Sergio NN 40x512

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
corres
Posts: 1916
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Sergio NN 40x512

Post by corres » Mon Jan 13, 2020 9:15 am

mwyoung wrote:
Mon Jan 13, 2020 9:00 am
...
My first round of testing is giving the Sergio 512 network a jump of 17 elo over the sergio 1705 network in early testing.
What is about your system and where are the tests for those "optimized parameters" and the detailed test results?

mwyoung
Posts: 1764
Joined: Wed May 12, 2010 8:00 pm

Re: Sergio NN 40x512

Post by mwyoung » Mon Jan 13, 2020 9:21 am

shrapnel wrote:
Mon Jan 13, 2020 4:16 am
corres wrote:
Mon Jan 13, 2020 12:48 am
The speed of a net is such an important parameter than the knowledge stored in NN. A "clever" but too slow net is not better then a fast but "stupid" net. The task of net-makers is for making nets with optimal speed and knowledge.
I would say that it is rather YOUR Task, not NN maker's, to make sure you have the fastest possible GPU to fully UTILIZE the large NN.
Its rather silly to use a slow Graphic Card for Testing and then say that large NN is bad !
You neglect the importance of Hardware for testing NNs.
Corres has no idea what he is saying. After what he is saying here. I have doubts he even knows how a computer works. The speed of a net is irrelevant in terms of NPS. What matters is does it give you superior performance over any other networks.
Early testing is showing the 512 network is giving us better analysis over the smaller networks in early testing. Right now it is up +17 elo in early testing over the previous best net of 384x30-t40-1705
Professing themselves to be wise, they became fools,
Take on me. foes 0

mwyoung
Posts: 1764
Joined: Wed May 12, 2010 8:00 pm

Re: Sergio NN 40x512

Post by mwyoung » Mon Jan 13, 2020 9:31 am

corres wrote:
Mon Jan 13, 2020 9:15 am
mwyoung wrote:
Mon Jan 13, 2020 9:00 am
...
My first round of testing is giving the Sergio 512 network a jump of 17 elo over the sergio 1705 network in early testing.
What is about your system and where are the tests for those "optimized parameters" and the detailed test results?
Just jump over to the Computer Chess Club: Tournaments and Matches. As I always do. I will be broadcasting the best nets and engines including the 40x512.

I am not sure what your problem is, but you act like you don't want better nets....
Professing themselves to be wise, they became fools,
Take on me. foes 0

mwyoung
Posts: 1764
Joined: Wed May 12, 2010 8:00 pm

Re: Sergio NN 40x512

Post by mwyoung » Mon Jan 13, 2020 9:52 am

corres wrote:
Mon Jan 13, 2020 9:15 am
mwyoung wrote:
Mon Jan 13, 2020 9:00 am
...
My first round of testing is giving the Sergio 512 network a jump of 17 elo over the sergio 1705 network in early testing.
What is about your system and where are the tests for those "optimized parameters" and the detailed test results?
GPU RTX 2080 ti.

DESKTOP-CORSAIR, Blitz 0m+10s 0


1 Lc0 v0.23.2+git.c8d9095 512 n +17 +2/=18/-1 52.38% 11.0/21
2 Lc0 v0.23.2+git.c8d9095 Sergi -17 +1/=18/-2 47.62% 10.0/21
Professing themselves to be wise, they became fools,
Take on me. foes 0

corres
Posts: 1916
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Sergio NN 40x512

Post by corres » Mon Jan 13, 2020 10:32 am

mwyoung wrote:
Mon Jan 13, 2020 9:52 am
corres wrote:
Mon Jan 13, 2020 9:15 am
mwyoung wrote:
Mon Jan 13, 2020 9:00 am
...
My first round of testing is giving the Sergio 512 network a jump of 17 elo over the sergio 1705 network in early testing.
What is about your system and where are the tests for those "optimized parameters" and the detailed test results?
GPU RTX 2080 ti.
DESKTOP-CORSAIR, Blitz 0m+10s 0
1 Lc0 v0.23.2+git.c8d9095 512 n +17 +2/=18/-1 52.38% 11.0/21
2 Lc0 v0.23.2+git.c8d9095 Sergi -17 +1/=18/-2 47.62% 10.0/21
What are the "optimized parameters" of Leela and how you get them?
What about your CPU and RAM?
And you can state 17 Elo difference after 21 games...
Note
Nobody are obliged to see your "broadcast" about your matches.
But if somebody report test result that person is obliged to write down the used parameters for tests.
I am afraid technically and practically weak contemplation could not help the creation of nets.

mwyoung
Posts: 1764
Joined: Wed May 12, 2010 8:00 pm

Re: Sergio NN 40x512

Post by mwyoung » Mon Jan 13, 2020 10:41 am

corres wrote:
Mon Jan 13, 2020 10:32 am
mwyoung wrote:
Mon Jan 13, 2020 9:52 am
corres wrote:
Mon Jan 13, 2020 9:15 am
mwyoung wrote:
Mon Jan 13, 2020 9:00 am
...
My first round of testing is giving the Sergio 512 network a jump of 17 elo over the sergio 1705 network in early testing.
What is about your system and where are the tests for those "optimized parameters" and the detailed test results?
GPU RTX 2080 ti.
DESKTOP-CORSAIR, Blitz 0m+10s 0
1 Lc0 v0.23.2+git.c8d9095 512 n +17 +2/=18/-1 52.38% 11.0/21
2 Lc0 v0.23.2+git.c8d9095 Sergi -17 +1/=18/-2 47.62% 10.0/21
What are the "optimized parameters" of Leela and how you get them?
What about your CPU and RAM?
And you can state 17 Elo difference after 21 games...
Note
in early testing
But if somebody report test result that person is obliged to write down the used parameters for tests.
Here is what I stated. "in early testing"

You are correct no is obliged to see anything. But I guess if you want to know go see for yourself. Because I am also not obliged to tell you settings. Go do your own homework.
Professing themselves to be wise, they became fools,
Take on me. foes 0

corres
Posts: 1916
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Sergio NN 40x512

Post by corres » Mon Jan 13, 2020 11:04 am

mwyoung wrote:
Mon Jan 13, 2020 10:41 am
corres wrote:
Mon Jan 13, 2020 10:32 am
mwyoung wrote:
Mon Jan 13, 2020 9:52 am
...
GPU RTX 2080 ti.
DESKTOP-CORSAIR, Blitz 0m+10s 0
1 Lc0 v0.23.2+git.c8d9095 512 n +17 +2/=18/-1 52.38% 11.0/21
2 Lc0 v0.23.2+git.c8d9095 Sergi -17 +1/=18/-2 47.62% 10.0/21
...
What are the "optimized parameters" of Leela and how you get them?
What about your CPU and RAM?
And you can state 17 Elo difference after 21 games...
Note
Nobody are obliged to see your "broadcast" about your matches.
But if somebody report test result that person is obliged to write down the used parameters for tests.
Here is what I stated. "in early testing"
You are correct no is obliged to see anything. But I guess if you want to know go see for yourself. Because I am also not obliged to tell you settings. Go do you own homework.
Na ja, everybody have right to their own stupidity. These are all about "correct testing" and "optimized parameters"...
But if you really want to contribute to making better and stronger nets you should keep some basic rule of reporting the own results. The most basic rule is to make acquainted with the circumstances of your tests.

mwyoung
Posts: 1764
Joined: Wed May 12, 2010 8:00 pm

Re: Sergio NN 40x512

Post by mwyoung » Mon Jan 13, 2020 11:24 am

corres wrote:
Mon Jan 13, 2020 11:04 am
mwyoung wrote:
Mon Jan 13, 2020 10:41 am
corres wrote:
Mon Jan 13, 2020 10:32 am
mwyoung wrote:
Mon Jan 13, 2020 9:52 am
...
GPU RTX 2080 ti.
DESKTOP-CORSAIR, Blitz 0m+10s 0
1 Lc0 v0.23.2+git.c8d9095 512 n +17 +2/=18/-1 52.38% 11.0/21
2 Lc0 v0.23.2+git.c8d9095 Sergi -17 +1/=18/-2 47.62% 10.0/21
...
What are the "optimized parameters" of Leela and how you get them?
What about your CPU and RAM?
And you can state 17 Elo difference after 21 games...
Note
Nobody are obliged to see your "broadcast" about your matches.
But if somebody report test result that person is obliged to write down the used parameters for tests.
Here is what I stated. "in early testing"
You are correct no is obliged to see anything. But I guess if you want to know go see for yourself. Because I am also not obliged to tell you settings. Go do you own homework.
Na ja, everybody have right to their own stupidity. These are all about "correct testing" and "optimized parameters"...
But if you really want to contribute to making better and stronger nets you should keep some basic rule of reporting the own results. The most basic rule is to make acquainted with the circumstances of your tests.
When I am ready as the network is new and being tuned for best performance. Testing will be shown as usual and posted on tournament and matches. Because it is all about correct testing and optimized parameters. And I am not here to contribute to make better or stronger nets. I am here to test if they have made a better and stronger net or any other A/B engine.
Professing themselves to be wise, they became fools,
Take on me. foes 0

mwyoung
Posts: 1764
Joined: Wed May 12, 2010 8:00 pm

Re: Sergio NN 40x512

Post by mwyoung » Mon Jan 13, 2020 12:32 pm

Nordlandia wrote:
Sun Jan 12, 2020 10:54 pm
In my case 2x 2080 Super is not strong enough to benefit it to the fullest. I think CCC can benefit from it.

TimeControl "3600+30"

It is too early to tell if the new network is good or bad. Clearly not an impressive game. I just started my testing against Stockfish 130120. I will see if mine will do better.
Professing themselves to be wise, they became fools,
Take on me. foes 0

mwyoung
Posts: 1764
Joined: Wed May 12, 2010 8:00 pm

Re: Sergio NN 40x512

Post by mwyoung » Mon Jan 13, 2020 12:46 pm

mwyoung wrote:
Mon Jan 13, 2020 12:32 pm
Nordlandia wrote:
Sun Jan 12, 2020 10:54 pm
In my case 2x 2080 Super is not strong enough to benefit it to the fullest. I think CCC can benefit from it.

TimeControl "3600+30"

It is too early to tell if the new network is good or bad. Clearly not an impressive game. I just started my testing against Stockfish 130120. I will see if mine will do better.
Mine did better and survived the first game against Stockfish with a draw as black.

https://www.youtube.com/watch?v=ZNgE-T6yE3E
Professing themselves to be wise, they became fools,
Take on me. foes 0

Post Reply