Stockfish 3 Official JA Windows/Linux builds available

[gerold]

Thanks, so far half of the test is completed and I confess I'm disappointed by the result, a meager 10 ELO increase over v2.3.1

I'm even hoping something is wrong in this testing

I have been test it vs. 2.3.1. Just 25 games so far in Arena.
Bliz 1/2 . Plus 60elo . That may change with more games. Some really nice games.

Best,
Gerold.

[Matthias Gemuh]

Thanks, so far half of the test is completed and I confess I'm disappointed by the result, a meager 10 ELO increase over v2.3.1

I'm even hoping something is wrong in this testing

You are disappointed because of an ELO increase of a +3000 ELO engine ?

[gerold]

Thanks, so far half of the test is completed and I confess I'm disappointed by the result, a meager 10 ELO increase over v2.3.1

I'm even hoping something is wrong in this testing

You are disappointed because of an ELO increase of a +3000 ELO engine ?

Happy for the SF team. Now running vs. Houdini 1.5.

After 40 games Its 11-11-18 . Few games the new SF 3 is looking good .

[melajara]

Thanks, so far half of the test is completed and I confess I'm disappointed by the result, a meager 10 ELO increase over v2.3.1

I'm even hoping something is wrong in this testing

You are disappointed because of an ELO increase of a +3000 ELO engine ?

No, I know it's hard to improve at this level BUT

1, there are a lot of people involved in the improvement of Stockfish, some with interesting lateral thinking too.
2, there is now an excellent testing framework to separate the wheat from the chaff in all those "improvements".
3, There has been approximately 5 months to pile up improvements.
4, Houdini 3 raised the bar so much that it could expose previously hidden eval flaws in the top contenders tier.

So, all in all, I was expecting something like 30-40 ELO. Of course, there is the confidence margin and others factors like perhaps not easily quantifiable improvements in "style" but of course this is debatable.

For me it makes Houdini 3 even more impressive as seemingly NOBODY from the top contenders is able to catch up (if not, maybe, the elusive Komodo 6 MP)

[Albert Silver]

Finally great news! But note: JA compile 4% SLOWER than autocompile from http://abrok.eu/stockfish/ dated 30.4. in Pentium 4 prosessor (bench ran many times with same difference).

Actually I can confirm this on my i5-2500k Sandy Bridge.

[ernest]

The auto-compile is still 3% faster...

I found 4% faster on my Intel Dualcore.
After all, nor using the auto-compile version is a loss of the same order as not using the popcnt/SSE4.2...

I wonder why Jim couldn't do as well as the auto-compile!

[M ANSARI]

That would seem strange that the auto compile would be faster than a custom hand tuned compile. I guess the new compilers are much more efficient than previously thought.

[Modern Times]

Only problem is, the SSE4 autocompiles do not run on AMD Phnom II X6, whereas Jim's Intel popcount exes do.

[Jim Ablett]

Finally great news! But note: JA compile 4% SLOWER than autocompile from http://abrok.eu/stockfish/ dated 30.4. in Pentium 4 prosessor (bench ran many times with same difference).

Actually I can confirm this on my i5-2500k Sandy Bridge.

Hi Albert,

Not sure if you are running the recompiled version. Jouni's comment was about about the first release which was bad.

Anyway I can give an explanation why the autocompile popcount versions are running faster on i5/i7. They are compiled with pure SSE4.2 which takes advantage of i3/i5/i7's built in hardware support for it which will give about a 4% increase in speed.

Problem for me is SSE4.2 will not work with a lot of AMD cpus which many testers use. So for compatibilty between Intel/AMD I dont use it.

Actually can someone confirm that JA popcount builds actually run faster than the JA non-popcount ones. They may not be using popcount at all.

Jim.

[Modern Times]

Problem for me is SSE4.2 will not work with a lot of AMD cpus which many testers use. So for compatibilty between Intel/AMD I dont use it.

Jim.

Yes, the Phenom II X6 was a very popular CPU. I have two of them, Ingo has 6 or 7..

I bought the X6 in preference to Intel Sandybridge. The X6 was and is a far better CPU for my purposes. But as you say, it supports popcount but not the full SSE4.2 instruction set. So those who compile have to be careful if they want an engine to run on the X6 with popcount.