Steroid Leela (2x faster leela) will be playing TCEC 20 Sufi

[Nay Lin Tun]

This New Leela version (partial implementaion of Ceres idea ) will be playing in this Sufi!

S20 DivP game 188 original and replay:
moveno: DivP -> PR1483
8: 97976 -> 168736 +72%
9: 90965 -> 194676 +114%
10: 89835 -> 196105 +118%
11: 84191 -> 94452 +12%
12: 64423 -> 91580 +42%
13: 71462 -> 109162 +53%
14: 58383 -> 74005 +27%
15: 54888 -> 80172 +46%
16: 51251 -> 64629 +26%
17: 45737 -> 91620 +100%
18: 46406 -> 109427 +136%
19: 46435 -> 114499 +147%
20: 47080 -> 113862 +142%
21: 46438 -> 119358 +157%
22: 47095 -> 121823 +159%
23: 42571 -> 139637 +228%
24: 32561 -> 150161 +361%
25: 42568 -> 152421 +258%
26: 49700 -> 150044 +202%
27: 50835 -> 151096 +197%
28: 56339 -> 142850 +154%
29: 53202 -> 115259 +117%
30: 55850 -> 151795 +172%
31: 56406 -> 168804 +199%
32: 72960 -> 176068 +141%
33: 30407 -> 143665 +372%
34: 33008 -> 165158 +400%
35: 42120 -> 186246 +342%
36: 81295 -> 192384 +137%
37: 90779 -> 136479 +50%
38: 82527 -> 176389 +114%
39: 88814 -> 204311 +130%
40: 64670 -> 157910 +144%
41: 73636 -> 129925 +76%
42: 87663 -> 156076 +78%
43: 93506 -> 193646 +107%
44: 77348 -> 154174 +99%
45: 90982 -> 190812 +110%
46: 94618 -> 187786 +98%
47: 87852 -> 184555 +110%

As you can see, the speed gain in some positions will be up to 400% and minimum of 12% .
Average is at least 200% faster!

[Dann Corbit]

Does it play stronger?
The reason I ask is that Ceres did not do well in Pohl's testing.

[Jouni]

Found this info:
Regarding our submission to SuFi:
• Given that it's unlikely, that Lc0 in the current configuration would win SuFi (don't let DivP fool you), it's quite likely that we'll send pr1483 binary there.
• PR 1483 is a port of one of Ceres features (parallel node gathering), implemented by Tilps.
• It gives ~+40% nps increase, at the cost of worse nodes selected (5 Elo loss at fixed nodes).
• It's also quite a tricky asynchronous code, so hang/crash-causing bugs are not unlikely -- it was tested reasonably well but only on a few configuration -- prepare for a drama.

[Nay Lin Tun]

Does it play stronger?
The reason I ask is that Ceres did not do well in Pohl's testing.

This is not Ceres, just Leela.
https://github.com/LeelaChessZero/lc0/pull/1483
I dont know which built Pohl used. And there is no official release of Ceres yet.
Crem's fixed node test shows around 5 elo loss.

[Nay Lin Tun]

Found this info:
Regarding our submission to SuFi:
• Given that it's unlikely, that Lc0 in the current configuration would win SuFi (don't let DivP fool you), it's quite likely that we'll send pr1483 binary there.
• PR 1483 is a port of one of Ceres features (parallel node gathering), implemented by Tilps.
• It gives ~+40% nps increase, at the cost of worse nodes selected (5 Elo loss at fixed nodes).
• It's also quite a tricky asynchronous code, so hang/crash-causing bugs are not unlikely -- it was tested reasonably well but only on a few configuration -- prepare for a drama.

40% speed increase info was based on " Nav's machine" .
200% speed is based on actual TCEC machine by Aloril test.( TCEC gets more benefits from this version because there is severe cpu bottle neck with 2.5 GHZ machine)

[Dann Corbit]

Does it play stronger?
The reason I ask is that Ceres did not do well in Pohl's testing.

This is not Ceres, just Leela.
https://github.com/LeelaChessZero/lc0/pull/1483
I dont know which built Pohl used. And there is no official release of Ceres yet.
Crem's fixed node test shows around 5 elo loss.

I think Pohl used the "all in one" clump someone made available.
It seems logical that re-coded as C++ it should do much better.
I am keen to see the result.

[Nay Lin Tun]

Version 27 beta is released.
https://github.com/LeelaChessZero/lc0/releases/

[Jouni]

The download for cudann is 270 MB, but only lc0.exe with 3 MB is changed . I got promising results even with GTX card:

Code: Select all

                           
1   Lc0 v0.27.0-rc1   +18  +20/=82/-14 52.59%   61.0/116
2   Stockfish 12      -18  +14/=82/-20 47.41%   55.0/116

SF was classic, but 4 cores.

[Ozymandias]

What were your results with 0.26.3 like?

[Jouni]

About 50% for Lc0. Different nets have no big effect. I used J94-80.