Alpha subZero - NN for CPU

pros:
-support polyglot book format
-suport 6 man Syzygy
-own net subzero.gz
-much faster than Lc0 on CPU
cons
-not working with all GUI
-work with arena,chessbase gui ,cutechess
-not work - Scid vs PC






[pgn][Event "games, Blitz 3m+2s"]
[Site "komp"]
[Date "2020.01.09"]
[Round "28"]
[White "Alpha subZero"]
[Black "Stockfish 070120 64 POPCNT"]
[Result "1-0"]
[ECO "B48"]
[Annotator "0.31;0.00"]
[PlyCount "105"]
[TimeControl "180+2"]

{Intel(R) Xeon(R) CPU  3.93GHz W=16.9 plies; 93kN/s; 312
TBAs B=28.4 plies; 9 366kN/s; 729 597 TBAs} 1. e4 c5 2. Nf3 e6 3. d4 cxd4 4.
Nxd4 Nc6 5. Nc3 Qc7 6. Be3 a6 7. Qd2 Nf6 8. O-O-O Be7 9. f3 b5 10. Kb1 Ne5 11.
g4 {Both last book move} b4 {[%eval 0,27] [%emt 0:00:20]} 12. Na4 {[%eval 31,
15] [%emt 0:00:06]} h6 {[%eval 0,29] [%emt 0:00:00]} 13. Bf4 {[%eval 105,16]
[%emt 0:00:13] (Rg1)} Rb8 {[%eval -14,24] [%emt 0:00:10] (Bb7)} 14. c4 {
[%eval 172,16] [%emt 0:00:04]} Bb7 {[%eval 42,27] [%emt 0:00:11] (bxc3)} 15. c5
{[%eval 134,16] [%emt 0:00:06]} O-O {[%eval 113,26] [%emt 0:00:03]} 16. Qc2 {
[%eval 130,17] [%emt 0:00:03]} Rbc8 {[%eval 96,28] [%emt 0:00:04]} 17. Nb6 {
[%eval 141,18] [%emt 0:00:00]} Qxb6 {[%eval 90,31] [%emt 0:00:08]} 18. cxb6 {
[%eval 141,18] [%emt 0:00:01]} Rxc2 {[%eval 99,31] [%emt 0:00:05]} 19. Kxc2 {
[%eval 159,18] [%emt 0:00:00]} Nxf3 {[%eval 103,31] [%emt 0:00:05]} 20. Nxf3 {
[%eval 161,18] [%emt 0:00:01]} Bxe4+ {[%eval 108,28] [%emt 0:00:03]} 21. Rd3 {
[%eval 158,18] [%emt 0:00:02]} Nd5 {[%eval 113,29] [%emt 0:00:02]} 22. Bc7 {
[%eval 185,17] [%emt 0:00:02]} Rc8 {[%eval 101,30] [%emt 0:00:02] (Nxc7)} 23.
Kb3 {[%eval 175,17] [%emt 0:00:07]} Nxc7 {[%eval 116,30] [%emt 0:00:03]} 24.
bxc7 {[%eval 172,18] [%emt 0:00:08]} Rxc7 {[%eval 117,30] [%emt 0:00:01]} 25.
Be2 {[%eval 185,17] [%emt 0:00:04]} a5 {[%eval 123,31] [%emt 0:00:12] (Bxd3)}
26. Rhd1 {[%eval 174,18] [%emt 0:00:20]} Bxd3 {[%eval 129,33] [%emt 0:00:00]}
27. Rxd3 {[%eval 172,18] [%emt 0:00:12]} d5 {[%eval 129,30] [%emt 0:00:04]
(Bc5)} 28. Ka4 {[%eval 173,17] [%emt 0:00:10]} Bc5 {[%eval 137,31] [%emt 0:00:
00] (Bd6)} 29. Rd1 {[%eval 203,17] [%emt 0:00:08] (Bd1)} Bb6 {[%eval 105,26]
[%emt 0:00:04]} 30. Bd3 {[%eval 195,17] [%emt 0:00:04] (Kb5)} g6 {[%eval 134,
31] [%emt 0:00:23] (Rc8)} 31. Rf1 {[%eval 237,17] [%emt 0:00:07]} Kg7 {[%eval
144,33] [%emt 0:00:13]} 32. Kb5 {[%eval 228,17] [%emt 0:00:02]} Bc5 {[%eval
112,31] [%emt 0:00:16]} 33. Kxa5 {[%eval 230,17] [%emt 0:00:00]} f5 {[%eval
124,30] [%emt 0:00:06] (Bd6)} 34. gxf5 {[%eval 215,17] [%emt 0:00:17]} gxf5 {
[%eval 138,31] [%emt 0:00:00]} 35. Re1 {[%eval 199,18] [%emt 0:00:26]} Kf6 {
[%eval 130,31] [%emt 0:00:00]} 36. Re2 {[%eval 209,16] [%emt 0:00:02]} Rb7 {
[%eval 126,30] [%emt 0:00:14]} 37. Ka4 {[%eval 216,17] [%emt 0:00:00]} Be7 {
[%eval 150,32] [%emt 0:00:40] (h5)} 38. h4 {[%eval 234,17] [%emt 0:00:06] (Ba6)
} Bc5 {[%eval 159,30] [%emt 0:00:21]} 39. h5 {[%eval 226,17] [%emt 0:00:17]
(Ba6)} b3 {[%eval 151,28] [%emt 0:00:10]} 40. a3 {[%eval 226,17] [%emt 0:00:00]
} Rg7 {[%eval 127,25] [%emt 0:00:01]} 41. Ne5 {[%eval 230,16] [%emt 0:00:02]
(Kb5)} f4 {[%eval 176,23] [%emt 0:00:08]} 42. Bg6 {[%eval 240,17] [%emt 0:00:
03]} Bd4 {[%eval 130,27] [%emt 0:00:00]} 43. Nc6 {[%eval 280,16] [%emt 0:00:09]
(Nf3)} Bc5 {[%eval 168,27] [%emt 0:00:09]} 44. Re1 {[%eval 286,17] [%emt 0:00:
07] (Ne5)} Rc7 {[%eval 165,24] [%emt 0:00:05] (Rg8)} 45. Nb4 {[%eval 316,16]
[%emt 0:00:04]} Rd7 {[%eval 169,25] [%emt 0:00:00] (Rb7)} 46. Rc1 {[%eval 330,
15] [%emt 0:00:03] (Be8)} Bxb4 {[%eval 222,26] [%emt 0:00:08] (Bd4)} 47. axb4 {
[%eval 373,17] [%emt 0:00:05]} e5 {[%eval 388,23] [%emt 0:00:00]} 48. b5 {
[%eval 607,16] [%emt 0:00:03]} e4 {[%eval 516,25] [%emt 0:00:02]} 49. b6 {
[%eval 768,16] [%emt 0:00:00]} e3 {[%eval 539,26] [%emt 0:00:05]} 50. Kb5 {
[%eval 954,17] [%emt 0:00:02]} Kg5 {[%eval 698,24] [%emt 0:00:01]} 51. Rg1+ {
[%eval 1174,17] [%emt 0:00:02] (Ka6)} Kh4 {[%eval 785,26] [%emt 0:00:04]} 52.
Bf5 {[%eval 1391,17] [%emt 0:00:00]} Rd8 {[%eval 946,28] [%emt 0:00:02]} 53.
Bg4 {[%eval 1422,17] [%emt 0:00:03]} 1-0[/pgn]

Download:
https://www.filemail.com/d/grxtwvlbtpudhyr

Hello bookmaker

Thanks for the post. Seems to be a strong engine

Wonder if the networks used with lc0 can be used with alpha too

Could you, please inform the link of the engine site?

It works not with infinity chess gui. That's too bad!

@MadMax777 wrote:

Hello bookmaker

Thanks for the post. Seems to be a strong engine

Wonder if the networks used with lc0 can be used with alpha too

Could you, please inform the link of the engine site?

its not support other net ,website is under construction.

Fire_Fox ... yes Infinity GUI also dont work,i dont know what causing problem.maybe they have engine list that is supported and others not.

Cheers.

thank you! And with the ctg adapter works very well. Now all is good

the engine plays well and uses only the CPU. Unfortunately after about 10 games it crashes and does not go beyond the moves of the book. Strangely, the file is protected with an obfuscator .... I will try to open it with Reverse engineering programs.....

@Alexa1978 wrote:

the engine plays well and uses only the CPU. Unfortunately after about 10 games it crashes and does not go beyond the moves of the book..

crashes can be caused by syzygy timing (too slow for large files), try if engine crashes change SyzygyProbe Limit from 6 to 5 for both engines in engine - engine testing.

Even after changing SyzygyProbe Limit engine crashes.Not so fast but after 10 match games its freezing.

https://wetransfer.com/downloads/01a27a7e5f6ec00fd15288c0e8432b2d20200111201150/fe70ce

Ok ,so this should work,i recompiled it with different settings ,just use this exe instead old one.I am testing it now in arena and cutechess and its not crash.

Now in Arena engine play much stably.In CB still nervous. I hope this project going forward, because engine play interesting and strong chess. Any information about network?

As I suspected, this is a clumsy compilation of Stockfish ....

I wouldn't say clumsy, but very skillful

well yes this is based on Sf indeed like any other strong modern engine but its not exactly the same engine,it plays different besides GPL licence which is SF and Lc0 allow change of code so this is no problem.

Here is example after 1200s (20 min) analysis of the same position (startposition) output is as follows



if it's the same engine the analysis tree variants after long time  will be also the same.SF candidat move after 20 min is 1.e4 Subzero 1.Nf3

Cheers

here another example (because in startposition is very limited possibilities) in this position SF by 10 mins analyze mostly 1 move Bc1-g5 ,Subzero at this time analyze 3 possibilities Bf4 ,a3 ,Nf5 and even not consider SF move Bg5.

How can I put 2048 hash?

Hash can be set to 0 ,this not using hash only cache which is a type of hash but self allocated.You can change cache size this way
- NN cache size limit --- is max value that cant be across
- ThreadCache multi values min 100 MB max 500 MB

and it works this way
if number of threads x cache multti is bigger than Cache size limit it sets up  this limit ,otherwise its allocate via thread cache multi option.

dear... bookmaker.... would be possible a compilation with more than 4 cpus? ... best

@bookmaker wrote:

Ok ,so this should work,i recompiled it with different settings ,just use this exe instead old one.I am testing it now in arena and cutechess and its not crash.

thanks bookmaker,
what is client.exe in the archive needed for?

I set it to 36 Lc0 cores (the max for my pc) and I got a whopping 51 nps! Could only reach depth 15 after a few mins. I'd have reached depth 30 in that same time with SF. Something wrong with this engine.

https://fil.email/FWknbjXz
try this compilation its new and should be ok ,but look if you got system 36 threads then you shouldn't set it to 36 but 32 because of 4 cumulative threads ,4 + 32 Lc0 threads = 36.My system is only 24 threads so i test it max on settings 4 + 20 Lc0

@mcdiniz wrote:

dear... bookmaker.... would be possible a compilation with more than 4 cpus? ... best

you can set number of threads whatever you want but you should set it in this way that Threads + Lc0 Threads should not ecxeed number of threads on your system.
I.e i have 24 thread so max settings i 4 + 20 Lc0.

@bookmaker wrote:

https://fil.email/FWknbjXz
try this compilation its new and should be ok ,but look if you got system 36 threads then you shouldn't set it to 36 but 32 because of 4 cumulative threads ,4 + 32 Lc0 threads = 36.My system is only 24 threads so i test it max on settings 4 + 20 Lc0

And what NPS do you get?

I doubt that if I set Threads to 4 and lower Lc0 threads to 32 that I'll get the expected speed. I should've got a lot more than just 51 nps....closer to 51,000 nps.

i get 65 -70 kns on max settings, on fast system you can adjust option OutOfOrder eval to limit selectivity then kn/s should jump a little.

Forgive my question, what does it mean that the other threads are reserved for Lc0? I ask you this because I noticed that the engine requires only the presence in the folder of subzero.zip, libopenblas.dll and client.exe without using them in the least, the engine only shows error if they are not present in folder, but otherwise doesn't use them.
Thank you