Unlimited, stronger game analysis on lichess.org

That's why you see 200x for Stockfish 11 HCE. The "value" of nodes computed by pre-NNUE engines are much weaker.
That 200x factor is still a placeholder value (more accurate than 1x but still a guess). Only Stockfish 19 1MB has been truly measured vs Fishnet at this point - it needs roughly 7x more nodes.
General advice - if your browser won't let you run Stockfish 19 94MB, stick to timed analysis.

It doesn't need to reach rating of SF19, it needs to be set to a reasonable value. Most players that use SF11+local analysis are people who don't like annotations based on SF19 calculations. Also, how strong does engine doing evaluations need to be on default? Is 3000 enough? 3200?

If adding an option to choose depth is lot of work, may I suggest setting it to 2M or 2,5M nodes per move (for SF11)

p.s. Also I doubt there is a noticeable difference from 10Mn to 200Mn per move. I'm not an expert, but some graph convinced me that SF19 has GM strength on 25Kn, and is near it owns peak around 25 Mn per move.

@gamblej said [^](/forum/redirect/post/muoqr3Ag) >That's why you see 200x for Stockfish 11 HCE. The "value" of nodes computed by pre-NNUE engines are much weaker. >That 200x factor is still a placeholder value (more accurate than 1x but still a guess). Only Stockfish 19 1MB has been truly measured vs Fishnet at this point - it needs roughly 7x more nodes. >General advice - if your browser won't let you run Stockfish 19 94MB, stick to timed analysis. It doesn't need to reach rating of SF19, it needs to be set to a reasonable value. Most players that use SF11+local analysis are people who don't like annotations based on SF19 calculations. Also, how strong does engine doing evaluations need to be on default? Is 3000 enough? 3200? If adding an option to choose depth is lot of work, may I suggest setting it to 2M or 2,5M nodes per move (for SF11) p.s. Also I doubt there is a noticeable difference from 10Mn to 200Mn per move. I'm not an expert, but some graph convinced me that SF19 has GM strength on 25Kn, and is near it owns peak around 25 Mn per move.

gamblej edited

It doesn't need to reach rating of SF19, it needs to be set to a reasonable value.

I may just remove it for weaker engines and leave them with timed evals.

You don't want depth on modern engines. And I don't want to make different UIs for every engine.

If you had a depth slider for Stockfish 19, positions would be a box of chocolates that could either take half a second or half an hour.

> It doesn't need to reach rating of SF19, it needs to be set to a reasonable value. I may just remove it for weaker engines and leave them with timed evals. You don't want depth on modern engines. And I don't want to make different UIs for every engine. If you had a depth slider for Stockfish 19, positions would be a box of chocolates that could either take half a second or half an hour.

PirateOfAdria edited

I found Stockfish 11 (beta) strength scaling per nodes per move. They are very close to how i remembered them (in my last post).

It reaches GM level around 25.000 nodes per move, scaling slows down around 4 million and falls bellow 100 elo per doubling of nodes. Finally it hits ceiling somewhere between 8 and 16 million, so 200m is well into placebo territory (graph stops at 67m).

Graph is made with FastGM data from testing "Stockfish 190509". SF11 was published Jan2020 so this data is probably SF 11 beta from May2019.

https://talkchess.com/viewtopic.php?t=72834
https://imgur.com/w79nMgc

I am aware that Stockfish 11 beta and Lichess in-browser implementation of SF11 may have different characteristics. I'm just "lobbying" for reconsideration of changing the 'standard' instead of removing it. Also, why not name it X million nodes, instead of "standard".

I found Stockfish 11 (beta) strength scaling per nodes per move. They are very close to how i remembered them (in my last post). It reaches GM level around 25.000 nodes per move, scaling slows down around 4 million and falls bellow 100 elo per doubling of nodes. Finally it hits ceiling somewhere between ***8 and 16 million***, so 200m is well into ***placebo*** territory (graph stops at 67m). Graph is made with FastGM data from testing "Stockfish 190509". SF11 was published Jan2020 so this data is probably SF 11 beta from May2019. https://talkchess.com/viewtopic.php?t=72834 https://imgur.com/w79nMgc I am aware that Stockfish 11 beta and Lichess in-browser implementation of SF11 may have different characteristics. I'm just "lobbying" for reconsideration of changing the 'standard' instead of removing it. Also, why not name it ***X million nodes***, instead of "standard".

gamblej

Also, why not name it X million nodes, instead of "standard".

Good question. For small searches like this (around 1,000,000 nodes), Stockfish introduces inefficiency by partitioning the search space across multiple threads. So a 32 thread search would actually need 2,000,000 nodes to achieve strength parity with a 1 thread search at 1,000,000 nodes, all other things being equal. And this has been measured.

This tax on parallelism tends towards zero on longer timed searches, but it's still significant below 5,000,000 nodes.

Short answer, "Standard" is the best I could do to convey "roughly the same quality you'd get from server analysis".

> Also, why not name it X million nodes, instead of "standard". Good question. For small searches like this (around 1,000,000 nodes), Stockfish introduces inefficiency by partitioning the search space across multiple threads. So a 32 thread search would actually need 2,000,000 nodes to achieve strength parity with a 1 thread search at 1,000,000 nodes, all other things being equal. And this has been measured. This tax on parallelism tends towards zero on longer timed searches, but it's still significant below 5,000,000 nodes. Short answer, "Standard" is the best I could do to convey "roughly the same quality you'd get from server analysis".