Author: Jay Urbanski
Date: 19:39:40 04/10/03
Go up one level in this thread
On April 10, 2003 at 11:18:43, Charles Worthington wrote: >To me, at this point, the jury is still out on 4 threads vs. two. I could run >the machine on 2 threads on the server and see what results I get but those >results would be meaningless because I have no way of knowing if the machine >would have played any better or worse using all 4 threads in identical positions >against identical opponents. The deepfritzmark test clearly shows an increase in >the performance of Deep Fritz 7 on dual threads vs four_but_ a significant >slowdown in nodes per second. This seems contradictory and had I gotten the same >result from Shredder I would have been suspicious as to the accuracy of that >particular benchmark test....but I didn't. Shredder showed a significant benefit >in both nps and time-to-solution with hyperthreading enabled. But the fritz >result is baffling: On the surface a 10% speedup in nodes per second should >result in a 10% increase in the number of positions reviewed by the program. It >should also result in greater ply depth. The faster the machine searches, the >faster it should be able to solve the fritzmark position. This seems like common >sense to me. So, my question is this: Is the fritz benchmark somehow more >innacurate than shredder's or is the reasoning I am using here somehow flawed? > >Charles This should be easily testable with a little hardware. Get two identical machines, enable Hyperthreading on one machine, disable it on the other. Run a tournament with a statistically significant number of games between them. I haven't taken the time to do this, but maybe I will at some point. Just watching the NPS of various engines (Fritz, Shredder, DJ, etc) they all seem to search anywhere from 15-30% faster with HT enabled. I turn it on. Bob is also right when he points out that SMT is not the same as SMP. You don't incur all the same overheads so you can't make the same scaling arguments.
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.