Lc0 node count and playing strength

[Hamster]

With Stockfish or similar CPU engines it is easy to see how "deep" an engine has calculated, either by looking at the search depth or by looking at the total nodes calculated. In practice the search depth is the better indicator as the number is nicer and makes more sense for humans. One could day that a "normal" position for OTC play is well analyzed if Stockfish reached depth 30 (often depth 25 is sufficient for non-GM and non-CC).

How to go about it with lc0? I have a very slow GPU (NVIDIA GeForce GT 610) and using a network with 24 blocks and 320 filters I get 17 nodes/seconds.
How many nodes (or seconds) shall I search to get a "decent" depth?

[mbabigian]

Some tests were done showing ELO gain per node count doubling and posted on the LC0 discord channel. Basically the ELO graph showed rapid gains up to about 128000 nodes and then it flattens out. With your ultra slow hardware, I'd try to reach at least 128000 for strongish moves. Due to the search design of LC0, it becomes exceedingly difficult for it to change its mind even after it finds a huge flaw in its analysis. This means that after realizing a catastrophic problem in the PV it could take more time than is on the clock to switch to an alternative move. This makes super long thinks ineffective unless the visit count between the top 2 or few moves happen to be very close.

Hope that helps,
Mike

[Nay Lin Tun]

With Stockfish or similar CPU engines it is easy to see how "deep" an engine has calculated, either by looking at the search depth or by looking at the total nodes calculated. In practice the search depth is the better indicator as the number is nicer and makes more sense for humans. One could day that a "normal" position for OTC play is well analyzed if Stockfish reached depth 30 (often depth 25 is sufficient for non-GM and non-CC).

How to go about it with lc0? I have a very slow GPU (NVIDIA GeForce GT 610) and using a network with 24 blocks and 320 filters I get 17 nodes/seconds.
How many nodes (or seconds) shall I search to get a "decent" depth?

17 nps is seriously low. My 1060 GTX has 2k nps in T60 nets.

[Hai]

With Stockfish or similar CPU engines it is easy to see how "deep" an engine has calculated, either by looking at the search depth or by looking at the total nodes calculated. In practice the search depth is the better indicator as the number is nicer and makes more sense for humans. One could day that a "normal" position for OTC play is well analyzed if Stockfish reached depth 30 (often depth 25 is sufficient for non-GM and non-CC).

How to go about it with lc0? I have a very slow GPU (NVIDIA GeForce GT 610) and using a network with 24 blocks and 320 filters I get 17 nodes/seconds.
How many nodes (or seconds) shall I search to get a "decent" depth?

17 nps is seriously low. My 1060 GTX has 2k nps in T60 nets.

2k is seriously low.
I have most of the time 5-7k with 40x512 nets.

[Ovyron]

How to go about it with lc0? I have a very slow GPU (NVIDIA GeForce GT 610) and using a network with 24 blocks and 320 filters I get 17 nodes/seconds.

Have you tried CPU Leela? I've been having problems running her properly as she seems to hate my computer, but even I'm getting 24 nodes/second.

My take on this: You require at least some 11111 nodes to get decent analysis (as decent as Stockfish depth 30). The good news is that after reaching them for the first time, Leela will reach them much quicker for incoming positions.

With just those nodes I was able to predict all of Leela's moves coming from 800 Million nodes on mmt's GPU on our 1.g4 game, though those were for the sharpest positions in chess and possibly losing, it seems you should be able to extrapolate what much higher node counts would show by interacting with the positions (not unlike using low depth to extrapolate what Stockfish at Depth 60 would show.)

[Nay Lin Tun]

With Stockfish or similar CPU engines it is easy to see how "deep" an engine has calculated, either by looking at the search depth or by looking at the total nodes calculated. In practice the search depth is the better indicator as the number is nicer and makes more sense for humans. One could day that a "normal" position for OTC play is well analyzed if Stockfish reached depth 30 (often depth 25 is sufficient for non-GM and non-CC).

How to go about it with lc0? I have a very slow GPU (NVIDIA GeForce GT 610) and using a network with 24 blocks and 320 filters I get 17 nodes/seconds.
How many nodes (or seconds) shall I search to get a "decent" depth?

17 nps is seriously low. My 1060 GTX has 2k nps in T60 nets.

2k is seriously low.
I have most of the time 5-7k with 40x512 nets.

I am just comparing non RTX old cards speeds.

[Nay Lin Tun]

My estimation is , OP should get around 80-100 nps in his hardware.

1. The processing power = 3x slower
2. No cuda support = 8x slower

2000÷24= 83 nps.