Zappa on a 32 threads AMD (Barcelona core)

[Tony Thomas]

I would love to see some games!!!

Me too... 24 million nodes per second! And 28.7 ply, is that a depths average from the middlegame?

It would really be interesting if some games, hopefully with eval/depth infos per move, could be retrieved and posted here. I wonder if it ever loses.

(Unfortunately I cannot access the engine room, because my serial numbers expired.)

May be someone should make an anti-Rybka book from using the games of this monster versus Rybka.

[Dann Corbit]

Looking at the CCRL data, it does not appear that Zappa scales better than Rybka:

1 Rybka 2.2 64-bit 4CPU 3105 +31 -30 73.0% -156.8 38.2% 369 78.4%
Rybka 2.2 64-bit 2CPU 3052 +24 -24 74.3% -172.7 33.4% 665 65.3%
Rybka 2.2 64-bit 2999 +35 -34 70.0% -128.8 47.9% 267 61.6%
2 Zap!Chess Zanzibar 64-bit 4CPU 3044 +27 -26 66.9% -104.4 46.8% 453 62.3%
Zap!Chess Zanzibar 64-bit 2CPU 3032 +65 -64 58.3% -61.1 39.7% 78 64.8%
Zap!Chess Zanzibar 64-bit 2905 +28 -27 55.7% -34.6 52.0% 383 62.3%

To me it does. The 1-4 cpu difference in ELO for Rybka is 106, while in Zappa it is 139. The 2-4 cpu difference for Zappa can be explained by the fact that only 78 games were played.

Zach

The error bars are large enough to prevent drawing such a conclusion.
I suppose when we have several thousand games we will be able to see which one is better.

The CEGT 40/20 data shows an increase of about 70 Elo per doubling for both programs (not as many variants are available):
no Program Elo + - Games Score Av.Op. Draws
1 Rybka 2.2 x64 2CPU 3014 16 16 1205 72.4 % 2846 35.2 %
13 Rybka 2.2 w32 1CPU 2937 17 17 1184 69.0 % 2798 34.4 %
8 Zap!Chess Zanzibar x64 4CPU 2992 36 36 200 56.0 % 2951 45.0 %
15 Zap!Chess Zanzibar x64 2CPU 2923 15 15 1250 57.4 % 2871 41.8 %

I don't see anything particularly significant here either.
One thing that seems clear to me is that we have not got nearly enough information to extrapolate to 32 threads.

There is nothing wrong with forming approximate conclusions based on limited information. But I think certainty requires data.

[Spock]

One of the aims of our quad testing was to be able to measure the scaling from 1CPU - 2CPU - 4CPU but in many cases we just haven't got enough games of all three to be able to draw definite conclusions. Try looking at the blitz list, there are far more games there

http://www.computerchess.org.uk/ccrl/404/

eg look at Glaurung
http://www.computerchess.org.uk/ccrl/40 ... =by+rating