Certainly not a 2100 Chess engine! ID 231

[AlvaroBegue]

So many excited people giving unfair advantage to LCzero, like CCLS does giving a GTX 1080 while restraing all engines to a single CPU core.
That is like a 10x advantage for LCZero but he doesn't care... I care and so many others.

High-end desktop computers have multiple cores and a powerful GPU. The fact that I haven't put the effort to make RuyDos use the multiple cores, and that I have no good ideas on how to make it use the GPU means that it is fair to evaluate it on a single thread while LCZero gets to use the GPU. If an engine author doesn't think that's fair, he can change his engine to make use of modern hardware.

[Daniel Shawul]

Why is it so important to you to waste time with anger and saying "LCzero not good on equal hardware ?"

Again a strawman, I said "LCZero is 2100 ish on 1 core"

Nobody cares about the strenght of LCzero on CPU but you. Everyone interested in just use GPU.

CEGT cares, CCRL cares, anybody who plans to run it on a CPU cares. People who plan to run it on their phone (by far the majority of chess engine users) actually do not care because it will not be ready for that next century
This fact alone should make you pose and think.

You don't care because you want to live in your bubble, well atleast until the hype fades with your new baby.
So many excited people giving unfair advantage to LCzero, like CCLS does giving a GTX 1080 while restraing all engines to a single CPU core.
That is like a 10x advantage for LCZero but he doesn't care... I care and so many others.

Did you know that the latest models of smartphones are already coming with NPUs? (Neural processing units) If you didn't, now you do...

http://www.alphr.com/huawei/1008275/why ... rtphone-ai

https://www.androidauthority.com/what-i ... ns-824423/

Buddy those are not designed for intensive inference on a 40x256 network like every millisecond. They are meant for resuming AI assistance services while internet is off for cloud AI computation. Wake me up when they have an 11 TFlops gpu on mobile processors that LCzero needs.

[David Xu]

Why is it so important to you to waste time with anger and saying "LCzero not good on equal hardware ?"

Again a strawman, I said "LCZero is 2100 ish on 1 core"

Nobody cares about the strenght of LCzero on CPU but you. Everyone interested in just use GPU.

CEGT cares, CCRL cares, anybody who plans to run it on a CPU cares. People who plan to run it on their phone (by far the majority of chess engine users) actually do not care because it will not be ready for that next century
This fact alone should make you pose and think.

You don't care because you want to live in your bubble, well atleast until the hype fades with your new baby.
So many excited people giving unfair advantage to LCzero, like CCLS does giving a GTX 1080 while restraing all engines to a single CPU core.
That is like a 10x advantage for LCZero but he doesn't care... I care and so many others.

Did you know that the latest models of smartphones are already coming with NPUs? (Neural processing units) If you didn't, now you do...

http://www.alphr.com/huawei/1008275/why ... rtphone-ai

https://www.androidauthority.com/what-i ... ns-824423/

Buddy those are not designed for intensive inference on a 40x256 network like every millisecond. They are meant for resuming AI assistance services while internet is off for cloud AI computation. Wake me up when they have an 11 TFlops gpu on mobile processors that LCzero needs.

"Buddy", I can't think of a gentle way to put this, so I'll be blunt: you are starting to sound obsessed. There is not a single person here except you who keeps on talking about how LC0 is 2100 on a single core. This indicates that, with the exception of yourself, no one cares about LC0's single-core rating. You keep saying "wake me up when X happens", but as far as I can tell, you're doing an excellent job keeping yourself awake by replying over and over again with the same thing.

Give it a rest, would you?

[Uri Blass]

Ah, I said on one CPU core, not on a GPU or a 44-core TCEC machine.
Feel free to twist my words and keep on trolling along with the OP

Here, CEGT tests for your entertainment

Please make another tournament where all the above engines use just the GPU to make it fair.

And another rating list CCRL agrees with me about the 2100ish rating on 1 core.

Code: Select all

 	Leela Chess 0.7 64-bit w187	2231	+42	−40	75.9%	−207.0	15.1%	278	 
100.0%
1	Leela Chess 0.5 64-bit w123	1959	+32	−32	46.9%	+24.8	17.7%	373

Giraffe has a neural network eval and it is about 2500 on 1 core.

Even here the newer version(187) has more than 2200 and it is not the latest version.

I guess it is clearly better than 2100 also with 1 cpu.

[Daniel Shawul]

Why is it so important to you to waste time with anger and saying "LCzero not good on equal hardware ?"

Again a strawman, I said "LCZero is 2100 ish on 1 core"

Nobody cares about the strenght of LCzero on CPU but you. Everyone interested in just use GPU.

CEGT cares, CCRL cares, anybody who plans to run it on a CPU cares. People who plan to run it on their phone (by far the majority of chess engine users) actually do not care because it will not be ready for that next century
This fact alone should make you pose and think.

You don't care because you want to live in your bubble, well atleast until the hype fades with your new baby.
So many excited people giving unfair advantage to LCzero, like CCLS does giving a GTX 1080 while restraing all engines to a single CPU core.
That is like a 10x advantage for LCZero but he doesn't care... I care and so many others.

Did you know that the latest models of smartphones are already coming with NPUs? (Neural processing units) If you didn't, now you do...

http://www.alphr.com/huawei/1008275/why ... rtphone-ai

https://www.androidauthority.com/what-i ... ns-824423/

Buddy those are not designed for intensive inference on a 40x256 network like every millisecond. They are meant for resuming AI assistance services while internet is off for cloud AI computation. Wake me up when they have an 11 TFlops gpu on mobile processors that LCzero needs.

"Buddy", I can't think of a gentle way to put this, so I'll be blunt: you are starting to sound obsessed. There is not a single person here except you who keeps on talking about how LC0 is 2100 on a single core. This indicates that, with the exception of yourself, no one cares about LC0's single-core rating. You keep saying "wake me up when X happens", but as far as I can tell, you're doing an excellent job keeping yourself awake by replying over and over again with the same thing.

Give it a rest, would you?

Again all I am pointing out is 2100-ish on single core like it or not.

I will expect another LC0 fan to jump in and comment on how it is only me that cares about this over and over ...

[Daniel Shawul]

Ah, I said on one CPU core, not on a GPU or a 44-core TCEC machine.
Feel free to twist my words and keep on trolling along with the OP

Here, CEGT tests for your entertainment

Please make another tournament where all the above engines use just the GPU to make it fair.

And another rating list CCRL agrees with me about the 2100ish rating on 1 core.

Code: Select all

 	Leela Chess 0.7 64-bit w187	2231	+42	−40	75.9%	−207.0	15.1%	278	 
100.0%
1	Leela Chess 0.5 64-bit w123	1959	+32	−32	46.9%	+24.8	17.7%	373

Giraffe has a neural network eval and it is about 2500 on 1 core.

Even here the newer version(187) has more than 2200 and it is not the latest version.

I guess it is clearly better than 2100 also with 1 cpu.

2100-ish != exactly 2100 to the very last digit

[Michael Sherwin]

I have not done much reading about LC0. I understand that it is a lot like AZ. I see two camps in this thread. One camp I would term, "it is what it is, get over it--camp" and the "it is not what it is not camp, get used to it--camp". LC0 is not a 2700+ elo engine on a single core get used to it. LC0 is a 2700+ engine on multicore and a gpu get over it. If both statements are correct then what are people arguing about? They are arguing over the get over it and the get used to it part. They are arguing over personal preference. Which is quite silly. If I had a multicore alpha/beta engine that also uses the gpu and reinforcement learns then I would compare it to LC0 on similar hardware. Since my engine only uses one thread and no gpu then I naturally choose to compare my engine to LC0 that only uses one thread and no gpu. I'm not wrong for doing that. That is my preference. So the fact that Daniel points out a simple truth that is relevant to some of us should not be attacked. When I said that if SF had RL and had trained 44 million games that it would be hundreds of elo stronger and maybe as much as a 1,000 elo I was attacked. I guess that is just what people do. Didn't I read that the LC0 that is playing in TCEC has trained 10 million games? And is therefore playing 600+ elo stronger (factor in hardware also)? So I'm going to repeat what I said about AZ, "the magic is in the reinforcement learning". But I do not have the strength to study the situation. So if I got anything wrong then get over it.

[Milos]

I have not done much reading about LC0. I understand that it is a lot like AZ. I see two camps in this thread. One camp I would term, "it is what it is, get over it--camp" and the "it is not what it is not camp, get used to it--camp". LC0 is not a 2700+ elo engine on a single core get used to it. LC0 is a 2700+ engine on multicore and a gpu get over it. If both statements are correct then what are people arguing about? They are arguing over the get over it and the get used to it part. They are arguing over personal preference. Which is quite silly. If I had a multicore alpha/beta engine that also uses the gpu and reinforcement learns then I would compare it to LC0 on similar hardware. Since my engine only uses one thread and no gpu then I naturally choose to compare my engine to LC0 that only uses one thread and no gpu. I'm not wrong for doing that. That is my preference. So the fact that Daniel points out a simple truth that is relevant to some of us should not be attacked. When I said that if SF had RL and had trained 44 million games that it would be hundreds of elo stronger and maybe as much as a 1,000 elo I was attacked. I guess that is just what people do. Didn't I read that the LC0 that is playing in TCEC has trained 10 million games? And is therefore playing 600+ elo stronger (factor in hardware also)? So I'm going to repeat what I said about AZ, "the magic is in the reinforcement learning". But I do not have the strength to study the situation. So if I got anything wrong then get over it.

Well, problem is both statements cannot be true in the same time no matter what.
Lets take the newest network id241. It is 15x192 network.
With this network on a single CPU core LC0 has roughly 35nps and on 1060 roughly 1500nps.
That is 40x factor or 5.5 doublings. LC0 gains 50-60Elo per doubling in the range of 1k-32k playouts per move.
If we take CCRL 40/40 as rating reference LC0 running on single core of CPU would get more than 1k playouts per move.
So we can have pretty reliable estimate that LC0 on 1060 is 300-350Elo stronger than LC0 on a single core of CPU.
Therefore, it is absolutely impossible that these two are separated with 600+ Elo.
So someone must be wrong in their estimation.

[Michael Sherwin]

I have not done much reading about LC0. I understand that it is a lot like AZ. I see two camps in this thread. One camp I would term, "it is what it is, get over it--camp" and the "it is not what it is not camp, get used to it--camp". LC0 is not a 2700+ elo engine on a single core get used to it. LC0 is a 2700+ engine on multicore and a gpu get over it. If both statements are correct then what are people arguing about? They are arguing over the get over it and the get used to it part. They are arguing over personal preference. Which is quite silly. If I had a multicore alpha/beta engine that also uses the gpu and reinforcement learns then I would compare it to LC0 on similar hardware. Since my engine only uses one thread and no gpu then I naturally choose to compare my engine to LC0 that only uses one thread and no gpu. I'm not wrong for doing that. That is my preference. So the fact that Daniel points out a simple truth that is relevant to some of us should not be attacked. When I said that if SF had RL and had trained 44 million games that it would be hundreds of elo stronger and maybe as much as a 1,000 elo I was attacked. I guess that is just what people do. Didn't I read that the LC0 that is playing in TCEC has trained 10 million games? And is therefore playing 600+ elo stronger (factor in hardware also)? So I'm going to repeat what I said about AZ, "the magic is in the reinforcement learning". But I do not have the strength to study the situation. So if I got anything wrong then get over it.

Well, problem is both statements cannot be true in the same time no matter what.
Lets take the newest network id241. It is 15x192 network.
With this network on a single CPU core LC0 has roughly 35nps and on 1060 roughly 1500nps.
That is 40x factor or 5.5 doublings. LC0 gains 50-60Elo per doubling in the range of 1k-32k playouts per move.
If we take CCRL 40/40 as rating reference LC0 running on single core of CPU would get more than 1k playouts per move.
So we can have pretty reliable estimate that LC0 on 1060 is 300-350Elo stronger than LC0 on a single core of CPU.
Therefore, it is absolutely impossible that these two are separated with 600+ Elo.
So someone must be wrong in their estimation.

Maybe someone does have it wrong. I can't invest any energy or time into this question. I was only giving my 2 cents about the argument being largely a preferences argument. However, I do not see the effect of learning mentioned in your numbers. Am I to understand that the learning was the same for both configurations? Also the gain in elo from doubling may not be the same for MCTS as it is for AB. Are you saying it is?

[Jhoravi]

I heard that SSE/AVX extensions for matrix multiply improves speed by a factor of 4. Can the CPU version be speedup this much?