Ideal Syzygy Probe Depth ? (using SSD)

[shrapnel]

Hi
I wonder if anyone here could tell me what is the ideal syzygy probe depth for Stockfish and Houdini 4 ? I have all the Syzygy EGTB on a 360 GB Corsair SSD.
I know the recommendation is to lower it for SSD users, but then I start to get THOUSANDS of TB hits, eventually slowing down my Engine ! If I raise it too much, then very few TB hits.
(Are they useful for online engine-engine matches at all ?? I have my doubts. )
Any ideas ?

[syzygy]

but then I start to get THOUSANDS of TB hits, eventually slowing down my Engine !

If TB probing considerably lowers nps, then raise probe depth. If nps stays fine, then do not raise it. Provided nps stays fine, thousands or millions of probes is not a problem (on the contrary, the more the better).

What percentage of nps decrease is still acceptable I do not know.

Btw, it seems that in endgame positions (with just a few pieces left) nps of SF when running on many cores anyway is rather low even without TB usage (one would expect nps to almost double compared to the opening, but in e.g. TCEC on 16 cores it decreased quite significantly).

[shrapnel]

OK, thank you.
I suppose the solution is to see what is the average nps I get with EGTB probing turned off, and then accordingly set the probe depth by Trial and Error method to the ideal one which has least drop in nps, while at the same time, getting a reasonable number of TB hits.

[Ryan Benitez]

(Are they useful for online engine-engine matches at all ?? I have my doubts. )

For engine vs engine matches don't expect any real gain but they might make endgame more elegant. I would just use them at root position. Best case is to just have the GUI handle EGTBs.

[Laskos]

(Are they useful for online engine-engine matches at all ?? I have my doubts. )

For engine vs engine matches don't expect any real gain but they might make endgame more elegant. I would just use them at root position. Best case is to just have the GUI handle EGTBs.

On one core, Syzygy, Scorpio and Shredderbases all give 15 +/- 7 2SD points bonus in blitz. That's with 5 men on HDD.

[Houdini]

(Are they useful for online engine-engine matches at all ?? I have my doubts. )

For Houdini 4 I've measured 5 to 10 Elo improvement with 6-men Syzygy in engine vs engine games at fast TC (30 second games), using the default settings and with the table bases on a normal hard disk.
The gain should increase with longer TC, and with the files on a SSD.

Under the same conditions I've never been able to demonstrate a clear Elo gain with other table base solutions (Nalimov and Gaviota), which is why we recommend Syzygy for use with Houdini.

[Vinvin]

Hi
I wonder if anyone here could tell me what is the ideal syzygy probe depth for Stockfish and Houdini 4 ? I have all the Syzygy EGTB on a 360 GB Corsair SSD.
I know the recommendation is to lower it for SSD users, but then I start to get THOUSANDS of TB hits, eventually slowing down my Engine ! If I raise it too much, then very few TB hits.
(Are they useful for online engine-engine matches at all ?? I have my doubts. )
Any ideas ?

There's no good value for probe depth. In some position you can get 1000 access/sec on depth 25 and on others positions you can get 100 access/sec on depth 30.

Depth is not relevant for TB access, only number of access/second have to be limited.

[RJN]

(Are they useful for online engine-engine matches at all ?? I have my doubts. )

For Houdini 4 I've measured 5 to 10 Elo improvement with 6-men Syzygy in engine vs engine games at fast TC (30 second games), using the default settings and with the table bases on a normal hard disk.
The gain should increase with longer TC, and with the files on a SSD.

Under the same conditions I've never been able to demonstrate a clear Elo gain with other table base solutions (Nalimov and Gaviota), which is why we recommend Syzygy for use with Houdini.

With SSD, I don't measure more than a few % slowdown of H4 on average, when using Syzygy probe depth as low as 2 (the minimum), with 6-piece.

Also, I wonder at what point would would TB hits with better quality knowledge offset a slowdown in NPS of the tree with its many lower quality positions being evaluated? What's a ballpark figure, any opinions? 5%, 10%, 20%?

[syzygy]

There's no good value for probe depth. In some position you can get 1000 access/sec on depth 25 and on others positions you can get 100 access/sec on depth 30.

Depth is not relevant for TB access, only number of access/second have to be limited.

I cannot agree.

1. There is no reason in itself that the number of accesses per second should be limited. On the contrary, the more the better. In my view this is self-evident. However, IF (note: IF) nps goes way down, then it might be good to increase the probe depth in order to lessen the impact on nps.

2. Depth is very relevant for TB access. Probes with high remaining depth will on average be more useful than probes with low remaining depth (but the latter still are useful).

Increasing probe depth limits probing to nodes with higher remaining depth. If in endgame positions with relatively few remaining pieces this still results in a decrease of nps, this might be worth it, because the probes that cost nps may still have a decisive impact on the quality of the search.

So:
- in middlegame positions far away from 6-piece positions it is fine to have TB probes, as long as they don't affect nps.
- in endgame positions close to 6-piece positions, a decrease in nps is acceptable.
- the optimum probe depth is dependent on hardware and not so much on the type of position: if despite high probe depth there are still a lot of accesses in an endgame position that are affecting nps, those are probably worth it. It is more important to make sure that nps does not go down too much in positions where tablebases can hardly be of use (i.e. positions with very many pieces).