Leela CPU becomes Top 5 engine

[Nay Lin Tun]

As smaller net T59 10x128 become stronger and stronger, Lc0 without graphic cards will also be a top 5 engine.


T59 Leela should be on par or slightly stronger than Fire and Ethereal, that make Leela on top 5 engine behind Stockfish , Komodo and Houdini.

https://ibb.co/7YxHGhv

[mwyoung]

Leela CPU becomes Top 5 engine, or who needs a GPU!



Hardware 2950x, RTX 2080 ti

DESKTOP-CORSAIR, Blitz 0m+5s Ponder Off. Book Perfect 2019 to 8 moves.

Komodo 13.2
4Gb HT
32 threads
Default

Lc0 23.2 591066 - Custom
CS = 2000000
CPuct = 3.0
Policy Temperature = 1.8

[Ovyron]

As smaller net T59 10x128 become stronger and stronger

Where can I get that net?

[Nay Lin Tun]

You can take the same network as chess.com or the latest network with these settings.
https://ibb.co/hyHtHMJ
https://training.lczero.org/networks/2

[Ovyron]

So d23bb62f I guess, since it has the highest ELO?

[Alayan]

Leela CPU is incredibly slow, so its scaling curve with time/hardware means that while terrible at 1c bullet, it gets better and better with more time and cores and starts plateauing quite late.

On the other hand, T59 on 4xRTX2080Ti is badly CPU-limited, and not much stronger than Leela CPU despite a significant nps advantage, as past a certain npm the elo gain becomes pathetic.

Leela CPU will outperform Fire or Ethereal if you let it run long enough on enough hardware, so for analysis and suggesting alternatives it can be nice, but it will lose below some time/hardware threshold.

[Thomas Lagershausen]

Exist this new Leela CPU for android devices?

If yes please post a link.

[Laskos]

Leela CPU is incredibly slow, so its scaling curve with time/hardware means that while terrible at 1c bullet, it gets better and better with more time and cores and starts plateauing quite late.

On the other hand, T59 on 4xRTX2080Ti is badly CPU-limited, and not much stronger than Leela CPU despite a significant nps advantage, as past a certain npm the elo gain becomes pathetic.

Leela CPU will outperform Fire or Ethereal if you let it run long enough on enough hardware, so for analysis and suggesting alternatives it can be nice, but it will lose below some time/hardware threshold.

"T59 on 4xRTX2080Ti is badly CPU-limited" is a complete overkill. Even 1xRTX2080Ti is CPU-limited with T59 at some 100k nps in the openings/midgame. Do the Lc0 devs try to address the 100k nps CPU limitation? It is still the same for at least one year. Are there some tricks with batch sizes and collision events to mitigate the limitation in nps by the quality of nodes?

The best strength-wise use to LTC for 4x2080Ti would be with a late T60 net or with even larger SV distilled nets. Maybe some 70-80k nps in the openings/midgame with T60 nets, and maybe 30-40k nps with SV big nets.

[mwyoung]

Leela CPU is incredibly slow, so its scaling curve with time/hardware means that while terrible at 1c bullet, it gets better and better with more time and cores and starts plateauing quite late.

On the other hand, T59 on 4xRTX2080Ti is badly CPU-limited, and not much stronger than Leela CPU despite a significant nps advantage, as past a certain npm the elo gain becomes pathetic.

Leela CPU will outperform Fire or Ethereal if you let it run long enough on enough hardware, so for analysis and suggesting alternatives it can be nice, but it will lose below some time/hardware threshold.

On the other hand it is impressive that at 250 nps on 1 thread. Lc0 cpu is torching a near 3000 CCRL elo rated program Fritz 13. At 5s a move.

We use to say, wow deep blue searches 200,000,000 nps and can beat Kasparov. Lc0 at 250 nps will crush any human, and many A/B engines.

[dkappe]

...that make Leela on top 5 engine behind Stockfish , Komodo and Houdini.

In CCC’s own championship both Komodo and Houdini finished outside the top 4.