Author: J. Wesley Cleveland
Date: 10:48:43 12/09/03
Go up one level in this thread
On December 09, 2003 at 13:33:41, Robert Hyatt wrote: >On December 09, 2003 at 13:30:11, Slater Wold wrote: > >>On December 09, 2003 at 13:17:55, Robert Hyatt wrote: >> >>>On December 09, 2003 at 13:04:01, Slater Wold wrote: >>> >>>>On December 09, 2003 at 12:49:17, Robert Hyatt wrote: >>>> >>>>>On December 09, 2003 at 12:33:20, Slater Wold wrote: >>>>> >>>>>>On December 09, 2003 at 11:49:19, Mathieu Pagé wrote: >>>>>> >>>>>>>On December 09, 2003 at 11:16:02, stuart taylor wrote: >>>>>>> >>>>>>>>Is 2.2 Ghz. of a 64-bit computer a similar speed for chess as is 4.4 is it were >>>>>>>>a 32-bit one? >>>>>>>>If not, what? >>>>>>>>S.Taylor >>>>>>> >>>>>>>Hi! >>>>>>> >>>>>>>No, it is not. >>>>>>> >>>>>>>64-bit computer are not twice as fast as 32-bit ones. The number of bit >>>>>>>represent the natural lenght of an number on a cpu. Since chess engines use lot >>>>>>>of 64 bits numbers they will run faster on 64 bits machines because on 32-bit >>>>>>>machines they have to do some trick to do 64 bits maths that are natural on a 64 >>>>>>>bit cpu. >>>>>>> >>>>>>>I dont think the improvement will be in the range of 2x speed up. Anyway it will >>>>>>>vary from diffrents engines. >>>>>>> >>>>>>>Mathieu >>>>>> >>>>>>GCP reported 70% with Sjeng. >>>>>> >>>>>>Bob has reported about 50% with Crafty. >>>>> >>>>>Not exactly. I reported 1.0M with a 2.8ghz xeon, vs 1.6M with a 1.8ghz >>>>>opteron. If you factor in a clock speed equalization, the xeon slows to >>>>>1.8ghz and would produce about 650K nodes per second and the opteron would >>>>>be more than 2x faster. >>>>> >>>>>I have not done a direct comparison of 32 bit code vs 64 bit code on the >>>>>opteron as I have no 32 bit compiler available there. If I get to do that >>>>>at some point it time, it would be interesting. It would be more interesting >>>>>to be able to say "use only 32 bit ops, but use all 16 registers" to get a >>>>>_real_ feel for what 64 bits offers over 32 bits, but that looks even >>>>>harder to test. >>>> >>>>Well, we can always deduct. :) >>>> >>>>An opteron 144 (1.8Ghz) running SuSe and gcc33, using -m32 to produce 32 bit >>>>code, got these results on 186.crafty: >>> >>>We can't compare with that at all. That is a _way_ old version, obviously. >>> >>>I can't do -m32 on this machine, as the libraries are all -m64 and they >>>become incompatible (I have already tried this a few days back in fact.) >>> >>> >>>> >>>>90.1 1109 >>>> >>>>The fastest 2.8Ghz Xeon on SPEC's website does: >>>> >>>>92.0 1087 (2k AS IC++ 7.0 compiler) >>>> >>>> >>>>For all practical purposes, we can say that a O144 a P4 2.8Ghz Xeon are >>>>'comparible'. >>> >>>OK. Can't argue there with no data of my own to rely on.. >>> >>> >>> >>>> >>>>1.0M to 1.6M = 60% speedup >>> >>>Not directly attributable to 64 bit however. -m32 restricts you to 8 >>>registers, while -m64 adds the other 8. That also factors in and makes >>>this less clear. >> >>Like I said, it was just a 'deduction'. I know you're pretty scientific, but >>this was just a rough comparison. >> >>I'd be confident in this statement however: "64-bits, depending on application, >>can speedup a typical chess program from 40% to 70%." >> >>Which is nice, considering it is practically 'free'. > > >the only unknown is the 40%. IE out of my 60% boost, what part comes >from 16 vs 8 registers as opposed to 64 bit vs 32 bit registers? I'll >answer that precisely one day, I hope. If you look at the specint results for the opteron with 64 bit gcc, crafty is the only test that gains significantly over opteron with 32 bit Intel. This suggests that the extra 8 registers are not contributing much to the other tests, so probably don't to crafty either.
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.