Author: Vincent Diepeveen
Date: 10:04:09 08/25/02
Go up one level in this thread
On August 25, 2002 at 12:52:38, Dan Andersson wrote: >> >>I need to note that the difference between 150ns and 80ns is nearly >>a factor 2, if that's not *considerable slower* then what is considerable >>slower? Memory that gets 2 times slower, that's a *big* slowdown. >> >Two hops or more will happen on systems with *many* CPUs. And the average cost >will be something like 1/n*80+a/n*115+b/n*150 ns (1+a+b=n). So the average would >lie in the ballpark of 120 ns on an average system (and it will be less than 190 >ns for a huge system). Thats ca 50% more than a normal memory access. Not to >shabby. Thus algoritmic alterations could be possible that ammortize the cost of >that extra 50%. I.e. Multi Probe, PV length two or more, et.c.But we will have >to see the actual performance of the chip to be sure. >MvH Dan Andersson please don't quote theoretic numbers. a chess program is not a theoretic 'read only a single stream' program. A chessprogram is both writing and reading. That's not going to be 190ns at all. I'm not interested in a dual of course. I'm not interested in a 4 processor hammer even. 8 processor hammer is more interesting though. SGI has about 1200 ns for 128 processors (each node = dual) see: http://techpubs.sgi.com/library/tpl/cgi-bin/getdoc.cgi?coll=0650&db=bks&srch=&fname=/SGI_Developer/OrOn2_PfTune/sgi_html/ch01.html That's a read LATENCY. Just as interesting is of course the costs of a write.
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.