Author: Robert Hyatt
Date: 07:31:11 10/16/99
Go up one level in this thread
On October 16, 1999 at 08:27:03, Jeremiah Penery wrote: >On October 15, 1999 at 23:21:29, Robert Hyatt wrote: > >>On October 15, 1999 at 18:29:28, Eelco de Groot wrote: >> >>>Finally some real news on this! Thanks Scott! Is it really true that the >>>original chip didn't use extensions at all in the last 4 ply search, where the >>>chips were used, does anybody know? I read something like that in a Dutch post >>>by Vincent on the semi official CSVN mailinglist. >>> >>>Regards,Eelco >> >> >>The chess chip didn't do 'singular extensions'. I have not seen anything >>about whether it used in check extensions or anything else. Had I done the >>chip I probably would have not done _any_ extensions in the hardware for >>reasons that are a bit complex to get into here unless someone wants to >>discuss the issue in more detail. > > >I would be interested in hearing about it (provided it isn't _too_ technical :). > >Jeremiah The issue is load-balancing. They do a two-level parallel search. The first 4 plies are searched serially (one cpu). The next 4 are searched in parallel on the SP cpus. The last 4 are searched in parallel on each SP's 16 chess processors. If the chess processors do a lot of extensions, then the trees they search will vary in size too much and make the load balancing issue very complex. I would do the extensions in the first 2/3 of the search, and then want the hardware to simply give me a search result in a very predictable amount of time. No extensions would help that. And if the extensions are done normally in the first 8 plies, those 8 plies turn into 30+ plies in some variations anyway...
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.