Test Position

[Vinvin]

Even The King from CM9000 (of course, with only one thread) can find it in less than a minute:
...

That makes me angry with the modern chess engines : they can be MUCH weaker on some tactical shots despite the fact their rating is 700 points higher

+ modern engines seems to forget the purpose of the game : mate the king !
When there are complex sacrifices to extract the king and mate him, Stockfish is in the fog

[Dann Corbit]

Even The King from CM9000 (of course, with only one thread) can find it in less than a minute:
...

That makes me angry with the modern chess engines : they can be MUCH weaker on some tactical shots despite the fact their rating is 700 points higher

+ modern engines seems to forget the purpose of the game : mate the king !
When there are complex sacrifices to extract the king and mate him, Stockfish is in the fog

The pure NNUE version with the current code base finds it instantly (if you use a good net, not the awful one they bundle).
SF alpha-beta can't find it with 10 threads in 3 minutes.
SF hybrid can't find it with 10 threads in 3 minutes. (No surprise there, because the SF search is used when the position is imbalanced)
Sting-sf-24 takes a minute and a half to find the key move (but the eval never gets above draw in 3 minutes)
Bluefish-FD-XI finds the solution in one second.

Sting:
r1b1qr1k/2p3pp/4p3/1pb1PpN1/pn3N1P/8/PPP1QPP1/2KR3R w - - acd 30; acs 171 ; acn 1298201k; ce 0 ; bm Rd8; pv Rd8 Qxd8 Qh5 h6 Qg6 hxg5 Nxe6 Bxe6 hxg5+ Kg8 Qxe6+ Rf7 g6 Qg5+ Kb1 Raf8 gxf7+ Rxf7 Qe8+ Bf8 e6 Rf6 e7 Qxg2 exf8Q+ Rxf8 Qe6+ Rf7 Qe8+ Rf8;

Bluefish-FD-XI:
r1b1qr1k/2p3pp/4p3/1pb1PpN1/pn3N1P/8/PPP1QPP1/2KR3R w - - acd 26/50; acs 157 ; acn 1319790k; ce 480 ; bm Rd8; pv Rd8 Bd7 Rxe8 Bxe8 Qf3 Rb8 Nfxe6 Be7 Nxf8 Bxf8 a3 Nc6 Qxf5 Bg6 Qe6 Be8 h5 Be7 Nf7+ Bxf7 Qxf7 Rf8 Qd5 Nd8 Qxb5 Ne6 Kb1 Nc5 h6 Ne6 hxg7+ Kxg7;

Bleeding edge SF with nnue-only search:
r1b1qr1k/2p3pp/4p3/1pb1PpN1/pn3N1P/8/PPP1QPP1/2KR3R w - - acd 34/57; acs 172 ; acn 762217k; ce 922 ; bm Rd8; pv Rd8 Bd7 Rxe8 Bxe8 a3 Nc6 Nfxe6 Be7 Qxb5 Nd8 Qd3 Nxe6 Nxe6 Rg8 Qxf5 Rb8 Ng5 Bg6 Nf7+ Bxf7 Qxf7 Bc5 Rd1 Rbe8 Qxc7 Bxf2 Kb1 Bxh4 Qc4 Bg5 Rd5 Ra8 Qe4 h6 Ka2 Rgb8 Rd7 Bc1 Rb7 Rxb7 Qxb7 Re8 Qe4 Bg5 Qxa4 Rxe5 Qc4 Bf6 b4;

[Dann Corbit]

You can try the experiment yourself.
Get this:

use the binary in the source folder and the network in the net folder under the src folder.
You have to set the AlwaysNNUE UCI option to true.

[peter]

You can try the experiment yourself.
Get this:

use the binary in the source folder and the network in the net folder under the src folder.
You have to set the AlwaysNNUE UCI option to true.

Thank you, Dann!
Compile is even for SSE4.2 well working with quite good n/s on an old Xeon.

[MikeB]

From https://github.com/official-stockfish/S ... -670598949

Easy for Stockfish-nnue, but seems surprisingly hard for Stockfish (with classical evaluation)

[d]r1b1qr1k/2p3pp/4p3/1pb1PpN1/pn3N1P/8/PPP1QPP1/2KR3R w - - 0 1

[Edit for clarity] This position is difficult for Stockfish-dev even if Use NNUE is set to true. That's because the NNUE evaluation is not actually used because of the "NNUE evaluation threshold" patch.

Bluefish Nn on the lowly RPI 4 with 4 cores:

info depth 15 seldepth 23 multipv 1 score cp -38 nodes 742014 nps 781067 tbhits 0 time 950 pv d1d8 e8d8 e2h5 h7h6 h5g6 b4a2 c1b1 h6g5 h4g5 h8g8 f4h5 a2c3 b2c3 d8e7 h5f6 f8f6 e5f6
info depth 16 seldepth 25 multipv 1 score cp -33 nodes 814795 nps 801963 hashfull 2 tbhits 0 time 1016 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 b4d3 c2d3 d8e7 h5f6 f8f6 e5f6 e7f7 g6h7 g8f8 h7h8 f7g8
info depth 17 seldepth 29 multipv 1 score cp 0 nodes 887176 nps 813176 hashfull 2 tbhits 0 time 1091 pv d1d8 c5e3 c1b1 e8d8 e2h5 h7h6 h5g6 d8g5 h4g5 e3f4 h1h6 g7h6 g6h6 h8g8 h6g6 g8h8 g6h6
info depth 18 seldepth 29 multipv 1 score cp 54 nodes 977285 nps 838151 hashfull 3 tbhits 0 time 1166 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 d8d7 h5f6 f8f6 g5f6 c5e3 f2e3 b4a2 c1b1 g8f8 g6h7
info depth 19 seldepth 32 multipv 1 score cp 197 nodes 1302234 nps 903701 hashfull 3 tbhits 0 time 1441 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 b4a2 c1b1 f8f7 h5f6 g8f8 h1h8 f8e7 h8d8 g7f6 e5f6 e7d8 g6f7 a2b4 g5g6 b4d5 g6g7
info depth 20 seldepth 36 multipv 1 score cp 242 nodes 1606699 nps 946230 hashfull 4 tbhits 0 time 1698 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 f8f7 h5f6 g8f8 h1h8 f8e7 h8d8 g7f6 e5f6 e7d8 g6f7 b4a2 c1b1 a2b4 g5g6 c8d7 g6g7
info depth 21 seldepth 45 multipv 1 score cp 334 nodes 2706743 nps 996958 hashfull 4 tbhits 0 time 2715 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 f8f7 h5f6 g8f8 h1h8 f8e7 h8d8 e7d8 g6f7 c7c6 a2a3 b4d5 g5g6 a8a7 f7g8 d8c7 g8g7 c7b6 f6d5 e6d5 g7h8 b5b4 a3b4
info depth 22 seldepth 42 multipv 1 score cp 279 nodes 3208920 nps 977733 hashfull 5 tbhits 0 time 3282 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 f8f7 h5f6 g8f8 h1h8 f8e7 h8d8 e7d8 g6f7 c7c6 f7g8 d8c7 g8g7 c7b6 g7h8 c8b7 h8h4 a8d8 a2a3 b4a6 g5g6 b5b4 a3b4 c5d4 g6g7 d4e5 g7g8q
info depth 23 seldepth 44 multipv 1 score cp 302 nodes 3859816 nps 983893 hashfull 7 tbhits 0 time 3923 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 f8f7 h5f6 g8f8 h1h8 f8e7 h8d8 e7d8 g6f7 c7c6 f7g8 d8c7 g8g7 c7b6 g7h8 c8b7 h8h4 a8d8 a2a3 d8d4 h4h8 b4c2 c1c2 b5b4
info depth 24 seldepth 53 multipv 1 score cp 305 nodes 5564607 nps 963735 hashfull 14 tbhits 0 time 5774 pv d1d8 e8d8 e2h5 h7h6 h5g6 h6g5 h4g5 h8g8 f4h5 f8f7 h5f6 g8f8 h1h8 f8e7 h8d8 e7d8 g6f7 c7c6 f7g8 d8c7 g8g7 c7b6 g7g8 c8b7 g8e6 c5f2 a2a3 b7c8 e6g8 b4a6 g5g6 a6c7 g8h8 c8e6 g6g7 e6c4 g7g8q c4g8 f6g8