SFNNue 190720 available

[MMarco]

nodchip just released new binaries with some speedup apparently: https://github.com/nodchip/Stockfish/re ... 2020-07-19

Code: Select all

Changes:

Enabled profile build. Thank you, discord is a piece of XXXX.
Updated README.md. Thank you, xXH4CKST3RXx.
Changed the default eval file path so that more GUIs can use Stockfish+NNUE.
Added AVX-512 support. Thank you, mstembera.
Added SSE3+POPCNT support.
Load the parameter set on an `isready' as well. Thank you, discord is a piece of XXXX.

[zullil]

nodchip just released new binaries with some speedup apparently: https://github.com/nodchip/Stockfish/re ... 2020-07-19

Code: Select all

Changes:

Enabled profile build. Thank you, discord is a piece of XXXX.
Updated README.md. Thank you, xXH4CKST3RXx.
Changed the default eval file path so that more GUIs can use Stockfish+NNUE.
Added AVX-512 support. Thank you, mstembera.
Added SSE3+POPCNT support.
Load the parameter set on an `isready' as well. Thank you, discord is a piece of XXXX.

Note that these are Windows binaries only. But Linux users should be able to compile from source with:

make profile-nnue ARCH=x86-64

[MikeB]

nodchip just released new binaries with some speedup apparently: https://github.com/nodchip/Stockfish/re ... 2020-07-19

Code: Select all

Changes:

Enabled profile build. Thank you, discord is a piece of XXXX.
Updated README.md. Thank you, xXH4CKST3RXx.
Changed the default eval file path so that more GUIs can use Stockfish+NNUE.
Added AVX-512 support. Thank you, mstembera.
Added SSE3+POPCNT support.
Load the parameter set on an `isready' as well. Thank you, discord is a piece of XXXX.

I would be interested in other benchmarks,

$ stockfish
Stockfish+NNUE 190720 64 POPCNT by T. Romstad, M. Costalba, J. Kiiski, G. Linscott, H. Noda, Y. Nasu, M. Isozaki
ucinewgame
info string NNUE ./eval/nn.bin found & loaded
bench
info string Hash table allocation: Windows large pages used.
info string Hash table allocation: Windows large pages used.
quit
$ stockfish bench >/dev/null

===========================
Total time (ms) : 2042
Nodes searched : 3377227
Nodes/second : 1653881

For those that want to try my version - might be good for Ryzen CPUs
https://github.com/MichaelB7/Stockfish/ ... ckfish.exe

and here is a bin obtained from links on Talkchess that has tested very well against SF - place it in a folder name 'eval'
https://github.com/MichaelB7/Stockfish/ ... val/nn.bin

The classic ipman benchmark
$ stockfish bench 2048 64 26 >/dev/null

===========================
Total time (ms) : 61019
Nodes searched : 3907470324
Nodes/second : 64036944

[zullil]

I would be interested in other benchmarks,

$ stockfish
Stockfish+NNUE 190720 64 POPCNT by T. Romstad, M. Costalba, J. Kiiski, G. Linscott, H. Noda, Y. Nasu, M. Isozaki
ucinewgame
info string NNUE ./eval/nn.bin found & loaded
bench
info string Hash table allocation: Windows large pages used.
info string Hash table allocation: Windows large pages used.
quit
$ stockfish bench >/dev/null

===========================
Total time (ms) : 2042
Nodes searched : 3377227
Nodes/second : 1653881

We must be using the same nn.bin, since our node counts match:

===========================
Total time (ms) : 2735
Nodes searched : 3377227
Nodes/second : 1234817

But, since everyone has different CPU speeds, I'm more interested in how the Stockfish-nnue nps compares to the corresponding Stockfish-dev nps.

For Stockfish-dev, I get

===========================
Total time (ms) : 2218
Nodes searched : 4578298
Nodes/second : 2064155

So I see 1234817 / 2064155 ~ 60%

[MikeB]

I would be interested in other benchmarks,

$ stockfish
Stockfish+NNUE 190720 64 POPCNT by T. Romstad, M. Costalba, J. Kiiski, G. Linscott, H. Noda, Y. Nasu, M. Isozaki
ucinewgame
info string NNUE ./eval/nn.bin found & loaded
bench
info string Hash table allocation: Windows large pages used.
info string Hash table allocation: Windows large pages used.
quit
$ stockfish bench >/dev/null

===========================
Total time (ms) : 2042
Nodes searched : 3377227
Nodes/second : 1653881

We must be using the same nn.bin, since our node counts match:

===========================
Total time (ms) : 2735
Nodes searched : 3377227
Nodes/second : 1234817

But, since everyone has different CPU speeds, I'm more interested in how the Stockfish-nnue nps compares to the corresponding Stockfish-dev nps.

For Stockfish-dev, I get

===========================
Total time (ms) : 2218
Nodes searched : 4578298
Nodes/second : 2064155

So I see 1234817 / 2064155 ~ 60%

about 1653/2600 - so about 53%, I have to run

but I did successfully convert Honey and convert to NN engines

results are early but quite promising:

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   1 Honey-XI-NN                    65      36      92   59.2%   72.8%
   2 Stockfish-XI-NN                42      37      92   56.0%   72.8%
   3 Bluefish-XI-NN                -11      39      94   48.4%   69.1%
   4 cfish                         -20      47      88   47.2%   58.0%
   5 Black-Diamond-XI-NN           -75      44      94   39.4%   59.6%

They each have a different bench due to different search parameters

[peter]

but I did successfully convert Honey and convert to NN engines

results are early but quite promising:

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   1 Honey-XI-NN                    65      36      92   59.2%   72.8%
   2 Stockfish-XI-NN                42      37      92   56.0%   72.8%
   3 Bluefish-XI-NN                -11      39      94   48.4%   69.1%
   4 cfish                         -20      47      88   47.2%   58.0%
   5 Black-Diamond-XI-NN           -75      44      94   39.4%   59.6%

They each have a different bench due to different search parameters

[chrisw]

I would be interested in other benchmarks,

$ stockfish
Stockfish+NNUE 190720 64 POPCNT by T. Romstad, M. Costalba, J. Kiiski, G. Linscott, H. Noda, Y. Nasu, M. Isozaki
ucinewgame
info string NNUE ./eval/nn.bin found & loaded
bench
info string Hash table allocation: Windows large pages used.
info string Hash table allocation: Windows large pages used.
quit
$ stockfish bench >/dev/null

===========================
Total time (ms) : 2042
Nodes searched : 3377227
Nodes/second : 1653881

We must be using the same nn.bin, since our node counts match:

===========================
Total time (ms) : 2735
Nodes searched : 3377227
Nodes/second : 1234817

But, since everyone has different CPU speeds, I'm more interested in how the Stockfish-nnue nps compares to the corresponding Stockfish-dev nps.

For Stockfish-dev, I get

===========================
Total time (ms) : 2218
Nodes searched : 4578298
Nodes/second : 2064155

So I see 1234817 / 2064155 ~ 60%

about 1653/2600 - so about 53%, I have to run

but I did successfully convert Honey and convert to NN engines

results are early but quite promising:

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   1 Honey-XI-NN                    65      36      92   59.2%   72.8%
   2 Stockfish-XI-NN                42      37      92   56.0%   72.8%
   3 Bluefish-XI-NN                -11      39      94   48.4%   69.1%
   4 cfish                         -20      47      88   47.2%   58.0%
   5 Black-Diamond-XI-NN           -75      44      94   39.4%   59.6%

They each have a different bench due to different search parameters

Are the PGNs available for style-analysis?

[lkaufman]

Using this SFNNue 190720 together with jjoshua SFNN02 net I ran 150 games vs very recent (july 6) Stockfish, one thread, 3' + 2", 5 Ghz i7, Hert 250 lowdraw book. I got a +37 elo result for SFNNue.

[MikeB]

I would be interested in other benchmarks,

$ stockfish
Stockfish+NNUE 190720 64 POPCNT by T. Romstad, M. Costalba, J. Kiiski, G. Linscott, H. Noda, Y. Nasu, M. Isozaki
ucinewgame
info string NNUE ./eval/nn.bin found & loaded
bench
info string Hash table allocation: Windows large pages used.
info string Hash table allocation: Windows large pages used.
quit
$ stockfish bench >/dev/null

===========================
Total time (ms) : 2042
Nodes searched : 3377227
Nodes/second : 1653881

We must be using the same nn.bin, since our node counts match:

===========================
Total time (ms) : 2735
Nodes searched : 3377227
Nodes/second : 1234817

But, since everyone has different CPU speeds, I'm more interested in how the Stockfish-nnue nps compares to the corresponding Stockfish-dev nps.

For Stockfish-dev, I get

===========================
Total time (ms) : 2218
Nodes searched : 4578298
Nodes/second : 2064155

So I see 1234817 / 2064155 ~ 60%

about 1653/2600 - so about 53%, I have to run

but I did successfully convert Honey and convert to NN engines

results are early but quite promising:

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   1 Honey-XI-NN                    65      36      92   59.2%   72.8%
   2 Stockfish-XI-NN                42      37      92   56.0%   72.8%
   3 Bluefish-XI-NN                -11      39      94   48.4%   69.1%
   4 cfish                         -20      47      88   47.2%   58.0%
   5 Black-Diamond-XI-NN           -75      44      94   39.4%   59.6%

They each have a different bench due to different search parameters

Are the PGNs available for style-analysis?

Final Results

Code: Select all

ResultSet>rp C:\cluster.mfb\Popcnt-LP\071920.pgn
10000 game(s) loaded
ResultSet>elo
ResultSet-EloRating>mm 01
00:00:00,00
ResultSet-EloRating>r
Rank Name                 Rating         +    -                      W    L    D   W%    =%   OppR
---------------------------------------------------------------------------------------------------------
   1 Stockfish-XI-NN       3117   0.0    8    8  4000 2162.0  54.0  676  352 2972  16.9  74.3  3096
   2 Honey-XI-NN           3112   5.0    8    8  4000 2116.0  52.9  623  391 2986  15.6  74.7  3097
   3 Bluefish-XI-NN        3105   7.7    8    8  4000 2042.5  51.1  533  448 3019  13.3  75.5  3099
   4 cfish                 3087  17.1    8    8  4000 1884.5  47.1  595  826 2579  14.9  64.5  3103
   5 Black-Diamond-XI-NN   3078   9.1    8    8  4000 1795.0  44.9  380  790 2830   9.5  70.8  3105
---------------------------------------------------------------------------------------------------------

It is truly remarkable how quickly SF-NN advanced so far. Kudos to all those involved.

Chris -

Here is a pgn file you requested - link expires in 7 days. TC was 50 sec base + .5 sec inc. Contains all 10,000 games

https://www.dropbox.com/t/SQ5xCWN8oSJ3KGgF

[MMarco]

Hi Mike,

Does cfish here include all latest patches?

I thought that Bluefish eval was different than Stockfish (Edit: Was this right?). Does that mean that now, SF, Bluefish, Honey and Black-Diamond share the same eval and differ only by search?