Lc0

[jmartus]

lc0 with a nvidia 4090 vs stockfish on a i7-12700K 12 cores who wins

[Krzysztof Grzelak]

it all depends on the many factors you use in the match.

[Werewolf]

lc0 with a nvidia 4090 vs stockfish on a i7-12700K 12 cores who wins

Probably SF. But for certain analysis, I'd prefer Lc0 if it's not about finding tactics.

[mehmet123]

lc0 with a nvidia 4090 vs stockfish on a i7-12700K 12 cores who wins

Lc0 at Nvidia 4090 is 2.5x faster than Lc0 at Nvidia 3080. At tournament time control Lc0 beats latest Stockfish Dev without difficulty.

[Krzysztof Grzelak]

Lc0 at Nvidia 4090 is 2.5x faster than Lc0 at Nvidia 3080. At tournament time control Lc0 beats latest Stockfish Dev without difficulty.

You're writing some good shit.

[mehmet123]

lc0 with a nvidia 4090 vs stockfish on a i7-12700K 12 cores who wins

Lc0 at Nvidia 4090 is 2.5x faster than Lc0 at Nvidia 3080. At tournament time control Lc0 beats latest Stockfish Dev without difficulty.

Lc0 at Nvidia 4090 is 4x faster than Lc0 at Nvidia 3060 ti. The power of Lc0 611213 at Nvidia 3060 ti isn't weaker than Stockfish 14 ( 8 cores) at tournament time control.
https://ssdf.bosjo.net/list.htm

Lc0 784xxx/785xxx is ~+20 elo stronger than Lc0 611xxx at blitz / rapid time control. i7-12700K is ~2x faster than Ryzen 7 1800X. 2x speed is a little more effective at Lc0 according to Stockfish. At long time control elo difference between Stockfish 15(4 cpu) and Stockfish 14(4 cpu) is + 20 elo at Cegt 40/20 and 0 elo at CCRL (40/15). As the playing time increases the elo power between Stockfish 15 and Stockfish 14 is decreases.
In summary, ı think we need minimum Intel Core i9-13900K (24cores /32thread) to beat Lc0 at Nvidia 4090 at rapid / long time control match.

[Jouni]

CSS forum test SF 16 cores against Lc0 2xA100(!):

1 stockfish : 0.0 1076.0 / 2000
2 lc0-T78 : -16.2 480.0 / 1000
3 lc0-BT1024 : -45.5 444.0 / 1000

[mehmet123]

CSS forum test SF 16 cores against Lc0 2xA100(!):

1 stockfish : 0.0 1076.0 / 2000
2 lc0-T78 : -16.2 480.0 / 1000
3 lc0-BT1024 : -45.5 444.0 / 1000

I think that this result can only be achieved in ultrablitz/blitz time control games, this result is impossible for me in a rapid / tournament time control match.

[mehmet123]

I found this match at Lc0 Discord. Lc0 plays at Nvidia RTX 4090 and Dragon 3.1 plays at a strong cpu (32 cores/ 64 threads)

windfishballadatch: 784968 and 806322 vs Komodo Dragon
LC0 version: dag-master 9f9d9490
Network: t80=806322, t78=784968
Komodo Dragon version: 3.1
LC0 options:--minibatch-size=192, different tuned search params for t78 and t80 (see attachment)
Komodo Dragon options: Threads=64, Hash=65536, Table Memory=1024
Hardware: 3990X 3.6 GHz (processes bound to 32 physical cores), 2x RTX 4090
Software: Cutechess GUI
Time control: 5 min + 2.5s
Book: TCEC SuFi 23
Tablebase: 6 men Syzygy
Adjudication: -draw movenumber=40 movecount=3 score=10 -resign movecount=2 score=700
Ordo: Book has some black advantage exits so results are pre-processed before running Ordo (logic: if sum of first exit evals in the two games of a pair is <0, results are mirrored). -a 0 -A dragon3.1
Comment: These conditions are very close to TCEC Nps for both engines

# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%)
1 lc0-t80 : 86.0 41.5 60.0 100 60 98
2 lc0-t78 : 25.6 41.7 53.0 100 53 89
3 Dragon3.1 : 0.0 ---- 87.0 200 44 ---


In this test, the number of games are too low and the games are blitz time match, but it still gives us some information about the power of Lc0 at Nvidia RTX 4090.

[Modern Times]

Depending on the test conditions I'd say probably Lc0.