Leela Chess Zero 42565 vs Stockfish 140619

[mwyoung]

.... Fourth problem is that he make people agry because of his stubbornness, implying he is right when he is obviously wrong.

Forums are build, to connect peoples with the same kind of interests. So there is a great way of sharing experiences and bring the theme on a great level.
So its better to deal with the winners, the ones you feel good and connected, the ones that bring all forward. And there are a lot of them here.

Wish a good day to all computerchess enthusiasts

Clemens Keck

I think some in the community protest a bit too much. As my ponder on matches play on flawlessly....hmmm

[Nordlandia]

mwyoung: please explain why you play with ponder enabled as if ponder is off. Lc0 need CPU time to get fed, and that CPU time is allocated for SF. If you make the necessary adjustments then everything is fine and some people will actually follow your matches.

[mwyoung]

mwyoung: please explain why you play with ponder enabled as if ponder is off. Lc0 need CPU time to get fed, and that CPU time is allocated for SF. If you make the necessary adjustments then everything is fine and some people will actually follow your matches.

I have already told you I am not. My processor is best leverage in a different way. Then testers running other types of processors.

It is not efficient to carve up the CPUs like in other processors when running Lc0 and Stockfish in SMT.

HINT: Remember Lc0 is not a AB engine. And THREADS ARE NOT PROCESSES.

[ankan]

Pondering makes sense when playing against Leela on a many-core machine, as Leela wouldn't use many CPU threads while thinking, and all other threads would then go to waste. Likewise the GPUs would be idle during Stockfish' turn when Leela is not pondering.

If Leela need two unshared cores, you can just set an affinity for it for hyper-threads 0-3. You can then set affinity for Stockfish to the remaining HT. That way they won't compete for cores. They might still compete for memory bandwidth, though; not sure how important that is for Leela.

I don't see why Leela couldn't run its two threads on the same physical core. That could of course be a disadvantage compared to running 2 threads on 2 physical cores, but that also holds for Stockfish' threads sharing physical cores. Yet some people claim that hyper-threading is beneficial compared to running 1 thread per physical core, and testing both under conditions with 2 active HT per core does not seem particularly unfair. It is just like all HT are somewhat slower physical cores. This doesn't involve any scheduling, so it should not be noisy.

Of course if you don't also reserve some cores for the OS, that would cause noise.

This creature is running 34 threads on a 16 core machine, and it IS affecting heavily the engines, much more Leela than SF. I don't know, I have to perform plain experiment to show black on paper what is obvious:

4 cores i7 CPU, 8 logical cores.
Leela on RTX 2070 GPU using 2 threads.

1/ Leela to 1 million nodes on 2 threads (using one of the latest nets) without the interference from SF:

info depth 16 seldepth 52 time 35832 nodes 1021332 score cp 27 hashfull 481 nps 28503
info depth 16 seldepth 51 time 35866 nodes 1017115 score cp 27 hashfull 481 nps 28358
info depth 16 seldepth 52 time 36044 nodes 1027829 score cp 27 hashfull 483 nps 28515

Very high stability, within 0.5% deviation speeds.


2/ SF on 8 threads AND Leela to 1 million nodes on 2 threads:

info depth 16 seldepth 53 time 43960 nodes 1058510 score cp 27 hashfull 498 nps 24078
info depth 16 seldepth 52 time 48555 nodes 1017708 score cp 27 hashfull 480 nps 20959
info depth 16 seldepth 52 time 43277 nodes 999935 score cp 27 hashfull 474 nps 23105
info depth 16 seldepth 52 time 58460 nodes 1063476 score cp 27 hashfull 498 nps 18191

The speeds on average are some 30% lower and they are very erratic, some 20-30% deviation on average from one run to another.
The issue would be even graver with his 2080ti, as it uses a bit more of CPU resources than mine 2070. And more severe in shorter runs, as there are bursts of slowdowns. All in all, this creature's tests and posts here are plain garbage.

You know are comparing a 4 core i7. To a 2950x designed to run heavy process multi threaded workloads. With huge L1, L2, and L3 cache, and quad channel 8 dimm memory. With a much better memory controller, and process and thread scheduler. In other words a much more advanced architecture to run this kind of workload.

I am just saying

Can you try similar experiment on your system (2950x)?
Shouldn't take much time.

Run lc0 with 2 threads without SF running and check nps after 1m nodes. Then keep SF running with 32 threads in the background (e.g: analysis mode/go infinite) AND run lc0 with 2 threads and check lc0 nps after 1m nodes?
I am curious to know if your high end system is really so much better.

[mwyoung]

Pondering makes sense when playing against Leela on a many-core machine, as Leela wouldn't use many CPU threads while thinking, and all other threads would then go to waste. Likewise the GPUs would be idle during Stockfish' turn when Leela is not pondering.

If Leela need two unshared cores, you can just set an affinity for it for hyper-threads 0-3. You can then set affinity for Stockfish to the remaining HT. That way they won't compete for cores. They might still compete for memory bandwidth, though; not sure how important that is for Leela.

I don't see why Leela couldn't run its two threads on the same physical core. That could of course be a disadvantage compared to running 2 threads on 2 physical cores, but that also holds for Stockfish' threads sharing physical cores. Yet some people claim that hyper-threading is beneficial compared to running 1 thread per physical core, and testing both under conditions with 2 active HT per core does not seem particularly unfair. It is just like all HT are somewhat slower physical cores. This doesn't involve any scheduling, so it should not be noisy.

Of course if you don't also reserve some cores for the OS, that would cause noise.

This creature is running 34 threads on a 16 core machine, and it IS affecting heavily the engines, much more Leela than SF. I don't know, I have to perform plain experiment to show black on paper what is obvious:

4 cores i7 CPU, 8 logical cores.
Leela on RTX 2070 GPU using 2 threads.

1/ Leela to 1 million nodes on 2 threads (using one of the latest nets) without the interference from SF:

info depth 16 seldepth 52 time 35832 nodes 1021332 score cp 27 hashfull 481 nps 28503
info depth 16 seldepth 51 time 35866 nodes 1017115 score cp 27 hashfull 481 nps 28358
info depth 16 seldepth 52 time 36044 nodes 1027829 score cp 27 hashfull 483 nps 28515

Very high stability, within 0.5% deviation speeds.


2/ SF on 8 threads AND Leela to 1 million nodes on 2 threads:

info depth 16 seldepth 53 time 43960 nodes 1058510 score cp 27 hashfull 498 nps 24078
info depth 16 seldepth 52 time 48555 nodes 1017708 score cp 27 hashfull 480 nps 20959
info depth 16 seldepth 52 time 43277 nodes 999935 score cp 27 hashfull 474 nps 23105
info depth 16 seldepth 52 time 58460 nodes 1063476 score cp 27 hashfull 498 nps 18191

The speeds on average are some 30% lower and they are very erratic, some 20-30% deviation on average from one run to another.
The issue would be even graver with his 2080ti, as it uses a bit more of CPU resources than mine 2070. And more severe in shorter runs, as there are bursts of slowdowns. All in all, this creature's tests and posts here are plain garbage.

You know are comparing a 4 core i7. To a 2950x designed to run heavy process multi threaded workloads. With huge L1, L2, and L3 cache, and quad channel 8 dimm memory. With a much better memory controller, and process and thread scheduler. In other words a much more advanced architecture to run this kind of workload.

I am just saying

Can you try similar experiment on your system (2950x)?
Shouldn't take much time.

Run lc0 with 2 threads without SF running and check nps after 1m nodes. Then keep SF running with 32 threads in the background (e.g: analysis mode/go infinite) AND run lc0 with 2 threads and check lc0 nps after 1m nodes?
I am curious to know if your high end system is really so much better.

You can see the performance from the live stream and videos of ponder off matches.

I am on vacation since last weekend and have been enjoying the great outdoors. Beavis, my African grey parrot is the only one home monitoring the system.

[MikeB]

Fourth problem is that he make people agry because of his stubbornness, implying he is right when he is obviously wrong.

That's the one thing that narcissists are really good at. Everything else that they think they are good at is delusional. It's s a sickness really. It's much easier just to let him continue in his fantasy world alone. He's not here for the chess but to impress.

[Nordlandia]

MikeB: like i said - an exercise in futility

"It is still an exercise in futility. Let's say the move prediction is 50%. Both of your engines would still be reaching the correct response line half of the time in the time control and in the other half will have wasted considerable time finding the wrong move for nothing to gain from it at only half of the potential depth. Your engines might agree on the same predicted reply at 15ply for example but may change course once 30 ply is reached. An exercise in futility."

[MikeB]

MikeB: like i said - an exercise in futility

"It is still an exercise in futility. Let's say the move prediction is 50%. Both of your engines would still be reaching the correct response line half of the time in the time control and in the other half will have wasted considerable time finding the wrong move for nothing to gain from it at only half of the potential depth. Your engines might agree on the same predicted reply at 15ply for example but may change course once 30 ply is reached. An exercise in futility."

You probably know this - but just in case there are others are tired of being subject to abuse. Say you want to ignore some here and not ever see their posts as you have determineded that the odds of them posting something you would be interested in now near 0 % ( using Lc0 type scoring output) .

I will just use somebody's account at random (big smile, wink wink):

Clink on the "Foe" button.

presto - all of the posts are hidden from your view

Narcissists are tolerated more by the mainstream for usually one or more reasons :

1. They have a lot of money , often they belong to the lucky sperm club and had nothing to do with how much money they do have.
2, They are born with good looks , male or female.
3. They might be one of the best at what they do, classic example is Bobby Fischer.
4. Or they hold an important political office, such as the current POTUS.

If they do not have any of these qualities , they can be safely ignored and you will not miss much.

[MikeB]

MikeB: like i said - an exercise in futility

"It is still an exercise in futility. Let's say the move prediction is 50%. Both of your engines would still be reaching the correct response line half of the time in the time control and in the other half will have wasted considerable time finding the wrong move for nothing to gain from it at only half of the potential depth. Your engines might agree on the same predicted reply at 15ply for example but may change course once 30 ply is reached. An exercise in futility."

You probably know this - but just in case there are others are tired of being subject to abuse. Say you want to ignore some here and not ever see their posts as you have determineded that the odds of them posting something you would be interested in now near 0 % ( using Lc0 type scoring output) .

I will just use somebody's account at random (big smile, wink wink):

Clink on the "Foe" button.

presto - all of the posts are hidden from your view

Narcissists are tolerated more by the mainstream for usually one or more reasons :

1. They have a lot of money , often they belong to the lucky sperm club and had nothing to do with how much money they do have.
2, They are born with good looks , male or female.
3. They might be one of the best at what they do, classic example is Bobby Fischer.
4. Or they hold an important political office, such as the current POTUS.

If they do not have any of these qualities , they can be safely ignored and you will not miss much.

[mwyoung]

MikeB: like i said - an exercise in futility

"It is still an exercise in futility. Let's say the move prediction is 50%. Both of your engines would still be reaching the correct response line half of the time in the time control and in the other half will have wasted considerable time finding the wrong move for nothing to gain from it at only half of the potential depth. Your engines might agree on the same predicted reply at 15ply for example but may change course once 30 ply is reached. An exercise in futility."

Yep. You have to ask the cost and return. It cost stockfish 6% and lc0 nothing. But stockfish can take advantage of the CPU overclock, but not lc0. As the GPU is the bottle neck. Not the CPU.