How much Rtx4080 help lc0

[Uri Blass]

I do not use lc0 but a friend asked me about the importance of Rtx4080 for lc0.
His plans are to buy Raptor Lake iG of new intel.

He claims that people who do not understand about cuda cannot find a good reply in google.

[Uri Blass]

He also asked me the following question

These are the best ners for lc0
https://lczero.org/dev/wiki/best-nets-for-lc0/

The question is how you define the best net and from what gpu

[Jouni]

You must find best lc0.exe, best net and best settings. And still You occasionally blunder and lose to Stockfish .

[smatovic]

Maybe this threads helps a bit:

"Using LC0 with one or two GPUs - a guide"
https://talkchess.com/forum3/viewtopic. ... 62#p794580

I would prefer an used RTX 2080 TI for ~500 bucks on eBay over new RTX 3080 or even RTX 4080, cos it has better FP32 to FP16 ratio on hardware level...

from: https://en.wikipedia.org/wiki/List_of_N ... sing_units

RTX 2080 TI: 11.750 TFLOPS FP32 and 23.500 TFLOPS FP16
RTX 3080: 25.06 TFLOPS FP32 and FP16
RTX 4090: 43 TFLOPS FP32 and FP16

In theory an doubling of processing power (FLOPS) should give an doubling of NPS with the same net.

In the past there were FP16 precision and FP32 precision options in the backend, dunno about current backend versions and nets and if they are optimized.

Smaller nets give more NPS, are faster, but contain less knowledge, you might want to try different nets for different use cases, like tournaments or analysing or less powerful hardware.

--
Srdja

[smatovic]

...

typo:

RTX 4080: 43.0 TFLOPS FP32 and FP16
RTX 4090: 73.1 TFLOPS FP32 and FP16

--
Srdja

[mehmet123]

Lc0 at Nvidia 4090 is 2.5x faster than Lc0 at Nvidia 3080,
Lc0 at Nvidia 4090 is 4x faster than Lc0 at Nvidia 3060 ti.


RTX 3080 : 13518 (T78) 18779 (T80)
RTX 4090 : 33612 (T78) 45966(T80)

[Magnum]

Maybe this threads helps a bit:

"Using LC0 with one or two GPUs - a guide"
https://talkchess.com/forum3/viewtopic. ... 62#p794580

I would prefer an used RTX 2080 TI for ~500 bucks on eBay over new RTX 3080 or even RTX 4080, cos it has better FP32 to FP16 ratio on hardware level...

from: https://en.wikipedia.org/wiki/List_of_N ... sing_units

RTX 2080 TI: 11.750 TFLOPS FP32 and 23.500 TFLOPS FP16
RTX 3080: 25.06 TFLOPS FP32 and FP16
RTX 4090: 43 TFLOPS FP32 and FP16

In theory an doubling of processing power (FLOPS) should give an doubling of NPS with the same net.

In the past there were FP16 precision and FP32 precision options in the backend, dunno about current backend versions and nets and if they are optimized.

Smaller nets give more NPS, are faster, but contain less knowledge, you might want to try different nets for different use cases, like tournaments or analysing or less powerful hardware.

--
Srdja

Buy 2x RTX 4090 Ti or 2x Titan RTX and be happy forever.
Or buy 4x to be on the safe side.
Or 7x.

[Uri Blass]

You must find best lc0.exe, best net and best settings. And still You occasionally blunder and lose to Stockfish .

The target is not to beat stockfish but get better analysis and lc0 may suggest different ideas that are not the same as stockfish.
There are positions when lc0 is better than stockfish and it did not lose all the matches against lc0 in TCEC.

[Werewolf]

You must find best lc0.exe, best net and best settings. And still You occasionally blunder and lose to Stockfish .

The target is not to beat stockfish but get better analysis and lc0 may suggest different ideas that are not the same as stockfish.
There are positions when lc0 is better than stockfish and it did not lose all the matches against lc0 in TCEC.

If the analysis is for OTB human games I’d say Lc0 is better than SF, partly because its search doesn’t assume optimal play, so it’s more trappy.

As an experiment I took a load of chess computers from the 1980s-1990s (max elo 2200) and made them play all the modern engines.
A very clear trend emerged regarding game length…

[Modern Times]

Lc0 at Nvidia 4090 is 4x faster than Lc0 at Nvidia 3060 ti.

Incredible. And the 3060 ti is no slouch.