Author: Robert Hyatt
Date: 13:26:41 12/17/02
Go up one level in this thread
On December 17, 2002 at 15:27:31, Matt Taylor wrote: >On December 17, 2002 at 15:19:34, Robert Hyatt wrote: > >>On December 17, 2002 at 14:21:10, Gian-Carlo Pascutto wrote: >> >>>On December 17, 2002 at 13:15:36, Robert Hyatt wrote: >>> >>>>On December 17, 2002 at 11:58:21, Gian-Carlo Pascutto wrote: >>>> >>>>>On December 17, 2002 at 11:27:18, Matt Taylor wrote: >>>>> >>>>>>Crafty gets better results with HT, >>>>> >>>>>In addition to what Vincent said, the data we currently >>>>>have is saying exactly the opposite. >>>>> >>>> >>>>I don't follow. I posted the following twice already: >>> >>>I've done my own tests. >>> >>>>1 thread, SMT disabled, 24 positions, run twice, 1001.5K nps >>>>2 threads, SMT disabled, same conditions, 1604.5K nps >>>>3 threads, SMT enabled, same conditions, 1820.25K nps >>>>4 threads, SMT enabled, same conditions, 1923.0K nps >>>> >>>>Hyperthreading took the 16.04.5K nps for two bare xeon processors and >>>>improved that by 20%. I can certainly post the raw data if it is important. >>>>I believe 20% is definitely better than 11%. And 20% is not something to wave >>>>off as unimportant. >>> >>>I measured similar numbers for 1->2 cpus (24%). But this is only >>>NPS increase, not the actual speedup which will be lower. >> >>So? That's outside the scope of the SMT/HT discussion. The question is "does >>it >>work" not "does it work well for chess applications." >> >>>-- >>>GCP > >Intel claims 30-40% for the former. Dr. Hyatt claims about 20% for the latter. >Vincent claims 11% for the latter. Eugene reports some things going 100% faster. > >The question of which application is entirely relevant for obvious reasons. In >the general case, most things will benefit. Whether or not they benefit -much- >is different, but it doesn't really matter. HT isn't costing Intel more than >pocket change to add to the CPU. Even a 1% speed gain for a dollar isn't too >bad. It's just unfortunate that you can't buy CPUs on that curve. AthlonXP 2100 >= $85. If only you could double performance for an extra $100! All an extra $100 >gets you is AthlonXP 2400. > >-Matt Maybe or maybe not. Notice that _all_ reported applications have run faster. QED it works. How well depends on the mix of applications. Just like the same question for a multiprogramming O/S...
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.