Gambit Rating List - May 20, 2021

[Rebel]

40/2 - http://rebel13.nl/grl-40-2.html
40/15 - http://rebel13.nl/grl-40-15.html

Blue : New Entries
Green : Version History or Engine Elo Progress

Comments
1. The new Stockfish-20-05-18 with a 45Mb network didn't make it to the top.

2. The elo difference between SF13 and SF12 = 51 elo. On CCRL (40/15) its 31 elo, if SF did not have to play against its derivatives its rating would be higher. On CCRL (40/2) the situation is even worse, for reasons unknown to me SF13 is rated lower than SF12.

3. Apparently something got wrong with Minic 3.06 (major regression), I will test the new 3.07 as first thing to do.

[xr_a_y]

You may have used Minic 3.06 without a net. We can check the log and make some tests if needed. In 3.07, the net is embeded, so you won't have this possible issue anymore ...

[Rebel]

I am pretty sure I did. But I am glad you embedded the net. I have Minic 3.07 running, 1100 games. You can watch the results at :

http://rebel13.nl/a/grl.htm

Page automatically refreshed every 30 seconds.

[xr_a_y]

What is the hardware used here and which executable is used ?
Minic using NNUE is really not good is AVX2 is not present, even worst than HCE on very old hardware.
An old test shows

Code: Select all

Rank Name                                Elo     +/-   Games   Score    Draw 
   1 minic_3.03_linux_x64_skylake_niny    56      11    2454   58.0%   39.2% 
   2 minic_3.03_linux_x64_nehalem_niny    13      11    2454   51.9%   40.9% 
   3 minic_3.03_linux_x64_nehalem         10      11    2455   51.5%   39.7% 
   4 minic_3.03_linux_x64_skylake          4      11    2455   50.6%   41.4% 
   5 minic_3.03_linux_x64_core2          -36      11    2456   44.9%   39.7% 
   6 minic_3.03_linux_x64_core2_niny     -48      11    2456   43.1%   37.3% 

[Rebel]

I am always using the same type of executable wherever possible to avoid inconsistencies, in your case always Nehalem.

Code: Select all

Results from file gauntlet-minic-307.pgn:

No. Name           Win Draw Loss Unf.  Score Games       %
----------------------------------------------------------
  1 Minic 3.07    +403 =368 -329   *0  587.0  1100   53.4%
  2 Seer 2.0.1     +46  =35  -19   *0   63.5   100   63.5%
  3 Halogen 10     +45  =36  -19   *0   63.0   100   63.0%
  4 Stash 29.0     +36  =30  -34   *0   51.0   100   51.0%
  5 Marvin 5.0     +29  =43  -28   *0   50.5   100   50.5%
  6 Danasah 8.8    +30  =38  -32   *0   49.0   100   49.0%
  7 Berserk 4.0.0  +29  =36  -35   *0   47.0   100   47.0%
  8 Orion 0.8      +23  =41  -36   *0   43.5   100   43.5%
  9 Topple 0.8.0   +29  =27  -44   *0   42.5   100   42.5%
 10 Tucano 9.0     +26  =26  -48   *0   39.0   100   39.0%
 11 Counter 3.7    +20  =29  -51   *0   34.5   100   34.5%
 12 Mr Bob_1.0.0   +16  =27  -57   *0   29.5   100   29.5%

Total Games:    1100
White Wins:      359 (32.6%)
Black Wins:      373 (33.9%)
Draws:           368 (33.5%)
Unfinished:        0 (0.0%)

Estimated elo gain for Minic_3.07 
Minic 3.04 : 3109.5
Minic_3.07  : 3078.5
Difference : -31.0

+23 to version 3.06

[xr_a_y]

Ok thanks. For sure Nehalem build is far from the best for NNUE evaluation so if your hardware allows to use a better one please do.
For instance Seer has no nehalem build, so you used ivybridge or skylake one (the core2 one would be very very slow), so I deduce that the sandybridge build of Minic can probably be used on your hardware it seems ?

I'd love to try your opening suite at home, is it available somewhere please ?

[connor_mcmonigle]

Ok thanks. For sure Nehalem build is far from the best for NNUE evaluation so if your hardware allows to use a better one please do.
For instance Seer has no nehalem build, so you used ivybridge or skylake one (the core2 one would be very very slow), so I deduce that the sandybridge build of Minic can probably be used on your hardware it seems ?

I'd love to try your opening suite at home, is it available somewhere please ?

The opening book might be relevant here, but I have Minic 3.07 at +33 elo in one of my tests to Seer v2.0.1 on a relatively high draw rate book. It seems a clear improvement to 3.04 which was maybe 40 elo behind Seer v2.0.1 under similar conditions.

[Rebel]

Ok thanks. For sure Nehalem build is far from the best for NNUE evaluation so if your hardware allows to use a better one please do.
For instance Seer has no nehalem build, so you used ivybridge or skylake one (the core2 one would be very very slow), so I deduce that the sandybridge build of Minic can probably be used on your hardware it seems ?

Realize that every NNUE engine has the same problem on a non AVX2 Pc.

I'd love to try your opening suite at home, is it available somewhere please ?

http://rebel13.nl/gambits.pgn

[xr_a_y]

Realize that every NNUE engine has the same problem on a non AVX2 Pc.

Sure, but if you use for instance a SSE4.2 (nehalem) build of one engine and a AVX (ivybridge) build for another one, results might look weird.

http://rebel13.nl/gambits.pgn

Thanks a lot ! This will be helpfull.

[xr_a_y]

Ok thanks. For sure Nehalem build is far from the best for NNUE evaluation so if your hardware allows to use a better one please do.
For instance Seer has no nehalem build, so you used ivybridge or skylake one (the core2 one would be very very slow), so I deduce that the sandybridge build of Minic can probably be used on your hardware it seems ?

I'd love to try your opening suite at home, is it available somewhere please ?

The opening book might be relevant here, but I have Minic 3.07 at +33 elo in one of my tests to Seer v2.0.1 on a relatively high draw rate book. It seems a clear improvement to 3.04 which was maybe 40 elo behind Seer v2.0.1 under similar conditions.

I see this also
Minic 3.04 < Seer v2.0.1
Minic 3.06 ~ Seer v2.0.1
Minic 3.07 > Seer v2.0.1