Lc0 + 320x24.J13B.2-swa-136000 vs. Stockfish 250919 TC= 30m+30s

[zullil]

Are you sure that 4 threads is better that 2 threads for Leela?

2 is the default and on my system it use 1.5 core.

Yes I am sure. You see the results. You can see the CPU and GPU load on my testing stream. The question you need to ask is why others are not tuning Lc0 to their rig for best performance. I guess because then do not know how, or don't want to take the time.

I have seen some very bad setups for Lc0.

I will tell you. If you are setting Lc0 for maximum NPS. You are on the wrong path. NPS with Lc0 is a false flag.

How did you determine that 4 worker threads are better than 2 for a single GPU system?

https://github.com/LeelaChessZero/lc0/b ... r/FLAGS.md

viewtopic.php?t=70035

https://groups.google.com/forum/#!searc ... IVlc2hBQAJ

[mwyoung]

The question you need to ask is why others are not tuning Lc0 to their rig for best performance. I guess because then do not know how, or don't want to take the time.

I have seen some very bad setups for Lc0.

I will tell you. If you are setting Lc0 for maximum NPS. You are on the wrong path. NPS with Lc0 is a false flag.

Hi
If NPS with Lc0 is a false flag, which agrees with my experience on Infinity Chess, what exactly are you Tuning it to and how ?

Good question!

I tune it by time to depth and GPU loading. I ignore NPS results. And test this over many positions for 5 mins for each position, and you need to use end games positions also, not just middle game positions to achieve the best results when tuning.

And then test against your other setups for improvement.

Each computer and GPU combo will have different settings.

The wild card setting is your cput setting. Tune this to what you think works best. My setting is 3.0, but some insist that others are better like 3.4. And this is dependent on your GPU speed and testing time controls. But tune it to your cput setting....

[shrapnel]

Interesting Idea !
Thank You

[George Tsavdaris]

I will tell you. If you are setting Lc0 for maximum NPS. You are on the wrong path. NPS with Lc0 is a false flag.
I tune it by time to depth and GPU loading. I ignore NPS results.

Tuning Leela according to depth is a very bad idea and it's pointless and dangerous. Dangerous that you will end up with a weaker Leela probably.
Depth is not a good metric at all for MCTS type programs like it is with AB type programs.
N/s is the only easy and sensible metric for performance.
As more N/s, is always better. For Leela that is.

So what you are doing so far with your tests, probably has a somewhat weaker Leela as you didn't try to achieve higher N/s.

[mwyoung]

I will tell you. If you are setting Lc0 for maximum NPS. You are on the wrong path. NPS with Lc0 is a false flag.
I tune it by time to depth and GPU loading. I ignore NPS results.

Tuning Leela according to depth is a very bad idea and it's pointless and dangerous. Dangerous that you will end up with a weaker Leela probably.
Depth is not a good metric at all for MCTS type programs like it is with AB type programs.
N/s is the only easy and sensible metric for performance.
As more N/s, is always better.

So what you are doing so far with your tests, probably has a somewhat weaker Leela as you didn't try to achieve higher N/s.

Then don't use GPU loading and TIME TO DEPTH. My results speak for themselves.

You act like lc0 is new, and we have not tried the NPS tuning. Like I said NPS is a false flag.

[George Tsavdaris]

I will tell you. If you are setting Lc0 for maximum NPS. You are on the wrong path. NPS with Lc0 is a false flag.
I tune it by time to depth and GPU loading. I ignore NPS results.

Tuning Leela according to depth is a very bad idea and it's pointless and dangerous. Dangerous that you will end up with a weaker Leela probably.
Depth is not a good metric at all for MCTS type programs like it is with AB type programs.
N/s is the only easy and sensible metric for performance.
As more N/s, is always better.

So what you are doing so far with your tests, probably has a somewhat weaker Leela as you didn't try to achieve higher N/s.

Then don't use GPU loading and TIME TO DEPTH. My results speak for themselves.

You act like lc0 is new, and we have not tried the NPS tuning. Like I said NPS is a false flag.

How exactly do they speak if you have nothing to compare?
I bet that using properly set up parameters for obtaining higher N/s would be better for Leela.
No need to assume things anyway, as depth is completely nonsensical metric for Leela. Optimizing for it is useless as i've said.

[mwyoung]

I will tell you. If you are setting Lc0 for maximum NPS. You are on the wrong path. NPS with Lc0 is a false flag.
I tune it by time to depth and GPU loading. I ignore NPS results.

Tuning Leela according to depth is a very bad idea and it's pointless and dangerous. Dangerous that you will end up with a weaker Leela probably.
Depth is not a good metric at all for MCTS type programs like it is with AB type programs.
N/s is the only easy and sensible metric for performance.
As more N/s, is always better.

So what you are doing so far with your tests, probably has a somewhat weaker Leela as you didn't try to achieve higher N/s.

Then don't use GPU loading and TIME TO DEPTH. My results speak for themselves.

You act like lc0 is new, and we have not tried the NPS tuning. Like I said NPS is a false flag.

How exactly do they speak if you have nothing to compare?
I bet that using properly set up parameters for obtaining higher N/s would be better for Leela.
No need to assume things anyway, as depth is completely nonsensical metric for Leela. Optimizing for it is useless as i've said.

I am good with you are anyone doing want they want to do to setup Lc0. If you think this is wrong, I am good with it. If you think I have not tested other settings you have not been paying attention.

It is to the end user to see if my method is better or worse.

"Whoever has ears, let them hear.”

[Nordlandia]

4 threads for Lc0 = exercise in futility.

And so with your ponder match.

[mwyoung]

4 threads for Lc0 = exercise in futility.

And so with your ponder match.

You always say this. But you can never answer why. Why is using 4 threads a exercise in futility?

[Nordlandia]

Because any more than default may hurt raw performance.