Author: Vincent Diepeveen
Date: 12:28:26 03/05/03
Go up one level in this thread
On March 05, 2003 at 13:30:05, Anthony Cozzie wrote: the basic problem of SMT/HT is that the P4 is not the ideal cpu to use SMT/HT. the k7 is a much better cpu for SMT/HT usage. therefore the speed wins for SMT/HT are not so convincing. Another problem of SMT/HT in general is that the OSes fire threads and processes at a speed which is just having a too big latency. For vaste majority of databases it is not true to say that SMT/HT helps there, because the weakest chain there is the speed/latency of the harddisks, not the IPC of a processor. >Suppose you are running a database like Oracle or DB/2. Most of the time what >you do is this: > >load word from some random place in memory >compare >jump > >In otherwords, databases spend a *large* amount of time in the memory subsystem. > Their instruction footprints are usually several MB (unlike chess programs >which are maybe 200-300KB) and they process GB of data. Hyperthreading is great >for this kind of application; while one thread is waiting on a cache miss, the >other can continue to do useful work. Another example is Nalimov's TB gen. >Lots of semi-random memory accesses, and not surprisingly he gets a great HT >speedup. Speaking of Nalimov, I could use a job Eugene :) > >This is not the case for a chess program though, which does lots of bit flipping >and computation. A chess engine is going to do only one realy semi-random >memory access per node - the transposition table. The result is that HT gives a >some speedup, but nothing exceptional. Quite frankly I am suprised Hyatt gets >25% in crafty. > >I think in the future we will see true single chip multiprocessors. With 90nm >and smaller processes, it should be cost effective to put two PIV cores on one >die. > >anthony
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.