StockfishMZ

[Peter Grayson]

Hi Oscar,
try with Material Eval Strategy -5 and Positional Eval Strategy +7
Move Overhead=1000

However, you did not specify whether you are using books or not

Hi, Marco
Thanks for your reply!
No book I'm using A00-A25 ECO file, with 8 moves
Thank you!

Marco,
as you can see at image below, SF MZ 2405 is not working right for me...
After leaving book (books mine as well as Exp file as book) engine start to move very badly (-1.50 even -2.00).
I'm playing 5min+2
Hash 2048
MES -5
PES 7
NNUE evaluation on
Classical Evaluation off
MCTS on
MCTS Threads 1
I don't understand which is wrong with this, but I've played 8 matches and this error is the same in all of them...
Can you help me find the bug or my error?
Thank you!

I am able to reproduce the lower ply depth and bad move output when the nnue network file is not in the engine folder.

When unpacked, the 230522 download included the game and ICCF playing engines with the network in the folder a level above the engine folders. The nnue network file nn-3c0aa92af1da.nnue must be copied and pasted into the game playing engine folder and the ICCF engine folder. Without the nnue file the engine will play sensibly when the Classic element is used but will produce strange moves and ply depths when nnue use is called for but the nnue file is absent.

Peter

[CrapCleaner]

Any chance that Marco or ChessMan can compile embeeded net of StockfishMZ Iccf 240522?
I have a friend who has problem setting path to the network. It would be of a great help to him if the engine already is compiled together with nnue.

[catalano7777]

Hi, Marco
Thanks for your reply!
No book I'm using A00-A25 ECO file, with 8 moves
Thank you!

Marco,
as you can see at image below, SF MZ 2405 is not working right for me...
After leaving book (books mine as well as Exp file as book) engine start to move very badly (-1.50 even -2.00).
I'm playing 5min+2
Hash 2048
MES -5
PES 7
NNUE evaluation on
Classical Evaluation off
MCTS on
MCTS Threads 1
I don't understand which is wrong with this, but I've played 8 matches and this error is the same in all of them...
Can you help me find the bug or my error?
Thank you!

I am able to reproduce the lower ply depth and bad move output when the nnue network file is not in the engine folder.

When unpacked, the 230522 download included the game and ICCF playing engines with the network in the folder a level above the engine folders. The nnue network file nn-3c0aa92af1da.nnue must be copied and pasted into the game playing engine folder and the ICCF engine folder. Without the nnue file the engine will play sensibly when the Classic element is used but will produce strange moves and ply depths when nnue use is called for but the nnue file is absent.

Peter

NNUE file is in the same folder as SF MZ. I've tried with NNUE evaluation On and Classical evaluation Off, both On, both Off (I mean all variants), but it's the same...
Thank you, Peter!

[Peter Grayson]

Marco,
as you can see at image below, SF MZ 2405 is not working right for me...
After leaving book (books mine as well as Exp file as book) engine start to move very badly (-1.50 even -2.00).
I'm playing 5min+2
Hash 2048
MES -5
PES 7
NNUE evaluation on
Classical Evaluation off
MCTS on
MCTS Threads 1
I don't understand which is wrong with this, but I've played 8 matches and this error is the same in all of them...
Can you help me find the bug or my error?
Thank you!

I am able to reproduce the lower ply depth and bad move output when the nnue network file is not in the engine folder.

When unpacked, the 230522 download included the game and ICCF playing engines with the network in the folder a level above the engine folders. The nnue network file nn-3c0aa92af1da.nnue must be copied and pasted into the game playing engine folder and the ICCF engine folder. Without the nnue file the engine will play sensibly when the Classic element is used but will produce strange moves and ply depths when nnue use is called for but the nnue file is absent.

Peter

NNUE file is in the same folder as SF MZ. I've tried with NNUE evaluation On and Classical evaluation Off, both On, both Off (I mean all variants), but it's the same...
Thank you, Peter!

It may be worth confirming the nnue network is being read correctly by the engine.

Open the engine folder and
Right click the engine you use.
Select "Run as administrator" This will open engine in command mode.

The information displayed on screen should be similar to ...
StockfishMZ 230522 by the Stockfish developers, M.Z (see AUTHORS file)

Compiled by : g++ (GNUC) 10.2.0 on MinGW64
Compile settings : 64bit SSE2
Build date/time : 2022-May-23 23:29:11

Operating System (OS) : Windows 10 Client Or Greater
CPU Brand : Intel(R) Core(TM)2 Quad CPU Q9550 @ 2.83GHz
NUMA Nodes : 1
Cores : 4
Threads : 4
Hyper-Threading : No
L1/L2/L3 cache size : 256KB/12MB/N/A
Memory installed (RAM): 8GB

info string StockfishMZ.exp -> Total moves: 8. Total positions: 6. Duplicate moves: 0. Fragmentation: 0.00%
Large Memory Pages : available

info string Evaluation network loaded successfully ... (My highlight)

The last line confirms the engine correctly loaded the nnue file. If the last text line is missing the engine has not identified the nnue file. In that case it may be worth overwriting the nnue file with the original or downloading a fresh copy from ...

https://tests.stockfishchess.org/api/nn/nn-3c0aa92af1da.nnue

If opening the engine in command mode and the initial screen confirms network has loaded then the issue would need more detailed investigation.

Peter

[catalano7777]

I am able to reproduce the lower ply depth and bad move output when the nnue network file is not in the engine folder.

When unpacked, the 230522 download included the game and ICCF playing engines with the network in the folder a level above the engine folders. The nnue network file nn-3c0aa92af1da.nnue must be copied and pasted into the game playing engine folder and the ICCF engine folder. Without the nnue file the engine will play sensibly when the Classic element is used but will produce strange moves and ply depths when nnue use is called for but the nnue file is absent.

Peter

NNUE file is in the same folder as SF MZ. I've tried with NNUE evaluation On and Classical evaluation Off, both On, both Off (I mean all variants), but it's the same...
Thank you, Peter!

It may be worth confirming the nnue network is being read correctly by the engine.

Open the engine folder and
Right click the engine you use.
Select "Run as administrator" This will open engine in command mode.

The information displayed on screen should be similar to ...
StockfishMZ 230522 by the Stockfish developers, M.Z (see AUTHORS file)

Compiled by : g++ (GNUC) 10.2.0 on MinGW64
Compile settings : 64bit SSE2
Build date/time : 2022-May-23 23:29:11

Operating System (OS) : Windows 10 Client Or Greater
CPU Brand : Intel(R) Core(TM)2 Quad CPU Q9550 @ 2.83GHz
NUMA Nodes : 1
Cores : 4
Threads : 4
Hyper-Threading : No
L1/L2/L3 cache size : 256KB/12MB/N/A
Memory installed (RAM): 8GB

info string StockfishMZ.exp -> Total moves: 8. Total positions: 6. Duplicate moves: 0. Fragmentation: 0.00%
Large Memory Pages : available

info string Evaluation network loaded successfully ... (My highlight)

The last line confirms the engine correctly loaded the nnue file. If the last text line is missing the engine has not identified the nnue file. In that case it may be worth overwriting the nnue file with the original or downloading a fresh copy from ...

https://tests.stockfishchess.org/api/nn/nn-3c0aa92af1da.nnue

If opening the engine in command mode and the initial screen confirms network has loaded then the issue would need more detailed investigation.

Peter

As you can see, nnue file is loaded successfully

I still cannot understand why such those weak moves...
Thank you, Peter!

[Peter Grayson]

[/quote]
As you can see, nnue file is loaded successfully

I still cannot understand why such those weak moves...
Thank you, Peter!
[/quote]

One other observation from the original snapshot is that the White BMI2 engine is dated 240522 that means it is the Chessman compile whereas the Black AVX2 engine is dated 230522 that is the original MZ compile. From what I can identify in the game score the issue is solely with the White engine. Having checked the Intel Xeon E5-1630 v3 specification it is not fully BMI2 compatible with the specification confirming AVX2 compatibility. See ...

https://www.intel.com/content/www/us/en/products/sku/82764/intel-xeon-processor-e51630-v3-10m-cache-3-70-ghz/specifications.html

I seem to recall comments in the past that BMI2 engines may run on an AVX2 machine but considerably slower and with poor results. Try changing the BMI2 engine to the Chessman AVX2 compile and check if the problem persists.

Alternatively copy and paste the MZ 230522 engine folder to create a second instance of the MZ compiles and use the second instance of the MZ avx2 compile to replace the BMI2 engine to compare performance with the Black AVX2 engine. They should then be the same.

Peter

[Anton101]

Hi! Catalano7777, please download CPU-Z a program that you will find in Google, this program will help you know which executables your PC works with.
As an additional comment, I used StockfishMZ ICCF only for three games on the playchess server and all three were draws. I left the parameters by default and, just increase Move Overhead to 1000, I did it just out of curiosity. The idea of building ICCF is another obviously, but total, it was only three games.

I hope I was able to help you.

Regards!.

[kramnik]

The same release with the embedded network, as you've requested:

https://pixeldrain.com/u/QvamtYtk

Any chance that Marco or ChessMan can compile embeeded net of StockfishMZ Iccf 240522?
I have a friend who has problem setting path to the network. It would be of a great help to him if the engine already is compiled together with nnue.

[Xalomon]

try with Material Eval Strategy -5 and Positional Eval Strategy +7

I tried these settings at playchess and suffered some loses (engine showing 0.00 for losing move while opponent shows big advantage), however then I tried Material Eval Strategy -5 and Positional Eval Strategy +6 and it seems to work well - so probably settings around that range are usable and make the engine play differently from others.

[catalano7777]

As you can see, nnue file is loaded successfully

I still cannot understand why such those weak moves...
Thank you, Peter!
[/quote]

One other observation from the original snapshot is that the White BMI2 engine is dated 240522 that means it is the Chessman compile whereas the Black AVX2 engine is dated 230522 that is the original MZ compile. From what I can identify in the game score the issue is solely with the White engine. Having checked the Intel Xeon E5-1630 v3 specification it is not fully BMI2 compatible with the specification confirming AVX2 compatibility. See ...

https://www.intel.com/content/www/us/en/products/sku/82764/intel-xeon-processor-e51630-v3-10m-cache-3-70-ghz/specifications.html

I seem to recall comments in the past that BMI2 engines may run on an AVX2 machine but considerably slower and with poor results. Try changing the BMI2 engine to the Chessman AVX2 compile and check if the problem persists.

Alternatively copy and paste the MZ 230522 engine folder to create a second instance of the MZ compiles and use the second instance of the MZ avx2 compile to replace the BMI2 engine to compare performance with the Black AVX2 engine. They should then be the same.

Peter
[/quote]
Peter, the black AVX2 is not mine. That belongs to my rival, from another country, with another machine... Mine is white BMI2, because I've tested all versions with Build Test, a tool to test engines, and BMI2 build is the fastest version for my machine..
Anyway I'll try AVX2, but I'm not sure where obtain the Chessman compiles from
Thank you, Peter!

[Peter Grayson]

Peter, the black AVX2 is not mine. That belongs to my rival, from another country, with another machine... Mine is white BMI2, because I've tested all versions with Build Test, a tool to test engines, and BMI2 build is the fastest version for my machine..
Anyway I'll try AVX2, but I'm not sure where obtain the Chessman compiles from
Thank you, Peter!

The Chessman compiles were posted earlier in this thread ...
http://outskirts.altervista.org/forum/viewtopic.php?p=45982&view=single_post#p45982

Marco's original release was
https://outskirts.altervista.org/forum/viewtopic.php?p=45951&view=single_post#p45951

It is worth trying one of the non BMI2 engines for comparison especially the AVX2 compile because that should run without issue on your Xeon CPU machine given its specification. The original StockfishMZ 230522 AVX2 Compile should also work too. If the same engine behaviour occurs with both AVX2 engines some further head scratching will be necessary.

Best regards,
Peter

[catalano7777]

Peter, the black AVX2 is not mine. That belongs to my rival, from another country, with another machine... Mine is white BMI2, because I've tested all versions with Build Test, a tool to test engines, and BMI2 build is the fastest version for my machine..
Anyway I'll try AVX2, but I'm not sure where obtain the Chessman compiles from
Thank you, Peter!

The Chessman compiles were posted earlier in this thread ...
http://outskirts.altervista.org/forum/viewtopic.php?p=45982&view=single_post#p45982

Marco's original release was
https://outskirts.altervista.org/forum/viewtopic.php?p=45951&view=single_post#p45951

It is worth trying one of the non BMI2 engines for comparison especially the AVX2 compile because that should run without issue on your Xeon CPU machine given its specification. The original StockfishMZ 230522 AVX2 Compile should also work too. If the same engine behaviour occurs with both AVX2 engines some further head scratching will be necessary.

Best regards,
Peter

Great, Peter,
I'll try AVX2 version and I'll be reportng the results

Thank you!!

[Peter Grayson]

Slow Ply depth issue recorded on dual Xeon machine.

The StockfishMZ 230522 engine issue experienced by catalano7777 has been reproduced on my dual Xeon E5-2687w machine.

The CPUs are AVX compatible but the closest engine in the MZ compiles was the sse41 engine. This was installed and appeared to run OK in command mode that pointed to it being the fastest engine of those that would run on the machine. However, when run in the Fritz 14 64 bit GUI and the Shredder GUI the engine indicated 8700 kN/s but ply depth increments were very slow indeed, similar to that recorded by catalano7777 and recorded in the attached CPU information, Engine configuration and output record file.

https://www.mediafire.com/file/28qtu2gt9i6vkkp/StockfishMZ_230522_sse41.zip/file

For comparison I installed the Chessman Stockfish 240522 AVX engine that matched the CPUs compatibility. That indicated 10800 kN/s and ply depths that looked consistent with that speed. That was also recorded in the file for comparison.

The Chessman sse41 engine was then installed and ran fine but slightly slower kN/s than the AVX compile and ply depths to match.

A further issue has been identified with the MZ and Chessman compile engines that is the same issue seen with the ShashChess 22 engine. When run in a 32 bit GUI, despite the configuration file recording 16 threads and the engine window also recording that number, the engine will only start using a single thread. That can be overcome by changing the number of threads and saving the new value from the engine window. The issue is present in the Fritz 8 through to Fritz 13 GUIs, Fritz 14 SE (32 bit) and the Shredder 10 GUI. Unfortunately it prevents the engine from being run in Tournament Mode or an Engine Match with more other than with 1 Thread.

With the Chessman compiles suffering the same issue it suggests the problem is with the UCI protocol code handling that is used for the engine initialisation.

I have not experienced any speed issues with the MZ BMI2 compile that has completed several matches and games without any noticeable issue. It may be nothing to do with the cause but both catalano7777 and my machine used here have Xeon CPUs fitted.

Peter

[Nemesis]

Bizarre .. to me anyway.
Zeus is my go-to for analysis - decided to try MZ (ChessMan compile).
Bf6 was not on my list of moves to analyze ??
This is just a for your information.

Incidentally .. 3950X .. 8 cores each .. W11 .. AVX2

[Peter Grayson]

''A further issue has been identified with the MZ and Chessman compile engines that is the same issue seen with the ShashChess 22 engine. When run in a 32 bit GUI, despite the configuration file recording 16 threads and the engine window also recording that number, the engine will only start using a single thread. That can be overcome by changing the number of threads and saving the new value from the engine window. The issue is present in the Fritz 8 through to Fritz 13 GUIs, Fritz 14 SE (32 bit) and the Shredder 10 GUI. Unfortunately it prevents the engine from being run in Tournament Mode or an Engine Match with more other than with 1 Thread.

With the Chessman compiles suffering the same issue it suggests the problem is with the UCI protocol code handling that is used for the engine initialisation.

I have not experienced any speed issues with the MZ BMI2 compile that has completed several matches and games without any noticeable issue. It may be nothing to do with the cause but both catalano7777 and my machine used here have Xeon CPUs fitted.''

Hi Peter,

from what you describe these are the symptoms of a GUI that is not updated.
For example I just made a quick test with MZ compile and had no problem.
I used Deep Fritz 13 GUI with last update (Deep_Fritz_13_Update_SP39_Setup) and it worked fine on my 11800H laptop.
When I installed Deep Fritz 13 after activation it said I was updated and that there where no updates.
Unfortunately the updates for older programs are not available anymore on the Chessbase server.
Fortunately I had saved the final update in an old hard drive and installed it manually.

In case you have already installed the final update and still have the same problem, then I'm out of suggestions.

I hope I helped!

Thanks Kerveros. I did the same as you, I saved the Deep Fritz 13 GUI update 39 and Deep Fritz 14 64 bit update 46 and installed them when I deactivated both from an older machine and installed them on the i7-10700 machine after I purchased it. The Fritz 14 update 46 is still available from Chessbase using the Activation\ Check for updates option.

The Stockfish engines run without issue in the older GUIs suggesting there is a difference in the ShashChess and MZ engine threads initialisaing code. Both the SugaR AI 2.40 and 2.50 engines initialise with the correct number of threads in the older GUIs that again hints at something being changed in the engine threads initialisation.

As for the difference in ply depths it was measured in the Deep Fritz 14 64 bit GUI and that issue was peculiar to the Stockfish 230522 MZ compile and was not repeated in the Chessman compile of the same code. Different compilers were used by MZ and Chessman but the issue was highlighted when running the engines in the Command window with "bench 512 16". The Chessman sse41 popcount compile inidcated 13450 kN/s whereas the MZ SSE41compile recorded 9460 kN/s, a 38% speed difference.

Peter