Leela Chess Zero Makes Mind-Blowing Piece Sacrifice!

[Milos]

I have only 128 GB RAM, but I have allocated a 2 TB swap file on super high speed SSD. LC0 has no problem analyzing overnight using two 2080 TI super cards.
I think it would be a great choice for correspondence analysis, especially during the opening phase.
Now, there is analysis time control longer than correspondence (I have analyzed a position for well over a week before). And perhaps LC0 would be a bad choice for that, but for most purposes you don't need to be a millionaire to use it for analysis.

That's because you use only the dumbest push button and forget type of analysis. I can spend 12 hours of compute time and get 10x better analysis than you leaving SF or Lc0 analysing something for a week. Sure, it requires some interaction and know-how, but then again, it's a matter of quality not quantity.

Additionally, I don't even wanna go there, but you don't seem to know much about CC. Even in early opening, when you are analyzing deep, most of analysis is about avoiding the traps and certain types of endgame, because most of your lines will be deep enough to reach there. And using Lc0 there would be a suicide. That's another reason why push analysis button and forget is pretty dumb and useless approach in CC.

Yes, all those TCEC wins are purely lucky.
Only a fool would consider LC0 engine analysis.
You already told us that the IM who liked the bishop sacrifice knows nothing, and I am sure you could give him some nice chess lessons.

Blah, blah. You have a very powerful hardware. So go ahead put your money where your mouth is and give us the winning line.
And lol about IM comment. Someone that is 2200 Elo and uses push button analysis on 3300 Elo chess engine, vs someone who is 0 OTB Elo, and adds 200 Elo to 3300 Elo engine with proper analysis knowledge is certainly comparable. I think not. Maybe you think chess Elo is an additive quantity so if you use both SF and Lc0 you end up with 6600 Elo analysis .

[Milos]

Meanwhile, Stoofvlees is giving Fire a bludgeoning in the current game like an angry Canadian beating the brains out of a baby seal.
Current standings have three programs above Stockfish, all of which are NN.

Another irrelevant comment. Take equal hardware (moneywise not stupid invented Leela ratio) so in your case (since you have both) SF on 3990X (with limited Brainfish book) vs Lc0 on 2x2080Ti and tell us the outcome (irrelevant of TC).

[Dann Corbit]

Since the Elo scale is arbitrary, and you can draw any baseline you like, I do hereby proclaim my Elo to be 10,000.
Yet my incredible modesty prevents me from using that phenomenal cosmic power, in all its beauty, splendor and glory.

I think that sometimes you have some very interesting opinions.
And other times, for instance in dismissal of LC0, they are way out of left field.
But that is OK, because there is room in the world for serious people and crackpots and everything in between.
I can be a little of both

[Dann Corbit]

Meanwhile, Stoofvlees is giving Fire a bludgeoning in the current game like an angry Canadian beating the brains out of a baby seal.
Current standings have three programs above Stockfish, all of which are NN.

Another irrelevant comment. Take equal hardware (moneywise not stupid invented Leela ratio) so in your case (since you have both) SF on 3990X (with limited Brainfish book) vs Lc0 on 2x2080Ti and tell us the outcome (irrelevant of TC).

Um...
Compare the price of 4 GPUs (about $4K) with 176 core CPU systems like TCEC.
In the words of Willy Wonka, "Strike that, reverse it"

The compute per dollar is much higher with GPUs than with CPUs. Everyone knows that

[Milos]

Meanwhile, Stoofvlees is giving Fire a bludgeoning in the current game like an angry Canadian beating the brains out of a baby seal.
Current standings have three programs above Stockfish, all of which are NN.

Another irrelevant comment. Take equal hardware (moneywise not stupid invented Leela ratio) so in your case (since you have both) SF on 3990X (with limited Brainfish book) vs Lc0 on 2x2080Ti and tell us the outcome (irrelevant of TC).

Um...
Compare the price of 4 GPUs (about $4K) with 176 core CPU systems like TCEC.
In the words of Willy Wonka, "Strike that, reverse it"

The compute per dollar is much higher with GPUs than with CPUs. Everyone knows that

Another topic change. Now jumping to TCEC. Again, do you have 3990X and 2x2080Ti system? It's easy to check. Just play the games .

[jp]

Compare the price of 4 GPUs (about $4K) with 176 core CPU systems like TCEC.

What are the reported nodes per second TCEC gets for Leela on those 4 GPUs and SF on those 176 cores?

[Dann Corbit]

Compare the price of 4 GPUs (about $4K) with 176 core CPU systems like TCEC.

What are the reported nodes per second TCEC gets for Leela on those 4 GPUs and SF on those 176 cores?

You do realize, that the NPS figures for AB and NN programs have nothing to do with each other.

But, currently, with a very large network, LC0 is getting 39K NPS.
Contrast that with Komodo, which is getting 150M NPS
(So 3846 times "faster" for K verses LC0).
Yet, it seems that those ponderous, laborious nodes are putting Komodo into difficulty.

Of course, this game is hardly decided yet. But make sure what your are measuring means something.
In this case, the measurement that matters is Elo.

[OliverBr]

LC0 is better at quiet positions than stockfish.
So for instance, for an opening, the analysis of LC0 is more reliable than Stockfish.
For tactical positions and endgame, Stockfish is much stronger.
That is why they are of equal strength. When there are no obvious or tactical moves, LC0 knows what to do.
When a tactical situation arises, or when we reach the endgame and SF is hitting 80 plies, SF knows what to do.

If SF were absurdly better at analysis than LC0, then how would it be possible for LC0 to win major tournaments?

This is a precise and correct summary. Thank you, @Dann!

[OliverBr]

Analysis is not just press a button and forget (even though you might be doing it that way).
Lc0 doesn't have hash (its cashed moves are in no way replacement for hash just for faster running) and that is a total killer in any analysis (and it's not only Lc0, MCTS engines are not useful for analysis, period). Since you can't go forward and backward and get any useful result of it.

And if you do just press a button and forget, Lc0 needs huge amounts of memory or it will run out and abort. Not everyone has 512GB RAM (but good luck to those who do!).

Hö? What?
Leela, compiled by me with CUDA libraries, runs reliably on a 32GB RAM System and doesn't use any significant part of it.

[OliverBr]

@Milos
you haven't answered to my question, if you use a version of Leela that uses CUDA?! You don't know it?