Elo difference for a 5 or 6-men syzygy

[Ferdy]

Is there any data for Stockfish NNUE with vs without syzygy?

[Dann Corbit]

Other tablebase systems did not deliver any Elo (e.g. Nalimov) result of a long contest was 0 Elo.)
The Syzygy tablebase files with 5 men gave 50 Elo. They are not very big and easily cached in total.
I doubt if a study has been made with 6 man, but for sure you need fast disk if you have a really fast machine.
I found that I was only getting 10% CPU throughput for some positions when I had my EGTB on mechanical disk.
A fast SSD improved that a lot.
A PCIE 4.0 gumstick SSD blew that out of the water (almost no slowdown).
So I guess that if you use PCIE 4.0 SSD with 6 man files and you have at least 64 GB fast RAM and 64 threads, then 6 men Syzygy tablebase files will add a lot but I did not measure it.

So my conclusion is 5 man +50 Elo, 6 man + ??? Elo (unless someone has carefully measured it now)
I also guess that Nalimov would also add Elo on gumstick SSD. I have mine configured that way, but I did not bother to measure it.

[Guenther]

Other tablebase systems did not deliver any Elo (e.g. Nalimov) result of a long contest was 0 Elo.)
The Syzygy tablebase files with 5 men gave 50 Elo.

...

I very much doubt this. Do you have a source for this?
IMHO in real game play over a sufficient number of games (no cherry picking - no endgame start positions)
it should not be more than may be 5 elo for an already very strong program.

[Jouni]

I just finished 5000 games syzygy test! SF12 has 5 piece and most important 6 tables on SSD. 8 moves book and 10+0,1 level. Result:

SF12 syzygy won SF12 no syzygy 2535,5/5000 = +4,9 ELO

With longer time control even less? Similar to one good tuning patch .

[Ferdy]

I got a small sample of games at tc=120s+1s on non-ssd disk. The slowdown is obvious.

Code: Select all

Result:
------------------------------------------------------------------------------
  #  name              games    wins   draws  losses   score    los%  elo+/-
  1. Stockfish 12        100       4      93       3    50.5    64.7     3.5
  2. Stockfish 12 ET6    100       3      93       4    49.5    35.3    -3.5

Cross table:
------------------------------------------------------------------------------
  #  name                 score   games         1         2
  1. Stockfish 12          50.5     100         x      50.5
  2. Stockfish 12 ET6      49.5     100      49.5         x

Tech:
------------------------------------------------------------------------------

Tech (average nodes, depths, time/m per move, others per game), counted for computing moves only, ignored moves with zero nodes:
  #  name                nodes/m         NPS  depth/m   time/m    moves     time
  1. Stockfish 12          2138K      811411     25.0      2.6     51.6    136.0
  2. Stockfish 12 ET6      1645K      606050     23.0      2.7     51.6    140.1
     all ---               1847K      707223     24.0      2.7     51.6    138.0

Banksia GUI is great.

[Ferdy]

In a 7-men ending test with single best move, Sf12 without ET performed better.

[Nordlandia]

I was told earlier that 5-men yield about +5 Elo while 6-men add about 10. 7-men is estimated to be around 15-20 elo. If i recall correctly.

[Dann Corbit]

Other tablebase systems did not deliver any Elo (e.g. Nalimov) result of a long contest was 0 Elo.)
The Syzygy tablebase files with 5 men gave 50 Elo.

...

I very much doubt this. Do you have a source for this?
IMHO in real game play over a sufficient number of games (no cherry picking - no endgame start positions)
it should not be more than may be 5 elo for an already very strong program.

No, you were right. It was 5.0 not 50.0

[Terje]

When Weiss was something like 2700 elo 6-men tables on a very fast SSD gave ~15-20 elo. Weiss had (and still has) very little chess knowledge so the gain might be smaller in engines with better evaluation, or just stronger engines in general.

[Engin]

Other tablebase systems did not deliver any Elo (e.g. Nalimov) result of a long contest was 0 Elo.)
The Syzygy tablebase files with 5 men gave 50 Elo. They are not very big and easily cached in total.
I doubt if a study has been made with 6 man, but for sure you need fast disk if you have a really fast machine.
I found that I was only getting 10% CPU throughput for some positions when I had my EGTB on mechanical disk.
A fast SSD improved that a lot.
A PCIE 4.0 gumstick SSD blew that out of the water (almost no slowdown).
So I guess that if you use PCIE 4.0 SSD with 6 man files and you have at least 64 GB fast RAM and 64 threads, then 6 men Syzygy tablebase files will add a lot but I did not measure it.

So my conclusion is 5 man +50 Elo, 6 man + ??? Elo (unless someone has carefully measured it now)
I also guess that Nalimov would also add Elo on gumstick SSD. I have mine configured that way, but I did not bother to measure it.

That is the reason now i removed Tornado nalimov probe because it's gain nothing even it's slow down the search and hanging sometimes with many threads running. so disable or remove is better then. On the other hand the NNUE eval is also detect winning, draw or losing endgames so what we need tablebases, just for 1 or 5 elo gain ?