6-men Syzygy from HDD and USB 3.0

[Laskos]

I finally downloaded 6-men Syzygy, performed CRC, put them on HDD, and WDL tables additionally on a 128 GB USB 3.0 stick. I played several thousands games to have a feel about them with both Stockfish and Komodo, but here I will present only Komodo 10.4 results, as Syzygy is more flexible there ("Smart Syzygy" option to probe 5-men from depth=2 irrespective of 6-men Probe Depth). Playing games is the only option to determine strength here, as for different Probe Depths, the search tree is different. What I found is that if keeping all (DTZ+WDL) tables on HDD, 6-men (with many different Probe Depth values) don't give improvement over 5-men on USB, oftentimes performing even worse for small Probe Depths. Both give about 10-15 ELO points over non-Syzygy (for Probe Depth=24 from HDD) from endgame starting positions. However, with 6-men WDL tables on USB stick (DTZ on HDD), I got 35-40 ELO points improvement from Komodo with 6-men (Probe Depth=12) over non-Syzygy Komodo from these endgame starting positions. Here is some example run (gauntlet against non-Syzygy). It is also worth noting NPS and Depth reached as shown by LittleBlitzer:

Code: Select all

Games Completed = 3000 of 3000 (Avg game length = 16.234 sec)
Settings = Gauntlet/32MB/200ms per move/M 600000cp for 1000 moves, D 120000 moves/EPD:C:\LittleBlitzer\End_789_13_17.epd(897)
Time = 13265 sec elapsed, 0 sec remaining
 1.  K NO Syzygy               1418.5/3000	736-899-1365    (L: m=899 t=0 i=0 a=0)		(tpm=199.9 d=24.84 nps=2736335)
 2.  K Syzygy-6 USB WDL PD=12   545.5/1000	325-234-441  	(L: m=234 t=0 i=0 a=0)		(tpm=189.0 d=24.50 nps=2321309)
 3.  K Syzygy-6 HDD WDL PD=24   513.5/1000	279-252-469  	(L: m=252 t=0 i=0 a=0)		(tpm=197.4 d=23.23 nps=1880441)
 4.  K Syzygy-5 USB PD=2        522.5/1000	295-250-455  	(L: m=250 t=0 i=0 a=0)		(tpm=193.6 d=24.99 nps=2573741)

"Smart Syzygy" was always "true" with 3-4-5-6-men tables. I would imagine that from SSD, the 6-men gains might be even more spectacular. With 128 GB SSD, it is probably advisable to put WDL on them. Also, with 6-men, my i7 PC (8 GB RAM) became almost unresponsive during some matches, due to RAM caching.

[Laskos]

Another interesting test. From 6-men endgame positions, hard wins, I tested Stockfish dev and Komodo 10.4 without tablebases against 6-men Syzygy enabled Stockfish, to see how many are converted by non-TB top engines.

1s/move

Stockfish dev: +34 =66 -0
Converts 34/100, loses 0/100 won positions.

Komodo 10.4: +26 =72 -2
Converts 26/100, loses 2/100 won positions.

Without tablebases, even the top engines have hard time with hard 6-men wins.

[Jouni]

Interesting results, thanks! BTW my test shows, that real ELO gain is about 20% of gains in endings (1/5 games is decided in endings) - so below 10 ELO from 6 piece then?

[Laskos]

Interesting results, thanks! BTW my test shows, that real ELO gain is about 20% of gains in endings (1/5 games is decided in endings) - so below 10 ELO from 6 piece then?

If your 20% stands, close to 10 Elo points, right. But from entire set on SSD, could be 15 or so Elo points from regular opening positions.

[Nordlandia]

Probing tablebases from SSD via USB 3.0 causes latency right?

SATA III is better than USB 3.0 for tablebases.

SATA or PCI-E is better for more TBHITS.

[Laskos]

Probing tablebases from SSD via USB 3.0 causes latency right?

SATA III is better than USB 3.0 for tablebases.

SATA or PCI-E is better for more TBHITS.

I was not saying one has to put SSD via USB 3.0. I simply haven't yet bought SSD storage, so cannot play with those.

[Jouni]

Are your 6-men endgame positions, hard wins from kirill-kryukov.com/chess/longest.../longest-checkmates.shtml? I have done some tests with them. Are you sure, that positions are really won with 50 move rule? I noticed, that even 5 piece positions with 5 piece syzygy ended sometimes as draw! 50 move draw from settings disabled (in syzy) didn't help.

[Nordlandia]

Houdini 5.01 performs better in the famous Spassky vs Fischer endgame position if EGTB Probe Depth is adjusted.

I tried -10 and -5 on my brand new Samsung 960 EVO.

If syzygy is stored on M.2 SSD or equivalent, default Probe Depth is probably not correct setting for H5.1.

[Dann Corbit]

Houdini 5.01 performs better in the famous Spassky vs Fischer endgame position if EGTB Probe Depth is adjusted.

I tried -10 and -5 on my brand new Samsung 960 EVO.

If syzygy is stored on M.2 SSD or equivalent, default Probe Depth is probably not correct setting for H5.1.

What depth worked best?
I have a big SSD also.

[Nordlandia]

Houdini 5.01 performs better in the famous Spassky vs Fischer endgame position if EGTB Probe Depth is adjusted.

I tried -10 and -5 on my brand new Samsung 960 EVO.

If syzygy is stored on M.2 SSD or equivalent, default Probe Depth is probably not correct setting for H5.1.

What depth worked best?
I have a big SSD also.

It's wrong to draw conclusions bases on few blitz games, but in my case -5 do seem reasonable.

Houdart argues that he don't like the speed penalty for probing tablebases, therby intelligent probing technique in Houdini 5.

For fast m.2 ssd, sweetspot is definitely between 0 and -10 somewhere.