Stockfish 2.3 update available

[Modern Times]

No, MSVCP110.dll

[Modern Times]

My current elo gain at 15" for 2.3 is 18, which implies around 12 or so at blitz levels. But this is with a speedup on Linux (about 8% now). So I'll be surprised if SF 2.3 shows even +10 on Windows if you show a decrease in NPS. As you say, we'll see. Frankly we were hoping for a bigger gain by SF, as currently only Ippo-related programs are close to Komodo at hyperbullet speeds, and we could use an unrelated opponent closer in strength for testing.

You could well be right, I guess we will see in a few days as results come in. FRC can be a little different too. But I can't run FRC at the moment due to the bug.

[Laskos]

If you send me a pm with an email address I can send to you directly a version compiled by me (not fast as Jim but enough to confirm bug is fixed).

P.S: For people testing in normal chess there is no need to wait, go with the fastest version and have fun

Non SSE compiles, ultra-bullet 2.5s + 0.04s games

Code: Select all

    Program                            Score       %     Elo    +   -    Draws

  1 Stockfish 2.3 JA 64bit         : 3227.5/6000  53.8   3213    7   7   42.9 %
  2 Stockfish 2.2.2 JA             : 2772.5/6000  46.2   3187    7   7   42.9 %

26 +/-7 Elo points improvement at this TC, probably 10-15 Elo points improvement at longer TC.

Kai

Using Linux, I show a huge gain in the direct matchup as you ran, at 5" + .05, about sixty elo. However when run against Komodo the gain drops to about thirty, and when run against Komodo at 15" + .15 it halves again to about 15. So I'll predict that it will show around 10 elo on the blitz rating lists.
I should point out that nodes per second is about 10% better for the new version. Can anyone tell us what they are getting as a nodes per second increase in Windows (over 2.2.2.)? It is also interesting that the new version is getting about 4/10s of a ply more depth at this level, far more than is due to the speed increase alone. But apparently this is not for free, as the elo gain I'm getting is fully explainable just by the speedup. Does anyone know the reason for the 10% speedup in Linux? Is it just a better compile, or has there been a true speedup in the code itself?

Windows x64.
It ended with 11% speedup, and in the middle of the 6,000 games test it was 15%. Depth was 0.4 ply deeper at this ultra-bullet control, which at these depths would mean about 20% speedup.

Kai

[Modern Times]

Depends exactly which x64 compiles you are using as well. There are at least three for both 2.2.2 and 2.3

[Laskos]

Depends exactly which x64 compiles you are using as well. There are at least three for both 2.2.2 and 2.3

No SSE, stockfish-23-64-ja and stockfish-222-64-ja compiles on AMD.

Kai

[gladius]

Using Linux, I show a huge gain in the direct matchup as you ran, at 5" + .05, about sixty elo. However when run against Komodo the gain drops to about thirty, and when run against Komodo at 15" + .15 it halves again to about 15. So I'll predict that it will show around 10 elo on the blitz rating lists.
I should point out that nodes per second is about 10% better for the new version. Can anyone tell us what they are getting as a nodes per second increase in Windows (over 2.2.2.)? It is also interesting that the new version is getting about 4/10s of a ply more depth at this level, far more than is due to the speed increase alone. But apparently this is not for free, as the elo gain I'm getting is fully explainable just by the speedup. Does anyone know the reason for the 10% speedup in Linux? Is it just a better compile, or has there been a true speedup in the code itself?

I'm quite surprised you see sixty elo at 5" even in self play, as that's what I usually test at, and I haven't seen anything close to that.

I'd expect about 10-15 ELO improvement overall. It seems some of the strength improvements didn't play nicely with each other sadly.

There definitely has been a speed-up in the code. Marco has been optimizing a bunch of stuff, and it all adds up. As well, though, some of the new evaluation features slow things down a little.

[lkaufman]

Using Linux, I show a huge gain in the direct matchup as you ran, at 5" + .05, about sixty elo. However when run against Komodo the gain drops to about thirty, and when run against Komodo at 15" + .15 it halves again to about 15. So I'll predict that it will show around 10 elo on the blitz rating lists.
I should point out that nodes per second is about 10% better for the new version. Can anyone tell us what they are getting as a nodes per second increase in Windows (over 2.2.2.)? It is also interesting that the new version is getting about 4/10s of a ply more depth at this level, far more than is due to the speed increase alone. But apparently this is not for free, as the elo gain I'm getting is fully explainable just by the speedup. Does anyone know the reason for the 10% speedup in Linux? Is it just a better compile, or has there been a true speedup in the code itself?

I'm quite surprised you see sixty elo at 5" even in self play, as that's what I usually test at, and I haven't seen anything close to that.

I'd expect about 10-15 ELO improvement overall. It seems some of the strength improvements didn't play nicely with each other sadly.

There definitely has been a speed-up in the code. Marco has been optimizing a bunch of stuff, and it all adds up. As well, though, some of the new evaluation features slow things down a little.

Bear in mind that I showed an 8% speedup in Linux, which at 5" self-play is probably worth at least 20 elo. Do you also observe a similar NPS increase? If not this might partly explain the differing results.

[gladius]

Bear in mind that I showed an 8% speedup in Linux, which at 5" self-play is probably worth at least 20 elo. Do you also observe a similar NPS increase? If not this might partly explain the differing results.

Interesting. No, I see 2.2.2 as being about the same speed as 2.3, although these are just my local compiles.

[Modern Times]

Bear in mind that I showed an 8% speedup in Linux, which at 5" self-play is probably worth at least 20 elo. Do you also observe a similar NPS increase? If not this might partly explain the differing results.

Interesting. No, I see 2.2.2 as being about the same speed as 2.3, although these are just my local compiles.

I see less for 2.3 but due to significant code changes I don't see any issue with that. But I'll check which compiles I'm using.

[geots]

Yes, with the GCC SSE42 version being just 3% faster than the generic x64 version, you will never see any Elo gain because if it exists, it will be lost in the error margins. But if you are doing analysis of some sort, then by all means take the 3%.

The problem in actual gameplay as you mention: I have never installed an exe without first checking their speed if there is a "plain" 64exe, an sse 64exe and/or a popcnt exe. And the diff. in nps speed I mentioned above that you called 3%- that is not the exception, but pretty much the rule. I don't remember one sse or popcnt that was fast enough to get it out of the margin for error that already exists.


gts