Smallnet (128x10) run1 progresses remarkably well

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Dann Corbit, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
Laskos
Posts: 10948
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: Smallnet (128x10) run1 progresses remarkably well

Post by Laskos » Thu Jan 03, 2019 4:41 am

After some stalling and rating inflation, test35 is the strongest of all at ultra-bullet, and only some 50 Elo points worse than test30 and test10 at bullet. It scales badly (partly due probably to NPS CPU bottleneck), but if test40 will start with bootstraped test35 nets, it will start at very high level, scaling well.

yanquis1972
Posts: 1766
Joined: Tue Jun 02, 2009 10:14 pm

Re: Smallnet (128x10) run1 progresses remarkably well

Post by yanquis1972 » Thu Jan 03, 2019 5:36 am

So for small nets is it a good idea to increase thread count or is that insignificant (or completely irrelevant)? I’m only typically getting ~30kns, fwiw

User avatar
Laskos
Posts: 10948
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: Smallnet (128x10) run1 progresses remarkably well

Post by Laskos » Thu Jan 03, 2019 7:01 am

yanquis1972 wrote:
Thu Jan 03, 2019 5:36 am
So for small nets is it a good idea to increase thread count or is that insignificant (or completely irrelevant)? I’m only typically getting ~30kns, fwiw
No, increasing thread count is no solution. It becomes a problem with an excess of some 90knps, so you should be fine. But smaller nets usually scale worse with TC.

yanquis1972
Posts: 1766
Joined: Tue Jun 02, 2009 10:14 pm

Re: Smallnet (128x10) run1 progresses remarkably well

Post by yanquis1972 » Fri Jan 04, 2019 6:03 pm

despite self-elo declining i think the network's continuing to strengthen; first round of 64 games against SF8 @ 5'3 were tied, second Lc0 was +4 wins, & currently:

1: Lc0 test35 37.0/66 ·······················
2: Stockfish 9 10.0/22 ===10=======01=1000===
3: Stockfish 8 1 core 9.5/22 =====0====0====10===0=?
3: Stockfish 10 1cpu 9.5/22 ===100=0=110==0100====

66 of 96 games played
Level: Blitz 5/3
lczero16start.pgn
i7@4.7ghz (all SFs on 1 core)
gtx1080

rolling nets, tried to update once per round when possible. the id running overnight (just updated) was 35773, which more or less marked the beginning of the latest decline/adjustment in self-elo, so i'll see how the remaining games go.

CCRL 40/40 has SF10 at +80 elo over SF8, so that score in particular is really surprising to me.

this was a record of won games as of many hours ago https://lichess.org/study/SDDYR2Y1 -- t35 seems remarkably good in the endgame, especially compared to previous nets. its winpercentage will hit 90% before the opponent eval is -1.00. (i think, though i haven' t checked, 95% = -9.00cp). >75% seems to be the point of no return, but that may've decreased -- again, still need to check how that translates to cp.

yanquis1972
Posts: 1766
Joined: Tue Jun 02, 2009 10:14 pm

Re: Smallnet (128x10) run1 progresses remarkably well

Post by yanquis1972 » Fri Jan 04, 2019 7:40 pm

can't examine it myself atm but the last round of king's indian games had an interesting result (+5 =1 -0): https://lichess.org/study/Bl1fGB3U

first win as white in the KGA --



it's becoming clear that SF needs the restrictor plate taken off

Post Reply