Super bullet chess: LC0 network 35932 vs newest Stockfish dev.

[Hai]

Super bullet chess: LC0 network 35932 vs newest Stockfish dev.

The 20 possible opening moves
40 games
Ponder on
LC0 with RTX 2080 Ti 32 kn/s
Stockfish with 4 cores @ 4.2 ghz
Time control: 1 min per game + 1 second per move.

Result:
LC0 19 points
Stockfish 21 points
+6 =28 -8
Winning percentage = 47.62%

= LC0 is 17 elo weaker than Stockfish.

http://www.mediafire.com/file/kpydwtgd5 ... 9.pgn/file

32390 is 154 elo stronger than 35932 viewtopic.php?f=6&t=69503

[Hai]

Can we train for a few days only t30 networks to get the strongest one into the superfinal vs Stockfish?
http://testserver.lczero.org/networks/
http://testserver.lczero.org/
It is still improving and 13000 selfplay elo are possible to reach.

Or why we are running 99% training games for the t35 networks?
t35 will never be ready in time for the superfinal.

And I don't see any reason why we should lose the superfinal if we can win it by using all power for the t30 networks.

[Werewolf]

Super bullet chess: LC0 network 35932 vs newest Stockfish dev.

The 20 possible opening moves
40 games
Ponder on
LC0 with RTX 2080 Ti 32 kn/s
Stockfish with 4 cores @ 4.2 ghz
Time control: 1 min per game + 1 second per move.

Result:
LC0 19 points
Stockfish 21 points
+6 =28 -8
Winning percentage = 47.62%

= LC0 is 17 elo weaker than Stockfish.

According to the Leela ratio your GPU is about 4x faster than your CPU. To make the match more balanced SF would need 4x as much time with ponder off. Lc0 still has a way to go.

[MonteCarlo]

Super bullet chess: LC0 network 35932 vs newest Stockfish dev.

The 20 possible opening moves
40 games
Ponder on
LC0 with RTX 2080 Ti 32 kn/s
Stockfish with 4 cores @ 4.2 ghz
Time control: 1 min per game + 1 second per move.

Result:
LC0 19 points
Stockfish 21 points
+6 =28 -8
Winning percentage = 47.62%

= LC0 is 17 elo weaker than Stockfish.

According to the Leela ratio your GPU is about 4x faster than your CPU. To make the match more balanced SF would need 4x as much time with ponder off. Lc0 still has a way to go.

Those "desired" ratios are based on the larger, slower 20 block networks.

This test was with the smaller, faster 10 block network from test35. How much speedup the smaller net gets varies a bit, but the hardware is probably about right to get the "normal" (a bit weird that the ratio defined in the A0 paper is simply taken as normal, but I digress) ratio with the 20 block network.

All this assumes that the stated speed is the speed with the 10 block, and not the 20 block, of course. If that's speed with the 20 block, then you are of course correct that this is a very favorable hardware setup relative to what DeepMind used.

Cheers!