Speed 2017

[Robert Pope]

Perft gives 149 kilo nodes per second. Uses same move iterator.

Counting material only gives 500KN/second

That makes no sense.

If perft uses other move iterator that does not do move ordering, it is 1.3MN/second. Perft calls the slow move iterator of search so it may help catching bugs.

Abbess isn't fast by any means, but it gets 6-10M nps in perft and 1M nps in search.

[AlvaroBegue]

Perft gives 149 kilo nodes per second. Uses same move iterator.

Counting material only gives 500KN/second

That makes no sense.

If perft uses other move iterator that does not do move ordering, it is 1.3MN/second. Perft calls the slow move iterator of search so it may help catching bugs.

I remain unconvinced. Why should move ordering slow down perft more than an actual search? And what on Earth are you doing in your move ordering that eats up nearly 90% of the CPU time?

[Henk]

Perft gives 149 kilo nodes per second. Uses same move iterator.

Counting material only gives 500KN/second

That makes no sense.

If perft uses other move iterator that does not do move ordering, it is 1.3MN/second. Perft calls the slow move iterator of search so it may help catching bugs.

I remain unconvinced. Why should move ordering slow down perft more than an actual search? And what on Earth are you doing in your move ordering that eats up nearly 90% of the CPU time?

For move ordering it ranks quiet moves. perft passes all moves so there is no fail high after a good capture or killer. It doesn't use Piecesquare table when there are no history statistics about the move.

[Luis Babboni]

Counting material only gives 500KN/second

In wich machine?

Mine count just material and have a speed of 200KN/s*!!
* in a Phenom 955.

[flok]

Embla does 16975 nodes per second on a raspberry pi 2. Jet it beats fairymax.

[Henk]

If it ranks quiet moves using history statistics only it's speed goes to 260 KN/s. That's 3 times slower than Fairy-max.

[Sven]

If it ranks quiet moves using history statistics only it's speed goes to 260 KN/s. That's 3 times slower than Fairy-max.

Don't bother with speed until your engine has become considerably stronger. Consider this reply:

Embla does 16975 nodes per second on a raspberry pi 2. Jet it beats fairymax.

So maybe you have to fix your bugs first?

[Henk]

Yes if it in evaluation I just comment out some terms. It diminishes possible bugs and improves speed as well.

Maybe Embla is using some clever pruning.