Delfi 5.3B released - memory bug fixed

[Mike S.]

The MP bug is NOT yet fixed for old "D" dual core cpus, at least not for the D945 in Vista 32 bit. This is a dual core cpu with Pentium 4 architecture. Users with this cpu type should compare single and 2 core speed (and should probably prefer the single core version unless the tests indicate that 2 cores are faster, with Delfi 5.3B).

Analysis by Delfi 5.3B (1x3,4 GHz D945):

1.Nf3 Nc6 2.d4 d5 3.e3 Nf6 4.Nc3 e6 5.Bd3 Bd6 6.0-0 0-0 7.e4 dxe4 8.Nxe4
Nxe4 9.Bxe4
= (0.21) Depth: 14 00:01:01 35950kN
² (0.31) Depth: 15 00:01:29 51527kN

Analysis by Delfi 5.3B (2x3,4 GHz D945):

1.e4 Nc6 2.Nf3 e5 3.Nc3 Nf6 4.d4 exd4 5.Nxd4 Bc5 6.Nf5
² (0.29) Depth: 9 00:00:53 442kN
² (0.41) Depth: 11 00:01:29 1964kN

[Tony Thomas]

Mike Sherwin says that he doesnt see any advantage from using large hashtables. I believe it was Christophe Theron who did an experiment with doubling hash tables, and he found that you gain 5 points for each doubling, or something small. I wonder if Terminator is always just a gloomy/grumpy gus.

There is no need to hardcode 64 Mb unless Delfi is intended for blitz chess. I haven't tested Delfi in the playchess engineroom but it would be whacked by progs like Glaurung which are more customisable. I'll see whether I can find the author's contacts to ask him to join this debate.

I dont think you really understand, even with umpteen hundred MB of Ram, Delfi would still lose to Glaurung, only Glaurung is Glaurung. It's really simple, the amount of strength increment that you get from increased hash is minimal. Why contact the author? We have had many hash debates at CCC before..Some claim that the higher is the better, but others like Bob thinks that too much RAM can hurt in certain situations.

[Tony Thomas]

May be you all can use 45 Gigs of hash and turn Delfi in to another Rybka. For any kind of practical use 64Mb of hash should be enough.

Correct and it's not hard for the author to increase the size of the hash table if he thinks that the engine will benifit from that....

Mike Sherwin says that he doesnt see any advantage from using large hashtables. I believe it was Christophe Theron who did an experiment with doubling hash tables, and he found that you gain 5 points for each doubling, or something small. I wonder if Terminator is always just a gloomy/grumpy gus.

Aren't you afraid to be terminated Tony

Tell him to bring it. Goat will protect me..Goat knows stuff only Chuck Norris knows, since Chuck Norris isnt defeatable...so am I..

[Ovyron]

Aren't you afraid to be terminated Tony

Tony can't be afraid, he's the Goatminator!

[Dr.Wael Deeb]

Aren't you afraid to be terminated Tony

Tony can't be afraid, he's the Goatminator!

[terminator]

The recommended hash setting is to set the hash to 1/2 of the available RAM i.e for 1024 set the hash to 512 for long games. To dictate 64 Mb as one size fits all is not realistic.

Was the MP bug fixed? Perhaps authors should make a beta release and ask people to test before releasing a final version. The big programs have plenty of testers and so on but for the small programs, users are the guinea pigs.

[Mike S.]

While it is of course more comfortable to get a bug free engine, there is no need to be very critical about this Delfi 5.3 release. I am following the Delfi release since 3.x versions, and almost NEVER, a "B" bug fix of a version was required. In general, the quality of Delfi releases is excellent.

(and we need to keep in mind that it's freeware)

While most engines can use big hash and benefit from big hash, this is not necessarily so with any engine. Maybe Delfi uses an unusual hash concept? It's called "smart data replacement" on the release notes page. Also, the impact of bigger hash is often overestimated. I don't think that Delfi would be much stronger with 512 MB than with 64 MB. Maybe a couple of Elo points, but nothing to worry about.

[terminator]

I did not follow computer chess in the past and was surprised to learn from http://www.msbsoftware.it/delfi/source.htm that a program called lion was a clone of Fruit/Toga at the Turin World Chess Championship. Would someone be daft enough to disguise Fruit knowing well that kibitzers of the event will likely have it?