Stockfish performing bad on ryzen cpu.

[awsmak]

I tried using Microsoft remote desktop app

you tell nothing about time controls (how it is organized) and books (which engine where and when start to play on its own and what book it uses before)

No opening books or endgame tablebase is used on both laptop and smartphone. Time control is set to 3seconds per move on both.

[awsmak]

Can you try this "exe" it was specifically compiled for AMD processors ...

https://www.dropbox.com/s/x4q40z21sx5on ... D.exe?dl=0
Note - binary is large 70MB - has NNUE and 55 EGTB's built in.
based on current Honey source:
https://github.com/MichaelB7/Stockfish/tree/honey

benches on a 3970x
current bench - classic mode 64 threads command:"Bluefish-12-W-AMD.exe b 2048 64 18 true >/dev/null" :
=================================
Total time (ms) : 41255
Nodes searched : 5387767058
Nodes/second : 130596k

NNUE mode 64 threads command "Bluefish-12-W-AMD.exe b 2048 64 18 true >/dev/null" :
=================================
Total time (ms) : 36933
Nodes searched : 4184508490
Nodes/second : 113299k

with 6 cores Classic mode:
"Bluefish-12-W-AMD.exe b 512 6 13 >/dev/null"
=================================
Total time (ms) : 4160
Nodes searched : 63355817
Nodes/second : 15229k

with 6 cores NNUE mode
Bluefish-12-W-AMD.exe b 512 6 13 true >/dev/null
=================================
Total time (ms) : 5122
Nodes searched : 68626044
Nodes/second : 13398k

I will surely check it out. I have tried some binaries that you referred on a different thread it never won but The Stockfish-Xlr5 always found draw. Let me check this out. Thanks for replying.

[yurikvelo]

Different compiles will give NPS variation in order of percents, which will give few ELO variation in strength.

Here we discuss enormous (in order of 200+ ELO) penalty, no matter which version was run on mobile.

I still do not understand how exactly tournament is run.
It is 0'+3" STC, run manually by 1 human operating two different interfaces - touch screen and keyboard/mouse

Both engines are run in "infinite analysis" mode and you move pieces as fast as you can? Which of 2 boards you move first?
All 200 games are from start position?

[awsmak]

Different compiles will give NPS variation in order of percents, which will give few ELO variation in strength.

Here we discuss enormous (in order of 200+ ELO) penalty, no matter which version was run on mobile.

I still do not understand how exactly tournament is run.
It is 0'+3" STC, run manually by 1 human operating two different interfaces - touch screen and keyboard/mouse

Both engines are run in "infinite analysis" mode and you move pieces as fast as you can? Which of 2 boards you move first?
All 200 games are from start position?

i ran it manually . engines are ran in time control 3 seconds per move. out of 200 the 100 are started with white by the laptop and 100 with black. every game was started from start position.

[Guenther]

Different compiles will give NPS variation in order of percents, which will give few ELO variation in strength.

Here we discuss enormous (in order of 200+ ELO) penalty, no matter which version was run on mobile.

I still do not understand how exactly tournament is run.
It is 0'+3" STC, run manually by 1 human operating two different interfaces - touch screen and keyboard/mouse

Both engines are run in "infinite analysis" mode and you move pieces as fast as you can? Which of 2 boards you move first?
All 200 games are from start position?

i ran it manually . engines are ran in time control 3 seconds per move. out of 200 the 100 are started with white by the laptop and 100 with black. every game was started from start position.

As long as you don't show a pgn file with depths/evals, which proves that your alleged 6 times faster ryzen(your measurement) 'loses' vs. your smartphone (against the same engine) no one can really help you.

[awsmak]

Can you try this "exe" it was specifically compiled for AMD processors ...

https://www.dropbox.com/s/x4q40z21sx5on ... D.exe?dl=0
Note - binary is large 70MB - has NNUE and 55 EGTB's built in.
based on current Honey source:
https://github.com/MichaelB7/Stockfish/tree/honey

benches on a 3970x
current bench - classic mode 64 threads command:"Bluefish-12-W-AMD.exe b 2048 64 18 true >/dev/null" :
=================================
Total time (ms) : 41255
Nodes searched : 5387767058
Nodes/second : 130596k

NNUE mode 64 threads command "Bluefish-12-W-AMD.exe b 2048 64 18 true >/dev/null" :
=================================
Total time (ms) : 36933
Nodes searched : 4184508490
Nodes/second : 113299k

with 6 cores Classic mode:
"Bluefish-12-W-AMD.exe b 512 6 13 >/dev/null"
=================================
Total time (ms) : 4160
Nodes searched : 63355817
Nodes/second : 15229k

with 6 cores NNUE mode
Bluefish-12-W-AMD.exe b 512 6 13 true >/dev/null
=================================
Total time (ms) : 5122
Nodes searched : 68626044
Nodes/second : 13398k

here are the some benches i tried
benches
1. bench 1024 16 18(classic)
=================================
Total time (ms) : 58783
Nodes searched : 1726863436
Nodes/second : 29376k

2. bench 1024 16 18(nn)
=================================
Total time (ms) : 83207
Nodes searched : 1845741017
Nodes/second : 22182k

3. bench 512 16 18(classic)
=================================
Total time (ms) : 54117
Nodes searched : 1595600521
Nodes/second : 29484k

4. bench 512 16 18(nn)
=================================
Total time (ms) : 77931
Nodes searched : 1720925110
Nodes/second : 22082k

5. bench 512 8 18(classic)
=================================
Total time (ms) : 43420
Nodes searched : 979584488
Nodes/second : 22560k

6. bench 512 8 18(nn)
=================================
Total time (ms) : 61939
Nodes searched : 1014954251
Nodes/second : 16386k

7. bench 512 6 13 ( classic )
=================================
Total time (ms) : 4051
Nodes searched : 60510698
Nodes/second : 14937k

8.bench 512 6 13(nnue)
=================================
Total time (ms) : 8240
Nodes searched : 100290774
Nodes/second : 12171k

I conducted a 10 rounds tournament at level '3 seconds per move' in arena with Bluefish-12-W-AMD.exe and stockfish_20090216_x64_avx2.exe both ran on 8 cores and 512 hash. bluefish lost 8 out of 10 games and 2 were draw.

[MikeB]

Can you try this "exe" it was specifically compiled for AMD processors ...

https://www.dropbox.com/s/x4q40z21sx5on ... D.exe?dl=0
Note - binary is large 70MB - has NNUE and 55 EGTB's built in.
based on current Honey source:
https://github.com/MichaelB7/Stockfish/tree/honey

benches on a 3970x
current bench - classic mode 64 threads command:"Bluefish-12-W-AMD.exe b 2048 64 18 true >/dev/null" :
=================================
Total time (ms) : 41255
Nodes searched : 5387767058
Nodes/second : 130596k

NNUE mode 64 threads command "Bluefish-12-W-AMD.exe b 2048 64 18 true >/dev/null" :
=================================
Total time (ms) : 36933
Nodes searched : 4184508490
Nodes/second : 113299k

with 6 cores Classic mode:
"Bluefish-12-W-AMD.exe b 512 6 13 >/dev/null"
=================================
Total time (ms) : 4160
Nodes searched : 63355817
Nodes/second : 15229k

with 6 cores NNUE mode
Bluefish-12-W-AMD.exe b 512 6 13 true >/dev/null
=================================
Total time (ms) : 5122
Nodes searched : 68626044
Nodes/second : 13398k

here are the some benches i tried
benches
1. bench 1024 16 18(classic)
=================================
Total time (ms) : 58783
Nodes searched : 1726863436
Nodes/second : 29376k

2. bench 1024 16 18(nn)
=================================
Total time (ms) : 83207
Nodes searched : 1845741017
Nodes/second : 22182k

3. bench 512 16 18(classic)
=================================
Total time (ms) : 54117
Nodes searched : 1595600521
Nodes/second : 29484k

4. bench 512 16 18(nn)
=================================
Total time (ms) : 77931
Nodes searched : 1720925110
Nodes/second : 22082k

5. bench 512 8 18(classic)
=================================
Total time (ms) : 43420
Nodes searched : 979584488
Nodes/second : 22560k

6. bench 512 8 18(nn)
=================================
Total time (ms) : 61939
Nodes searched : 1014954251
Nodes/second : 16386k

7. bench 512 6 13 ( classic )
=================================
Total time (ms) : 4051
Nodes searched : 60510698
Nodes/second : 14937k

8.bench 512 6 13(nnue)
=================================
Total time (ms) : 8240
Nodes searched : 100290774
Nodes/second : 12171k

I conducted a 10 rounds tournament at level '3 seconds per move' in arena with Bluefish-12-W-AMD.exe and stockfish_20090216_x64_avx2.exe both ran on 8 cores and 512 hash. bluefish lost 8 out of 10 games and 2 were draw.

WOW those are some results - care to provide the PGN file?

How long did it take to run that match and what machine were you using ( how many cores) , and how many games were being run in concurrency?

Edit: Thanks - Seems I like I just broke something, Your short test and my short test confirms it.

[awsmak]

I was just checking the Honey12 version from your link
https://github.com/MichaelB7/Stockfish/tree/honey
It always end in draw. conducted 10 matches against my smartphone and all of them were draws. 9mnps VS 1.5mnps the one with greater nps and depth should win atleast once right?
Settings on laptop were 8cores, 512mb hash table, nn used all other settings kept default.
Level 3 seconds/move, ponder on

[mwyoung]

Why your Droidfish is running < 1.6 mnps, are you using 1 core? 4 cores droidfish should be around 5 mnps.

Stockfish 11 gives 5000-6000knps stockfish 12 gives only 1500knps using 6 cores

All I know is you have some kind of issue. There is no way Sf 12 should be getting only 1.5 mnps on 6 cores. I get almost 1 mnps on 1 core at the start position.

You are using a laptop, and it may not have good thermals. And AVX creates much more heat and stress on the CPU. Try running the other version of SF 12.

[awsmak]

Why your Droidfish is running < 1.6 mnps, are you using 1 core? 4 cores droidfish should be around 5 mnps.

Stockfish 11 gives 5000-6000knps stockfish 12 gives only 1500knps using 6 cores

All I know is you have some kind of issue. The is no way Sf 12 should be getting only 1.5 mnps on 6 cores. I get almost 1 mnps on 1 core at the start position.

You are using a laptop, and it may not have good thermals. And AVX creates much more heat and stress on the CPU. Try running the other version of SF 12.

on my laptop on 8 cores sf12 gives 9+mnps. my smartphone on 6 cores sf12 gives 1.5-2mnps. but my laptop is always loosing or drawing against my smartphone.