16 (Rxf7)} 42. Rf1 {#12/71
20 (Rxh7+)} Rc7 {#12/35 13} 43. Qe6 {#11/77 20} Qxh6 {#11/22 0} 44. gxh6 {
#10/86 21} Kh7 {#10/20 0} 45. Ne4 {#9/90 23} Rfxf7 {#9/18 0} 46. Ng5+ {#8/116
19} Kxh6 {#8/20 0} 47. Nxf7+ {#7/180 19 (Rxf7)} Kh7 {#9/18 0} 48. Ng5+ {
#6/245 17} Kh6 {#7/16 0} 49. h4 {#5/245 2 (Qxd6)} Kg7 {#5/13 0 (Rg7)} 50. Qf6+
{#4/245 0} Kh6 {#4/8 0} 51. Qf8+ {#3/245 0 (Qh8+)} Rg7 {#4/8 0} 52. Ne6 {
#2/245 0 (Qh8+)} Kh7 {#1/2 0} 53. Qxg7# {#1/245 0} 1-0

[/pgn]

[bmp1974]

NNUE played with 1 book move. (It was on its own after 1st move)

[corres]

Longer TC give more help to the slower engine. It is obvious from Stockfish(dev) - StockfishNNUE games also.

[Milos]

NNUE plays beautiful positional chess at LTC (10min+20sec).

That is not LTC, that is called rapid.

[Gino Figlio]

NNUE and regular stockfish are the same in correspondence time control or analysis

[Ovyron]

NNUE and regular stockfish are the same in correspondence time control or analysis

I find them completely different, worlds apart. You can always guide the position to places where NNUE outperforms regular Stockfish (just give bonuses for positions where they disagree, as the disagreement grows you reach positions where regular Stockfish's eval isn't good at all.)

[Zenmastur]

NNUE plays beautiful positional chess at LTC (10min+20sec). Sample games

Yes, but does it play stronger is the question. Beauty is very subjective. ELO isn't.

NNUE and regular stockfish are the same in correspondence time control or analysis

NNUE has been around long enough for this to be determined. A single CC game can last 2 years or more. NNUE has been out for a few weeks. Thousands of games or even hundreds of thousands of CC time control games would need to be played, depending on the rating difference, before this could be accurately determined. So, I think your claims are a bit premature.

NNUE and regular stockfish are the same in correspondence time control or analysis

I find them completely different, worlds apart. You can always guide the position to places where NNUE outperforms regular Stockfish (just give bonuses for positions where they disagree, as the disagreement grows you reach positions where regular Stockfish's eval isn't good at all.)

In general I would agree. The question is how much ELO does this translate to. I think it's yet to be determined.

Regards,

Zenmastur