Stockfish 17/09/21 Scaling Test. 1 to 16 CPU Cores.

Hardware Threadripper 2950x, RTX 2080TI 64 Gb Ram, SSD 2 TB Evo 970 Plus.

Book Perfect Book 2021
7 man table bases
TC = 1m + 1s
1 to 16 Threads
4 GB Hash

Code:

Rank Name                          Elo     +/-   Games    Wins  Losses   Draws   Points   Score    Draw
   1 Stockfish 17/09/21 - 8 CPU      11       9     160       5       0     155     82.5   51.6%   96.9%
   2 Stockfish 17/09/21 - 16 CPU       9       8     160       4       0     156     82.0   51.2%   97.5%
   3 Stockfish 17/09/21 - 4 CPU       7       9     160       4       1     155     81.5   50.9%   96.9%
   4 Stockfish 17/09/21 - 2 CPU      -2       9     160       2       3     155     79.5   49.7%   96.9%
   5 Stockfish 17/09/21 - 1 CPU     -24      14     160       0      11     149     74.5   46.6%   93.1%

I guess this is going to be the new standard. As all testing is now done at blitz or faster time controls.    
And few cores by todays standards. I guess there is no incentive to check scaling factors like in the past.
So Programmers like Stockfish do not care.

A 24 hours test showed only a +33 Elo gain in self play going from 1 to 16 CPU cores at 1m + 1s time controls. Yikes!

The good news is I may never have to upgrade my PC again with this standard.

mwyoung wrote:

Code:

Rank Name                          Elo     +/-   Games    Wins  Losses   Draws   Points   Score    Draw
   1 Stockfish 17/09/21 - 8 CPU      11       9     160       5       0     155     82.5   51.6%   96.9%
   2 Stockfish 17/09/21 - 16 CPU       9       8     160       4       0     156     82.0   51.2%   97.5%
   3 Stockfish 17/09/21 - 4 CPU       7       9     160       4       1     155     81.5   50.9%   96.9%
   4 Stockfish 17/09/21 - 2 CPU      -2       9     160       2       3     155     79.5   49.7%   96.9%
   5 Stockfish 17/09/21 - 1 CPU     -24      14     160       0      11     149     74.5   46.6%   93.1%

I guess this is going to be the new standard. As all testing is now done at blitz or faster time controls.    
And few cores by todays standards. I guess there is no incentive to check scaling factors like in the past.
So Programmers like Stockfish do not care.

A 24 hours test showed only a +33 Elo gain in self play going from 1 to 16 CPU cores at 1m + 1s time controls. Yikes!

The good news is I may never have to upgrade my PC again with this standard.

The testing in the past also was at blitz or faster time control and I do not think that it is the problem.
The problem is that when you get closer to perfect chess it is harder to get rating advantage from more time or from more cores.

It is not that programs of the past earn more elo from doubling the number of cores when you start from the same playing strength.

I wonder if a better book can help to get a bigger rating difference when I do not mean inferior lines for one side but lines that increase the chances of both sides even if the opponent does not use the book.
In other words stockfish with book A get 50% against stockfish with book B but with book A there is a bigger difference in rating between 1 core and 16 cores.

I also wonder what happen if you use no book and start every game from the opening position and if you will get a bigger difference in rating or not.
Note that stockfish with more than 1 core is not deterministic so you will not get the same game again and again even if you use no book.

Uri Blass wrote:

mwyoung wrote:

Code:

Rank Name                          Elo     +/-   Games    Wins  Losses   Draws   Points   Score    Draw
   1 Stockfish 17/09/21 - 8 CPU      11       9     160       5       0     155     82.5   51.6%   96.9%
   2 Stockfish 17/09/21 - 16 CPU       9       8     160       4       0     156     82.0   51.2%   97.5%
   3 Stockfish 17/09/21 - 4 CPU       7       9     160       4       1     155     81.5   50.9%   96.9%
   4 Stockfish 17/09/21 - 2 CPU      -2       9     160       2       3     155     79.5   49.7%   96.9%
   5 Stockfish 17/09/21 - 1 CPU     -24      14     160       0      11     149     74.5   46.6%   93.1%

I guess this is going to be the new standard. As all testing is now done at blitz or faster time controls.    
And few cores by todays standards. I guess there is no incentive to check scaling factors like in the past.
So Programmers like Stockfish do not care.

A 24 hours test showed only a +33 Elo gain in self play going from 1 to 16 CPU cores at 1m + 1s time controls. Yikes!

The good news is I may never have to upgrade my PC again with this standard.

The testing in the past also was at blitz or faster time control and I do not think that it is the problem.
The problem is that when you get closer to perfect chess it is harder to get rating advantage from more time or from more cores.

It is not that programs of the past earn more elo from doubling the number of cores when you start from the same playing strength.

I wonder if a better book can help to get a bigger rating difference when I do not mean inferior lines for one side but lines that increase the chances of both sides even if the opponent does not use the book.
In other words stockfish with book A get 50% against stockfish with book B but with book A there is a bigger difference in rating between 1 core and 16 cores.

I also wonder what happen if you use no book and start every game from the opening position and if you will get a bigger difference in rating or not.
Note that stockfish with more than 1 core is not deterministic so you will not get the same game again and again even if you use no book.

"The testing in the past also was at blitz or faster time control and I do not think that it is the problem.
The problem is that when you get closer to perfect chess it is harder to get rating advantage from more time or from more cores."

You will have to prove that to me first! My testing shows we are far from perfect chess.

If you are right, then there is no point in developing chess engines any further. As we are so near perfect play.

My guess is like in the past. Today's chess engines will be crushed by a chess engine 5 years into the future. Just like Stockfish crushes chess engines from 5 years in the past.

mwyoung wrote:

Uri Blass wrote:

mwyoung wrote:

Code:

Rank Name                          Elo     +/-   Games    Wins  Losses   Draws   Points   Score    Draw
   1 Stockfish 17/09/21 - 8 CPU      11       9     160       5       0     155     82.5   51.6%   96.9%
   2 Stockfish 17/09/21 - 16 CPU       9       8     160       4       0     156     82.0   51.2%   97.5%
   3 Stockfish 17/09/21 - 4 CPU       7       9     160       4       1     155     81.5   50.9%   96.9%
   4 Stockfish 17/09/21 - 2 CPU      -2       9     160       2       3     155     79.5   49.7%   96.9%
   5 Stockfish 17/09/21 - 1 CPU     -24      14     160       0      11     149     74.5   46.6%   93.1%

I guess this is going to be the new standard. As all testing is now done at blitz or faster time controls.    
And few cores by todays standards. I guess there is no incentive to check scaling factors like in the past.
So Programmers like Stockfish do not care.

A 24 hours test showed only a +33 Elo gain in self play going from 1 to 16 CPU cores at 1m + 1s time controls. Yikes!

The good news is I may never have to upgrade my PC again with this standard.

The testing in the past also was at blitz or faster time control and I do not think that it is the problem.
The problem is that when you get closer to perfect chess it is harder to get rating advantage from more time or from more cores.

It is not that programs of the past earn more elo from doubling the number of cores when you start from the same playing strength.

I wonder if a better book can help to get a bigger rating difference when I do not mean inferior lines for one side but lines that increase the chances of both sides even if the opponent does not use the book.
In other words stockfish with book A get 50% against stockfish with book B but with book A there is a bigger difference in rating between 1 core and 16 cores.

I also wonder what happen if you use no book and start every game from the opening position and if you will get a bigger difference in rating or not.
Note that stockfish with more than 1 core is not deterministic so you will not get the same game again and again even if you use no book.

"The testing in the past also was at blitz or faster time control and I do not think that it is the problem.
The problem is that when you get closer to perfect chess it is harder to get rating advantage from more time or from more cores."

You will have to prove that to me first! My testing shows we are far from perfect chess.

If you are right, then there is no point in developing chess engines any further. As we are so near perfect play.

My guess is like in the past. Today's chess engines will be crushed by a chess engine 5 years into the future. Just like Stockfish crushes chess engines from 5 years in the past.

I agree that there are many positions when chess engines do not play perfectly but the question is if they practically get them in chess games.

I do not claim that it is impossible to beat chess engines of today but I suspect playing against opponent's weakness may be the only way to do it

I mean that it is possible that most games that the engine play today are perfect games with no mistakes and the way to beat chess engines is not to get advantage of mistakes in games but to play different moves that cause the engines of today to make mistakes.

These different moves are not objectively better ot worse than the moves the engines practically play so there is no reason for the engines to find them with more time.

mwyoung wrote:

Code:

Rank Name                          Elo     +/-   Games    Wins  Losses   Draws   Points   Score    Draw
   1 Stockfish 17/09/21 - 8 CPU      11       9     160       5       0     155     82.5   51.6%   96.9%
   2 Stockfish 17/09/21 - 16 CPU       9       8     160       4       0     156     82.0   51.2%   97.5%
   3 Stockfish 17/09/21 - 4 CPU       7       9     160       4       1     155     81.5   50.9%   96.9%
   4 Stockfish 17/09/21 - 2 CPU      -2       9     160       2       3     155     79.5   49.7%   96.9%
   5 Stockfish 17/09/21 - 1 CPU     -24      14     160       0      11     149     74.5   46.6%   93.1%

I guess this is going to be the new standard. As all testing is now done at blitz or faster time controls.    
And few cores by todays standards. I guess there is no incentive to check scaling factors like in the past.
So Programmers like Stockfish do not care.

A 24 hours test showed only a +33 Elo gain in self play going from 1 to 16 CPU cores at 1m + 1s time controls. Yikes!

The good news is I may never have to upgrade my PC again with this standard.

Very good research.