Hard-Talkchess-2020 set, final release

[Vinvin]

Position #169 is not good enough because 2 moves are clearly winning : Ng5 and Bxh7+

r1qn1rk1/p1pp1ppp/1p2p2b/1P2P3/P7/R2B1N2/1P4PP/3Q1R1K w - - bm Bxh7+; id "Hard-Talkchess-2020.169" ;

Code: Select all

Blue Marlin 14.4a-avx2:

 35/53	03:06	 5.439.122k	29.114k	+13,78	1.Ng5 f5 2.exf6 Bxg5 3.Bxh7+ Kxh7 4.Qh5+ Kg8 5.Qxg5 Rxf6 6.Rxf6 Nf7 7.Qg4 Ne5 8.Qf4 d6 9.Rg3 Qd7 10.Qh6 Qe7 11.Rf4 Rf8 12.Rxf8+ Qxf8 13.Qxe6+ Qf7 14.Qxf7+ Kxf7 15.Rc3 d5 16.Rxc7+ Ke6 17.Rxg7 Nd3 18.Kg1 Nxb2 19.Rxa7 d4 20.Kf2 Nd3+ 21.Kf1 Kd5 22.h4 Kc4 23.h5 Nf4 24.h6 d3 25.Ke1 Nxg2+ 26.Kd2 Nh4 27.h7
 36/58	03:06	 5.439.122k	29.114k	+21,43	1.Bxh7+ Kh8 2.Ng5 Nc6 3.bxc6 Qd8 4.Nxf7+ Rxf7 5.Rxf7 Kxh7 6.cxd7 Qg5 7.Raf3 Qc1 8.Qxc1 Bxc1 9.Rf8 Rxf8 10.Rxf8 Bg5 11.g4 Kg6 12.Rh8 Be7 13.h4 Kf7 14.d8Q Bxd8 15.Rxd8 Ke7 16.Rg8 g6 17.Rg7+ Kd8 18.g5 c5 19.Rxg6 Kc7 20.h5 Kc6 21.h6 Kd5 22.Rg7 Kxe5 23.Rf7 c4 24.g6 Kd6

[amanjpro]

Position #169 is not good enough because 2 moves are clearly winning : Ng5 and Bxh7+

r1qn1rk1/p1pp1ppp/1p2p2b/1P2P3/P7/R2B1N2/1P4PP/3Q1R1K w - - bm Bxh7+; id "Hard-Talkchess-2020.169" ;

Code: Select all

Blue Marlin 14.4a-avx2:

 35/53	03:06	 5.439.122k	29.114k	+13,78	1.Ng5 f5 2.exf6 Bxg5 3.Bxh7+ Kxh7 4.Qh5+ Kg8 5.Qxg5 Rxf6 6.Rxf6 Nf7 7.Qg4 Ne5 8.Qf4 d6 9.Rg3 Qd7 10.Qh6 Qe7 11.Rf4 Rf8 12.Rxf8+ Qxf8 13.Qxe6+ Qf7 14.Qxf7+ Kxf7 15.Rc3 d5 16.Rxc7+ Ke6 17.Rxg7 Nd3 18.Kg1 Nxb2 19.Rxa7 d4 20.Kf2 Nd3+ 21.Kf1 Kd5 22.h4 Kc4 23.h5 Nf4 24.h6 d3 25.Ke1 Nxg2+ 26.Kd2 Nh4 27.h7
 36/58	03:06	 5.439.122k	29.114k	+21,43	1.Bxh7+ Kh8 2.Ng5 Nc6 3.bxc6 Qd8 4.Nxf7+ Rxf7 5.Rxf7 Kxh7 6.cxd7 Qg5 7.Raf3 Qc1 8.Qxc1 Bxc1 9.Rf8 Rxf8 10.Rxf8 Bg5 11.g4 Kg6 12.Rh8 Be7 13.h4 Kf7 14.d8Q Bxd8 15.Rxd8 Ke7 16.Rg8 g6 17.Rg7+ Kd8 18.g5 c5 19.Rxg6 Kc7 20.h5 Kc6 21.h6 Kd5 22.Rg7 Kxe5 23.Rf7 c4 24.g6 Kd6

Zahak finds Ng5 fairly quickly, and sticks to it. Never changing to Bxh7+:

Code: Select all

info depth 33 seldepth 58 hashfull 970 tbhits 0 nodes 6877027458 nps 10004463 score cp 296 time 687395 pv f3g5 f7f5 e5f6 h6g5 d3h7 g8h7 d1h5 h7g8 h5g5 f8f6 f1f6 d8f7 g5g6 f7e5 g6g3 d7d6 f6f1 c8d7 a3c3 c7c5 b5c6 e5c6 g3g5 c6e5 c3g3 a8f8 f1f8 g8f8 h2h3 f8g8 b2b3 a7a6 h1g1 d7c7 g1h2

[Vinvin]

Position #177 is not good enough because 2 moves are clearly winning : N3h4 and Qxe1

r2qrbk1/1b3p2/3p2pB/1p1P1N2/1np1P3/5NRP/1P3PP1/1B1Qn1K1 w - - bm N3h4; id "Hard-Talkchess-2020.177" ;

Code: Select all

FEN: r2qrbk1/1b3p2/3p2pB/1p1P1N2/1np1P3/5NRP/1P3PP1/1B1Qn1K1 w - - 0 1

Blue Marlin 14.4a-avx2:
 51/19	37:56	 58.239.527k	25.587k	+9,90	1.Qxe1 Ra1 2.Kh2 Rxb1 3.Qxb1 Nd3 4.Ng5 Qf6 5.Bxf8 Rxf8 6.Qd1 Nxf2 7.Qe2 Nd3 8.Qg4 Bc8 9.Qh4 Qh8 10.Qxh8+ Kxh8
 51/77	37:56	 58.239.527k	25.587k	+11,36	1.N3h4 Qf6 2.Bg5 Qh8 3.Bd2 Qh7 4.Bxb4 Ra1 5.Qxe1 Bc8 6.Kh2 Bd7 7.f4 Qh5 8.Rg5 Qxg5 9.fxg5 Bxf5
---------------------------------------------------------------------------
 52/79	42:04	 64.952.113k	25.729k	+10,03	1.Qxe1 Nd3 2.Bxd3 cxd3 3.Qe3 Ra4 4.Bxf8 Rexe4 5.Rxg6+ fxg6 6.Qh6 Qxf8 7.Qxg6+ Kh8 8.Ng5 Re7 9.Ne6 Qf7
 52/17	42:04	 64.952.113k	25.729k	+11,36	1.N3h4 Qf6 2.Bg5 Qh8 3.Bd2 Qh7 4.Bxb4 Ra1 5.Qxe1 Bc8 6.Kh2 Bd7 7.f4 Qh5 8.Rg5 Qxg5 9.fxg5

[Vinvin]

A got my new computer 5 months ago :
CPU : AMD Ryzen 9 5950X 16-Core 3.40 GHz (running most of the time between 3.9 and 4.0 GHz)
Memory : DDR4 128GB PC 3200 CL16 G.Skill (8x16 GB)
OS : Windows 10 Professionnel

I first test it with Stockfish same version used on my old computer (POPNCT), after that with the AVX2 version and after that with the large page enable.
The result was as expected : around +10% speed with AVX2 and again +10% speed with large page enable.

Bench with SF 14.1 :
2 runs with 32 threads, CPU running around 3.5 GHz : 47.399.601 and 46.767.743 Nodes/second
2 runs with 16 threads, CPU running around 3.9 GHz : 35.689.329 and 35.268.145 Nodes/second

The numbers seem OK compared to the page : http://ipmanchess.yolasite.com/amd--int ... ckfish.php

With this new computer, I ran the Hard-Talkchess-2020 set revision 2021, the version with 114 positions.

Conditions :
- CPU : 5950X running around 3.9 GHz
- 16 threads used
- 5 minutes per position
- 6 man Syzygy + KRPPKRP
- 64 GB HashTable (still more than 50 GB for Syzygy cache)
- Several runs are needed because multithread is used. I fill the sheet and I use some formula to get average numbers.

Tests in the next message.

[Vinvin]

Results with the new computer.

Conditions :
- CPU : 5950X running around 3.9 GHz
- 16 threads used (to avoid hyper-threading)
- 5 minutes per position
- 6 man Syzygy + KRPPKRP
- 64 GB HashTable (still more than 50 GB for Syzygy cache)
- Several runs are needed because multithread is used. I fill the sheet and I use some formula to get average numbers.
- AVX2 version used everywhere when possible

List sorted from best to worst :
Name / #found-average / (Time with penalty in seconds) / number of runs
Blue Marlin 14.4a : 101,2 (84) on 11 runs
SugaR AI ICCF 1.90 : 101,0 (82) on 6 runs
ShashChess18.2 : 100,5 (92) on 11 runs
ShashChess20.1 : 98,7 (99) on 10 runs
ShashChess20 : 98,4 (101) on 10 runs
Crystal 3.2 : 98,2 (99) on 5 runs
Blue Marlin 14.5 : 98,0 (102) on 13 runs
ShashChess17.1 : 98,0 (111) on 6 runs
SugaR AI ICCF 2.40 : 96,8 (103) on 14 runs
Honey-v14 : 96,3 (116) on 8 runs
SugaR AI ICCF 2.50 : 93,9 (117) on 10 runs
Stockfish_21.11.23.21 : 92,4 (141) on 5 runs
Stockfish_21.08.05.16 : 91,2 (149) on 5 runs

Sheet with all results here (you can download the file if it doesn't display very well in the browser) : https://www.dropbox.com/s/wmgbzeesjcz4e ... 0.ods?dl=0

Still to do : some more runs with Crystal 3.2, test Crystal 3.1, remove some doubtful positions posted in this thread.

[AlexChess]

That is the demonstration that derivatives can be better than Stockfish, at least in tactical analysis.

Best regards, Alex

[Vinvin]

Added Crystal 3.1
+ one run with Crystal 3.2
+ some corrections in the sheet. Some formulas at far right was wrong.

Results with the new computer.

Conditions :
- CPU : 5950X running around 3.9 GHz
- 16 threads used (to avoid hyper-threading)
- 5 minutes per position
- 6 man Syzygy + KRPPKRP
- 64 GB HashTable (still more than 50 GB for Syzygy cache)
- Several runs are needed because multithread is used. I fill the sheet and I use some formula to get average numbers.
- AVX2 version used everywhere when possible

List sorted from best to worst :
Name / #found-average / (Time with penalty in seconds) / number of runs
Blue Marlin 14.4a : 101,2 (84) on 11 runs
SugaR AI ICCF 1.90 : 101,0 (82) on 6 runs
ShashChess18.2 : 100,5 (92) on 11 runs
ShashChess20.1 : 98,7 (99) on 10 runs
ShashChess20 : 98,4 (101) on 10 runs
Crystal 3.2 : 98,0 (99) on 6 runs
Blue Marlin 14.5 : 98,0 (102) on 13 runs
ShashChess17.1 : 98,0 (111) on 6 runs
SugaR AI ICCF 2.40 : 96,8 (103) on 14 runs
Crystal 3.1 : 96,4 (108) on 8 runs
Honey-v14 : 96,3 (116) on 8 runs
SugaR AI ICCF 2.50 : 93,9 (117) on 10 runs
Stockfish_21.11.23.21 : 92,4 (141) on 5 runs
Stockfish_21.08.05.16 : 91,2 (149) on 5 runs

Sheet with all results here (you can download the file if it doesn't display very well in the browser) : https://www.dropbox.com/s/ckc1fb1gfjm1h ... 5.ods?dl=0

Still to do : remove some doubtful positions posted in this thread.

[Jouni]

You have really invested for this testsuite (hardware + time)! How about testing tablebases influence? I am 99% sure, that they effect is less than SMP randomness.

[RubiChess]

That is the demonstration that derivatives can be better than Stockfish, at least in tactical analysis.

Best regards, Alex

I could write a SF derivate that solves 100% of these positions.
I could even write a Rubi derivate that solves 100%.

[jhellis3]

I could write a SF derivate that solves 100% of these positions.
I could even write a Rubi derivate that solves 100%.

I can turn lead into gold.