I wonder what is the reason that so many cores work for stockfish

Chris Whittington wrote:

Admin wrote:

mwyoung wrote:

Admin wrote:

3. Question for you.
Match-one, SF1 vs Dragon, 1000 games, result 550 - 450, W=100 | D=900 | L=0
Match-two, SF2 vs Dragon, 1000 games, result 550 - 450, W=200 | D=700 | L=100

Which engine is stronger, SF1 or SF2 ?

They are the same in regards to TPR.

I put the question on Talkchess.

http://talkchess.com/forum3/viewtopic.php?f=2&t=80384

9 votes so far and only one agrees with you , 6 votes go to me

Both are same strength, by definition.

Thank You!

mwyoung wrote:

1-0,0-1 = 1/2-1/2, 1/2-1/2

One has a 0% draw rate, one has 100% draw rate. But performance is equal.

I think that the above is true from the elo calculation perspective: however, as Thomas Bayes pointed out in the 1700s, you get a better answer when you take ALL the available information into account.

The extra information we now have that Arpad Elo didn't when he created his rating system is that when two chess players at approximately the same level play each other, the higher their rating, the more likely they are to draw. So under the following scenarios (assuming that A and B are good enough players to be able to actually get a win)...

Scenario 1: A and B play 100 games and get 100 draws

Scenario 2: A and B play 100 games and get 50 wins each

...it is very likely that:

1. A and B are of roughly equal standard under both scenarios

2. A and B are likely to much stronger players under scenario 1 than they are under scenario 2

mwyoung wrote:

Only 2 people out of 11 people on Talkchess got the correct answer!  That is telling.  

SF won the last 4 TCEC tournaments on even bigger hardware, how so SF scales bad?

Admin wrote:

mwyoung wrote:

Only 2 people out of 11 people on Talkchess got the correct answer!  That is telling.  

SF won the last 4 TCEC tournaments on even bigger hardware, how so SF scales bad?

There's an alternative possibility with scaling.

More cores = more threads filling the shared memory hash table.
This is good until the hash starts getting overfilled. Hash size important factor therefore.

The advantage to each thread search of more data in the hash is basically better move ordering. But, if your move ordering is already very good, then the incremental increase from more cores is less than a poor ordering engine.

So, it may not be that SF is scaling 'bad', it may be that SF move ordering is already very good.

Admin wrote:

mwyoung wrote:

Only 2 people out of 11 people on Talkchess got the correct answer!  That is telling.  

SF won the last 4 TCEC tournaments on even bigger hardware, how so SF scales bad?

Good Question, I see we have moved on from the what engine is stronger topic.  

And do not use TCEC, as that is not a testing tournament.

But the answer is simple!

Stockfish is the strongest chess engine, and a bad scaling engine.

Both can be true.

But testing conditions matter!

Chris Whittington wrote:

Admin wrote:

mwyoung wrote:

Only 2 people out of 11 people on Talkchess got the correct answer!  That is telling.  

SF won the last 4 TCEC tournaments on even bigger hardware, how so SF scales bad?

There's an alternative possibility with scaling.

More cores = more threads filling the shared memory hash table.
This is good until the hash starts getting overfilled. Hash size important factor therefore.

The advantage to each thread search of more data in the hash is basically better move ordering. But, if your move ordering is already very good, then the incremental increase from more cores is less than a poor ordering engine.

So, it may not be that SF is scaling 'bad', it may be that SF move ordering is already very good.

If that is true we need to reconcile the time to depth and core count relationship.

What is clear is Stockfish likes fewer faster cores over lots of cores.


What I hate to see is my Laptop computer beating up my 16 core desktop computer running Stockfish.