Maximum speed

[matthewlai]

I happen to have quite a few engines in my gauntlet, so I ran a search from start position on all of them. All single threaded.

This approach can be a little bit misleading if different engines count nodes in different ways.

Yeah certainly. I'm hoping they all count by number of calls to Search and QSearch. That's how most people do it nowadays I think?

[matthewlai]

What CPU?

Intel Core i5-4258U (Macbook Pro, Haswell, 2.4 GHz, 2.9 GHz turbo). Pretty sure max turbo was used because I wasn't running anything else, so the other core was idle.

[Joost Buijs]

Node-count depends upon many things, not only on the way you count nodes.

For instance using the transposition table in quiescence or not, or using lazy evaluation or not.
And there can be easily a difference in speed of a factor 2 between a simple evaluation function and a more advanced one.

Node-count doesn't mean anything with respect to playing strength.

[matthewlai]

Node-count doesn't mean anything with respect to playing strength.

I totally agree. That was not claimed . I would know that. Giraffe beats many engines searching 10x as fast.

[Henk]

I happen to have quite a few engines in my gauntlet, so I ran a search from start position on all of them. All single threaded.

This approach can be a little bit misleading if different engines count nodes in different ways.

Yes you are right for instance Skipper currently counts illegal nodes where king might be captured too. Perhaps I have to change that first.

In my opinion a node is a position with a value. No matter how you get that value. If you get value from hash it is still a node. If you do a null move it is still a node for position has changed because side to move has changed.

[Joost Buijs]

What CPU?

Intel Core i5-4258U (Macbook Pro, Haswell, 2.4 GHz, 2.9 GHz turbo). Pretty sure max turbo was used because I wasn't running anything else, so the other core was idle.

Seems a bit slowish, most engines in my gauntlet (on a Haswell that runs default at 3.6 GHz. turbo) run at ~2.2 mln. n/s single threaded.

The notebook probably uses 1 channel memory instead of 4.

[matthewlai]

What CPU?

Intel Core i5-4258U (Macbook Pro, Haswell, 2.4 GHz, 2.9 GHz turbo). Pretty sure max turbo was used because I wasn't running anything else, so the other core was idle.

Seems a bit slowish, most engines in my gauntlet (on a Haswell that runs default at 3.6 GHz. turbo) run at ~2.2 mln. n/s single threaded.

The notebook probably uses 1 channel memory instead of 4.

Yeah memory latency/bandwidth probably matters, too.

This one uses 2 channels.

[stegemma]

What CPU?

Intel Core i5-4258U (Macbook Pro, Haswell, 2.4 GHz, 2.9 GHz turbo). Pretty sure max turbo was used because I wasn't running anything else, so the other core was idle.

Seems a bit slowish, most engines in my gauntlet (on a Haswell that runs default at 3.6 GHz. turbo) run at ~2.2 mln. n/s single threaded.

The notebook probably uses 1 channel memory instead of 4.

Satana on i7 4790K @ 4 GHz runs at about 3.4 Mnps while my old Raffaela, on the same machine, runs at about 5.7 Mnps but Raffaela is weaker than Satana and Satana is weaker than any other engine...

[Joost Buijs]

I happen to have quite a few engines in my gauntlet, so I ran a search from start position on all of them. All single threaded.

This approach can be a little bit misleading if different engines count nodes in different ways.

Yes you are right for instance Skipper currently counts illegal nodes where king might be captured too. Perhaps I have to change that first.

In my opinion a node is a position with a value. No matter how you get that value. If you get value from hash it is still a node. If you do a null move it is still a node for position has changed because side to move has changed.

Every do-move() (also do-nullmove()) I call a node, and I evaluate every node except when the king is in check, which is only a small portion of the total number of nodes.

In the past I negated the evaluation from the underlying node when I did a nullmove. Recently my evaluation function became asymmetric, so that trick does not work anymore.

[WinPooh]

Node-count doesn't mean anything with respect to playing strength.

Yes, remember early Rybkas with their low nodecount.
More to say, depth of search also means almost nothing. Remember Junior