Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Does Hydra try to avoid a second Leiden?

Author: Vincent Diepeveen

Date: 07:51:41 05/19/04

Go up one level in this thread


On May 19, 2004 at 10:18:01, Ingo Althofer wrote:

>In the Leiden tournament (in April 2004) Hydra was terribly bashed by Shredder.
>Perhaps after that they only play in events without Shredder.  ;)
>
>Ingo Althofer.

I feel we must not be too hard judging the hydra team. It's the sheikh who
doesn't want to join in Israel. I'm sure the entire team wants to.

However any claim that Hydra is the strongest chess program on earth must be
forgotten. It was very lucky in Leiden in many games. I'm sure the quick level
of 90 0 and majority of participants not using big hardware has lead to Hydra
scoring more points than it based upon what it showed earned IMHO.

Basically the ultra agressive speculative evaluation which Hydra uses to play
means it will never win a world title. Ultra agressive behaviour has 2
advantages :
  a) you kill every bad tested program with big bugs left in eval (especially
king safety) thanks to agressive attack and big search depth.
  b) it has a nice playing style. attack attack without looking back...

It has 1 major disadvantage:
  a) it is incorrect and will simply lose in future

Note that there is a subtle but important difference between agressive play and
speculative agressive play.

So time is running out slowly for Hydra now. 1 chip is not so slow now compared
to software, but next year it will be when the dual opterons at 1 chip start to
arrive.

The software guys all will profit from hashtables a lot at opteron and each
generation getting faster we profit a lot.

Hardware however doesn't profit from all that. Last 6 plies no hashtables for
Hydra...

Hydra's parallel search doesn't scale to more processors easily when compared to
software using a hashtable. Moving from 8 to 16 processors i didn't see it
search deeper actually (i forgot to ask during ict4 how many processors used,
frankly i didn't care at that moment).

So in 2010 they probably will be still 16 processor clocked at 33Mhz.

Perhaps one time an update to 66Mhz and PCI-X myrinet cards (3 us one way ping
pong latency versus the current ones it has are like 8 us).

So in 2004 world champs the hardware advantage from hydra would be not there,
but in 2005 for sure also gone against many non-world champs participants.

What is left is a strong program definitely in world top and of course eyes
closed crushes software like crafty (even when using just 1 hardware processor,
because eval just matters).

However it is outdated right now IMHO.

It's advantage IMHO was using 8-16 processors, not for being in hardware.

And knowing the parallel skills of the university paderborn (feldman & co)
adding more processors we can only expect a slowdown.

When running at 460 processors with diep i did several experiments turning off
hashtable last few plies and had very discouraging results there.

Note that Feldmann correctly had already concluded this years ago (that one
*needs* to do hashtable lookups everywhere; for me it is amazing that Feldmann
could draw so many correct conclusions with such bad software implementations).







This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.