nodes match results

[Uri Blass]

I try to see how many nodes do engines need to get draw death for computer chess
It seems that 50M nodes per move vs 5M nodes per move for stockfish is usually draw
so I started some matches with node odds.


50000 nodes beat 5000 nodes 39.5-0.5

The only draw in the match is the following game.

[Event "Chess Tournament3"]
[Date "2021.10.03"]
[Round "1"]
[White "Stockfish 260921 dup"]
[Black "Stockfish 260921"]
[Result "1/2-1/2"]
[TimeControl "nodes: 5000000;nodes: 50000000"]
[Time "15:33:19"]
[Board "12"]
[Termination "repetition"]
[ECO "A03"]
[Opening "Bird's opening"]

1. f4 {End of opening} d5 {-0.25/8 438 5003; A03: Bird's opening} 2. e3 {-0.58/11 40 50051} Bf5 {-0.39/8 10 5006}
3. b3 {-0.25/14 56 50043} e6 {-0.33/9 10 5007} 4. Nf3 {-0.50/13 52 50037} Nf6 {-0.49/9 10 5005}
5. Bb2 {-0.47/13 55 50004} Be7 {+0.00/7 10 5001} 6. Nc3 {-0.65/13 53 50023} O-O {-0.66/9 10 5003}
7. Bd3 {-0.30/13 54 50005} Bxd3 {-0.50/8 10 5000} 8. cxd3 {-0.37/14 51 50049} c5 {-0.30/7 10 5005}
9. d4 {-0.25/14 50 50010} Nc6 {-0.19/9 10 5002} 10. dxc5 {-0.41/13 50 50025} Bxc5 {-0.22/9 10 5003}
11. O-O {-0.48/13 49 50026} a6 {-0.20/8 10 5002} 12. Ne2 {-0.44/12 51 50061} Ne4 {-0.25/8 10 5001}
13. Ng3 {-0.66/11 51 50022} f6 {-0.38/9 10 5000} 14. a3 {-0.70/11 50 50063} Ba7 {-0.75/8 10 5005}
15. Qc2 {-0.59/13 50 50031} Nxg3 {-0.93/9 10 5002} 16. hxg3 {-0.32/13 48 50004} Qd6 {-0.65/8 10 5002}
17. b4 {-0.24/13 52 50009} e5 {-0.42/9 10 5004} 18. fxe5 {-0.06/14 50 50030} fxe5 {-0.86/9 10 5006}
19. Ng5 {-0.25/13 52 50007} Qh6 {-0.74/9 10 5002} 20. Nh3 {-0.21/13 49 50013} d4 {-0.13/8 10 5000}
21. Rxf8+ {-0.08/13 53 50026} Rxf8 {-0.37/10 10 5010} 22. Qc4+ {+0.00/14 53 50066} Kh8 {-0.15/10 10 5006}
23. Rf1 {+0.00/14 50 50012} Rxf1+ {-0.01/9 10 5004} 24. Qxf1 {+0.08/14 51 50015} Qf6 {-0.12/8 10 5002}
25. Qd3 {+0.00/15 51 50008} Qd8 {-0.26/9 10 5001} 26. Qe4 {+0.00/14 50 50071} h6 {-0.20/8 10 5000}
27. Kh2 {+0.00/17 51 50011} dxe3 {+0.00/11 10 5011} 28. dxe3 {+0.00/20 45 50045} Qd2 {+0.00/12 10 5004}
29. Qg6 {+0.00/19 48 50048} Qxb2 {-2.20/10 10 5002} 30. Qe8+ {+0.00/27 37 50053} Kh7 {+0.00/11 10 5006}
31. Ng5+ {+0.00/31 44 50060} hxg5 {+0.00/23 10 5001} 32. Qh5+ {+0.00/38 38 50052} Kg8 {+0.00/25 10 5010}
33. Qe8+ {+0.00/41 32 50019} Kh7 {+0.00/26 10 5006} 34. Qh5+ {+0.00/44 31 50013} Kg8 {+0.00/28 10 5003}
35. Qe8+ {+0.00/50 30 50049} Kh7 {+0.00/245 39 245} 1/2-1/2

[Uri Blass]

Summary of results so far:
1)Stockfish 50K nodes beat 5K nodes 39.5-0.5
2)Stockfish 500K nodes per move-Stockfish 50K nodes per move 33-7(26 wins for the stronger side and 14 draws)
White as the weaker side made 9 draws and black as the weaker side made 5 draws.
3)Stockfish 5M nodes-Stockfish 50K nodes 37-3(34 wins for the stronger side and 6 draws)
5M nodes won 18-0 with black with 2 draws.

black could not win against 1.e4 or 1.f4 inspite of 100:1 nodes advantage.
Now same when 5M nodes per move get white.

5M nodes won with white 16-0 with 4 draws.
only opening it could not win are 1.f4 1.g4 1.h4 1.Nh3

4)Stockfish5M nodes-Stockfsih 500K nodes 27-13
5M won with white 3-0 and 17 draws.
5M won with black 11-0 and 9 draws
5M won with white only with 1.d4 1.f4 and 1.h3
5M won with black against 1.a3 1.a4 1.b3 1.b4 1.d3 1.e3 1.f3 1.f4 1.g3 1.g4 1.Na3


5)Stockfish50M nodes-Stockfish500K nodes
50M won with white 6-0 and 14 draws
50M won with 1.b4 1.d3 1.d4 1.e3 1.f4 1.Nh3
black is leading 3-0 and 5 draws and the 6th game with 1.e3 is probably another draw
50M won with black against 1.a4 1.b3 1.d4



Notes about matches that are not purely node matches

5M nodes per move is probably at similiar level to 60+0.6 based on my testing(60+0.6 won 20.5-19.5)
10M nodes per move could beat 60+0.6 22.5-17.5
5-0 with 35 draws(1 win with white and 4 wins with black)

It means that 25M nodes per move is probably at similiar level to 5+3 time control and that
5+3 time control with 7 cores is probably stronger than 50M nodes per move.

I tried stockfish with 7 cores and time control of 5+3 to see if it can beat stockfish 500K nodes per move in most games

Stockfish with 7 cores at time control 5+3 could get 12 wins and 8 draws with white against stockfish 500K nodes per move and I believe that black should be easier because there are some bad moves like 1.g4 for white.

[dangi12012]

Summary of results so far:
1)Stockfish 50K nodes beat 5K nodes 39.5-0.5
2)Stockfish 500K nodes per move-Stockfish 50K nodes per move 33-7(26 wins for the stronger side and 14 draws)
White as the weaker side made 9 draws and black as the weaker side made 5 draws.
3)Stockfish 5M nodes-Stockfish 50K nodes 37-3(34 wins for the stronger side and 6 draws)
5M nodes won 18-0 with black with 2 draws.

black could not win against 1.e4 or 1.f4 inspite of 100:1 nodes advantage.
Now same when 5M nodes per move get white.

5M nodes won with white 16-0 with 4 draws.
only opening it could not win are 1.f4 1.g4 1.h4 1.Nh3

4)Stockfish5M nodes-Stockfsih 500K nodes 27-13
5M won with white 3-0 and 17 draws.
5M won with black 11-0 and 9 draws
5M won with white only with 1.d4 1.f4 and 1.h3
5M won with black against 1.a3 1.a4 1.b3 1.b4 1.d3 1.e3 1.f3 1.f4 1.g3 1.g4 1.Na3


5)Stockfish50M nodes-Stockfish500K nodes
50M won with white 6-0 and 14 draws
50M won with 1.b4 1.d3 1.d4 1.e3 1.f4 1.Nh3
black is leading 3-0 and 5 draws and the 6th game with 1.e3 is probably another draw
50M won with black against 1.a4 1.b3 1.d4



Notes about matches that are not purely node matches

5M nodes per move is probably at similiar level to 60+0.6 based on my testing(60+0.6 won 20.5-19.5)
10M nodes per move could beat 60+0.6 22.5-17.5
5-0 with 35 draws(1 win with white and 4 wins with black)

It means that 25M nodes per move is probably at similiar level to 5+3 time control and that
5+3 time control with 7 cores is probably stronger than 50M nodes per move.

I tried stockfish with 7 cores and time control of 5+3 to see if it can beat stockfish 500K nodes per move in most games

Stockfish with 7 cores at time control 5+3 could get 12 wins and 8 draws with white against stockfish 500K nodes per move and I believe that black should be easier because there are some bad moves like 1.g4 for white.

I think you would need a graph of average elo gain per 10x increase of node count based on a starting count.
You cannot generalize from 500k to 50m == 50m to 5B!

I will help you create a graph with 16 Cores = 28 Cores of Zen3. I also have an extra Ryzen 3800X which I can let run overnight.
Can you create a script which I can just let run to get the results? (you would need the uci commands that your GUI uses to generate the matches)

I am interested in nodecount vs elo graph. So everyone knows how much elo current SF has with a given nodecount. Then we can graph older pre NNUE SF too to see how many nodes nnue needs less - to get the same elo.
Time control is irelevant since we talk about fixed nodecount per move. Also since a TT Table is build i will do 58 GB TT.

[Uri Blass]

Summary of results so far:
1)Stockfish 50K nodes beat 5K nodes 39.5-0.5
2)Stockfish 500K nodes per move-Stockfish 50K nodes per move 33-7(26 wins for the stronger side and 14 draws)
White as the weaker side made 9 draws and black as the weaker side made 5 draws.
3)Stockfish 5M nodes-Stockfish 50K nodes 37-3(34 wins for the stronger side and 6 draws)
5M nodes won 18-0 with black with 2 draws.

black could not win against 1.e4 or 1.f4 inspite of 100:1 nodes advantage.
Now same when 5M nodes per move get white.

5M nodes won with white 16-0 with 4 draws.
only opening it could not win are 1.f4 1.g4 1.h4 1.Nh3

4)Stockfish5M nodes-Stockfsih 500K nodes 27-13
5M won with white 3-0 and 17 draws.
5M won with black 11-0 and 9 draws
5M won with white only with 1.d4 1.f4 and 1.h3
5M won with black against 1.a3 1.a4 1.b3 1.b4 1.d3 1.e3 1.f3 1.f4 1.g3 1.g4 1.Na3


5)Stockfish50M nodes-Stockfish500K nodes
50M won with white 6-0 and 14 draws
50M won with 1.b4 1.d3 1.d4 1.e3 1.f4 1.Nh3
black is leading 3-0 and 5 draws and the 6th game with 1.e3 is probably another draw
50M won with black against 1.a4 1.b3 1.d4



Notes about matches that are not purely node matches

5M nodes per move is probably at similiar level to 60+0.6 based on my testing(60+0.6 won 20.5-19.5)
10M nodes per move could beat 60+0.6 22.5-17.5
5-0 with 35 draws(1 win with white and 4 wins with black)

It means that 25M nodes per move is probably at similiar level to 5+3 time control and that
5+3 time control with 7 cores is probably stronger than 50M nodes per move.

I tried stockfish with 7 cores and time control of 5+3 to see if it can beat stockfish 500K nodes per move in most games

Stockfish with 7 cores at time control 5+3 could get 12 wins and 8 draws with white against stockfish 500K nodes per move and I believe that black should be easier because there are some bad moves like 1.g4 for white.

I think you would need a graph of average elo gain per 10x increase of node count based on a starting count.
You cannot generalize from 500k to 50m == 50m to 5B!

I will help you create a graph with 16 Cores = 28 Cores of Zen3. I also have an extra Ryzen 3800X which I can let run overnight.
Can you create a script which I can just let run to get the results? (you would need the uci commands that your GUI uses to generate the matches)

I am interested in nodecount vs elo graph. So everyone knows how much elo current SF has with a given nodecount. Then we can graph older pre NNUE SF too to see how many nodes nnue needs less - to get the same elo.
Time control is irelevant since we talk about fixed nodecount per move. Also since a TT Table is build i will do 58 GB TT.

The GUI is free but I am not sure how I get the uci commands that it uses to generate the matches.
I can look at the uci protocol to learn possible commands but I think that it may be better to ask the creator of the gui.

[dangi12012]

Well just send me the commands - i am interested in a fixed nodecount SF vs SF NNUE matchup.
To see how many nodes nnue can prune better!

[Uri Blass]

Well just send me the commands - i am interested in a fixed nodecount SF vs SF NNUE matchup.
To see how many nodes nnue can prune better!

I am not sure if you got my private message so I will post the commands also here:

Use the command "go nodes n" when n is the number of nodes you want the engine to compute to that.

Here is the example BSG sends two commands to an engine, the first one to set up positon, the second one to compute:

position startpos moves g1f3 g8f6 d2d4 e7e6 c2c4 b7b6 e2e3 c8b7 f1e2 d7d5 e1g1 d5c4 e2c4 a7a6 c4e2 b8d7 c1d2 c7c5 d4c5 f8c5 b1c3 b6b5 b2b3 c5a3 e3e4
go nodes 1000

[dangi12012]

So I get this output:
info depth 1 seldepth 1 multipv 1 score cp 129 nodes 44 nps 22000 tbhits 0 time 2 pv a3b2
info depth 2 seldepth 2 multipv 1 score cp 417 nodes 95 nps 47500 tbhits 0 time 2 pv a3b2 a2a3 b2a1 d1a1
info depth 3 seldepth 3 multipv 1 score cp 361 nodes 156 nps 78000 tbhits 0 time 2 pv a3b2 h2h3 b2a1 d1a1
info depth 4 seldepth 4 multipv 1 score cp 242 nodes 496 nps 248000 tbhits 0 time 2 pv a3b2 e4e5 b2a1 d1a1
info depth 5 seldepth 5 multipv 1 score cp 188 nodes 928 nps 309333 tbhits 0 time 3 pv a3b2 e4e5 b2a1 d1a1 b7f3
info depth 6 seldepth 5 multipv 1 score cp 188 nodes 1001 nps 333666 tbhits 0 time 3 pv a3b2 e4e5 b2a1 d1a1 b7f3
bestmove a3b2 ponder e4e5

The only thing you need to do - and i will execute is to send me a list of commands and i will post the output here.
You sent me this position here
r2qk2r/1b1n1ppp/p3pn2/1p6/4P3/bPN2N2/P2BBPPP/R2Q1RK1 b kq - 0 13

[dangi12012]

How about that:
setoption name Threads value 30
setoption name Hash value 32000
position fen r2qk2r/1b1n1ppp/p3pn2/1p6/4P3/bPN2N2/P2BBPPP/R2Q1RK1 b kq - 0 13

go nodes 1000000000
.......
info depth 34 seldepth 47 multipv 1 score cp 220 upperbound nodes 1000000464 nps 31140051 hashfull 217 tbhits 0 time 32113 pv e8g8 d2c1
bestmove e8g8

So we get 32s for 1 billion nodes

[Uri Blass]

The problem in games is that the next command is based on the output you got from the engine so I cannot send a fixed list of commands to generate a game and you probably need to have some program to generate the commands.

giving position with no moves is also not the same as giving all the moves because the move may be also depdenent on the history of the game when there are rules that position can be draw by repetition because of some history.

After a move of one engine you need to send a command to the second engine based on the move that the first engine played(even if it is the engine against itself I do not want one engine with black to learn from the first engine with white when they search different number of nodes so I run the engine twice).

Example:
If I want to run manually match between Stockfish and itself that start with 1.e4 when white search 1000000 nodes per moves and black search 1000 nodes per move (usually I do it by GUI) then
I can give one engine something like the following commands:

Stockfish 071021 by the Stockfish developers (see AUTHORS file)
setoption name Threads value 3
setoption name Hash value 2048
position startpos moves e2e4
go nodes 1000
info string NNUE evaluation using nn-13406b1dcbe0.nnue enabled
info depth 1 seldepth 1 multipv 1 score cp -19 nodes 68 nps 34000 tbhits 0 time 2 pv e7e5
info depth 2 seldepth 2 multipv 1 score cp 95 nodes 212 nps 70666 tbhits 0 time 3 pv g8f6 d2d4
info depth 3 seldepth 3 multipv 1 score cp -9 nodes 390 nps 130000 tbhits 0 time 3 pv g8f6 b1c3 e7e5
info depth 4 seldepth 4 multipv 1 score cp -106 upperbound nodes 1002 nps 334000 tbhits 0 time 3 pv g8f6 e4e5
info depth 4 seldepth 4 multipv 1 score cp -22 nodes 1002 nps 334000 tbhits 0 time 3 pv e7e6 d2d4 d7d5
bestmove e7e6 ponder d2d4
position startpos moves e2e4 e7e6 d2d4
go nodes 1000
info string NNUE evaluation using nn-13406b1dcbe0.nnue enabled
info depth 1 seldepth 1 multipv 1 score cp -50 nodes 127 nps 63500 tbhits 0 time 2 pv c7c5
info depth 2 seldepth 2 multipv 1 score cp -71 nodes 339 nps 169500 tbhits 0 time 2 pv d7d5 b1c3
info depth 3 seldepth 3 multipv 1 score cp -71 nodes 636 nps 318000 tbhits 0 time 2 pv d7d5 b1c3 d5e4
info depth 4 seldepth 4 multipv 1 score cp -84 upperbound nodes 1006 nps 503000 tbhits 0 time 2 pv d7d5 b1c3
bestmove d7d5 ponder b1c3

Note that I know to write "position startpos moves e2e4 e7e6 d2d4" only because the second engine that I run at 1000000 nodes per move played this move and I gave the second engine the following commands

Stockfish 071021 by the Stockfish developers (see AUTHORS file)
setoption name Threads value 3
setoption name Hash value 2048
position startpos moves e2e4 e7e6
go nodes 1000000
info string NNUE evaluation using nn-13406b1dcbe0.nnue enabled
info depth 1 seldepth 1 multipv 1 score cp 74 nodes 103 nps 51500 tbhits 0 time 2 pv d2d4
info depth 2 seldepth 2 multipv 1 score cp 98 nodes 283 nps 94333 tbhits 0 time 3 pv b1c3 a7a6
info depth 3 seldepth 3 multipv 1 score cp 122 nodes 486 nps 162000 tbhits 0 time 3 pv b1c3 e6e5 d2d4 e5d4 d1d4
info depth 4 seldepth 4 multipv 1 score cp 71 nodes 1381 nps 460333 tbhits 0 time 3 pv d2d4 d7d5 b1c3 d5e4
info depth 5 seldepth 5 multipv 1 score cp 3 nodes 4535 nps 1133750 tbhits 0 time 4 pv d2d4 g8f6 d1d3 c7c5 d4d5
info depth 6 seldepth 7 multipv 1 score cp 66 nodes 6568 nps 1313600 tbhits 0 time 5 pv b1c3 d7d5 d2d4
info depth 7 seldepth 7 multipv 1 score cp 19 nodes 13000 nps 2166666 tbhits 0 time 6 pv b1c3 d7d5 d2d4 g8f6 e4e5 f6d7 c1e3
info depth 8 seldepth 12 multipv 1 score cp 29 nodes 20230 nps 2528750 tbhits 0 time 8 pv b1c3 d7d5 d2d4 f8b4 e4e5 c7c5 a2a3 c5d4
info depth 9 seldepth 11 multipv 1 score cp 54 nodes 26159 nps 2615900 tbhits 0 time 10 pv b1c3 d7d5 d2d4 f8b4 g1e2 g8f6 a2a3
info depth 10 seldepth 12 multipv 1 score cp 67 nodes 65093 nps 3254650 tbhits 0 time 20 pv d2d4 d7d5 e4d5 e6d5 g1f3 c7c6 f1d3 f8e7
info depth 11 seldepth 15 multipv 1 score cp 46 nodes 107544 nps 3258909 tbhits 0 time 33 pv d2d4 d7d5 b1c3 g8f6 c1g5 f8e7 e4e5 f6d7 g5e7 d8e7 f1d3 e8g8
info depth 12 seldepth 20 multipv 1 score cp 45 nodes 137648 nps 3357268 tbhits 0 time 41 pv d2d4 d7d5 e4e5 c7c5 c2c3 b8c6 g1f3 g8e7 f1d3 c5d4 c3d4 e7f5 d3f5 e6f5 b1c3 f8e7
info depth 13 seldepth 17 multipv 1 score cp 61 nodes 275521 nps 3444012 tbhits 0 time 80 pv d2d4 d7d5 b1c3 g8f6 e4e5 f6d7 f2f4 c7c5 g1f3 c5d4 f3d4 f8c5 c1e3 e8g8
info depth 14 seldepth 26 multipv 1 score cp 48 nodes 540250 nps 3441082 tbhits 0 time 157 pv d2d4 d7d5 e4e5 c7c5 c2c3 b8c6 g1f3 g8e7 f1d3 c5d4 c3d4 e7f5 d3f5 e6f5 b1c3 c8e6 h2h4 h7h6 c3e2 f8e7 h4h5
info depth 15 seldepth 23 multipv 1 score cp 32 nodes 739741 nps 3472962 tbhits 0 time 213 pv d2d4 d7d5 b1c3 g8f6 e4e5 f6d7 f2f4 c7c5 g1f3 b8c6 c1e3 a7a6 d1d2 b7b5 h2h4 d8a5 h4h5 h7h6 g2g4 c5d4 f3d4
info depth 16 seldepth 27 multipv 1 score cp 40 nodes 806503 nps 3461386 tbhits 0 time 233 pv d2d4 d7d5 b1c3 g8f6 e4e5 f6d7 f2f4 c7c5 g1f3 b8c6 c1e3 f8e7 a2a3 e8g8 f1d3 a7a6 e1g1 c5c4 d3e2
info depth 17 seldepth 24 multipv 1 score cp 43 nodes 960373 nps 3454579 tbhits 0 time 278 pv d2d4 d7d5 b1c3 g8f6 e4e5 f6d7 f2f4 c7c5 g1f3 b8c6 c1e3 a7a6 d1d2 d8b6 e1c1 b6a5 d4c5 f8c5 c1b1 b7b5 e3c5 d7c5 d2f2
info depth 18 seldepth 24 multipv 1 score cp 43 nodes 1004005 nps 3450189 tbhits 0 time 291 pv d2d4 d7d5 b1c3 g8f6 e4e5 f6d7 f2f4 c7c5 g1f3 b8c6 c1e3 a7a6 d1d2 d8b6 e1c1 b6a5 d4c5 f8c5 c1b1 b7b5 e3c5 d7c5 d2f2
bestmove d2d4 ponder d7d5

[Uri Blass]

For the starting positions you can simply try all legal moves of white so you start first game with the command
position startpos moves a2a3 and second game with position startpos moves a2a4...