Author: Robert Hyatt
Date: 13:17:26 02/17/03
Go up one level in this thread
On February 17, 2003 at 13:18:21, Brian Richardson wrote: >On February 17, 2003 at 13:12:29, Robert Hyatt wrote: > >>On February 17, 2003 at 12:47:24, Brian Richardson wrote: >> >>>On February 17, 2003 at 11:22:55, Robert Hyatt wrote: >>> >>>>On February 16, 2003 at 21:56:03, Brian Richardson wrote: >>>> >>>>>On February 16, 2003 at 19:08:03, Robert Hyatt wrote: >>>>> >>>>>>On February 16, 2003 at 16:50:48, Brian Richardson wrote: >>>>>> >>>>>>>I have been running some informal Itanium2 tests with Tinker and Crafty (18.15). >>>>>>>The results are not encouraging. I know Bob Hyatt has posted better numbers for >>>>>>>Itanium2, but I can’t come even close. Then again, I am no professional >>>>>>>compiler developer or performance engineer. Anyway, here is the data (Knps). >>>>>>> >>>>>>>Intel IA64 Itanium2 (1GHz) dual CPU system vs dual AMD 1900+ (1.6GHz) >>>>>>> >>>>>>>Crafty18.15 1 CPU 2 CPUs SMP Speedup >>>>>>>IA64 368 715 1.94 >>>>>>>AMD 615 1015 1.65 >>>>>>>IA64 Slower 40% 30% >>>>>>> >>>>>> >>>>>> >>>>>> >>>>>>That looks horrible. What compiler, etc? IE the best results I have seen >>>>>>come from the microsoft in-house compiler guys. Even a 1ghz processor should >>>>>>be faster than that using a raw PIII at 1ghz... >>>>> >>>>>That is the point. Most of use do not have access to Microsoft and Intel >>>>>compiler experts. I used the free IA-64 compiler updates from Intel and >>>>>Microsoft at: >>>>>http://www.intel.com/software/products/compilers/downloads/cwin.htm >>>>>http://www.microsoft.com/msdownload/platformsdk/sdkupdate/ >>>>> >>>> >>>>Can't offer much advice since I don't have one, but it is pretty strange that >>>>the >>>>1ghz Mckinley is less than 1/2 the speed of a 600mhz 21264 alpha, which does >>>>suggest something is badly wrong... >>> >>>No, the 1GHz Itanium2 is about the same speed as a 1GHz Alpha, per published >>>SPEC results. Again, the point is that much faster results may be possible with >>>hand optimizations, which Tim likely did, I assume. >> >> >>Not sure what you mean by "hand optimizations". If you mean changes to the >>source >>code, the answer is "no". He showed up one night on ICC with a gnuchess that >>was >>running impossibly fast. I chatted with him, sent him a crafty source tarball >>and he >>compiled it and had it on ICC within 15 minutes. It was the version I posted >>the WAC >>nps numbers for (16.18) and the NPS we saw out of crafty on that box (he did use >>dual >>cpus) was pretty amazing at the time. I'm pretty sure he knew how to compile >>with good >>optimization flags, but that was all he did. No assembly of any kind, no source >>changes, >>etc... > >Perhaps version differences (16.18 v 18.6?) are pretty significant in nps. > Yes. The most recent version is actually a tiny bit slower, but nowhere near even 10%, so that doesn't explain things. However, I'll try to compile 16.18 tonight and compare it on my xeon with the current version, just so we have accurate numbers rather than my rough estimate... >> >> >> >>>Here are the Crafty SPEC results: >>>Alpha 21264A 616MHz 219 >> >> >>Clearly your Mckinley numbers ought to beat the above significantly... >> >> >> >>>Alpha 21264C 1GHz 123 >>>Alpha 21264C 1.25GHz 98.5 >>> >>>> >>>> >>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>>>x86 binary >>>>>>>on IA64 170 337 1.98 >>>>>>>IA64 Slower 72% 67% >>>>>>> >>>>>>> >>>>>>>Tinker >>>>>>>IA64 260 >>>>>>>AMD 330 >>>>>>>IA64 Slower 21% >>>>>>> >>>>>>>x86 on IA64 103 >>>>>>>IA64 Slower 69% >>>>>>> >>>>>>> >>>>>>>Note that the 2 CPU SMP efficiency for the Itanium2 system is significantly >>>>>>>better than the dual AMD system, as has been reported before. >>>>>>> >>>>>>>Also, note that both the Intel and Microsoft Itanium2 compilers produced >>>>>>>different results, depending on debug (no optimization) vs optimization modes. >>>>>>>Also, I could not get Crafty to compile with the Intel compiler, and some Tinker >>>>>>>routines broke the Intel optimizer. Moreover, the Intel compiler produced code >>>>>>>about 5% or so faster than the Microsoft compiler for Tinker. I was not able to >>>>>>>get profiling to work, nor did I try Vtune. >>>>>>> >>>>>>>The results are more or less in line with the SPEC CPU2000int results for Crafty >>>>>>>(see www.spec.org), some of which are (run times in seconds): >>>>>>> >>>>>>>Hewlett-Packard Company hp server rx2600 (1000 MHz, Itanium 2) >>>>>>>186.crafty 128 >>>>>>> >>>>>>>Advanced Micro Devices Epox 8KHA+ Motherboard, AMD Athlon (TM) XP 1900+ (CPU >>>>>>>MHz: 1600) >>>>>>>186.crafty 106 >>>>>>> >>>>>>>Intel Corporation Intel D850EMVR motherboard (3.06 GHz, Pentium 4 processor with >>>>>>>HT Technology) >>>>>>>186.crafty 86 >>>>>>> >>>>>>>Advanced Micro Devices ASUS A7N8X (REV 1.02) Motherboard, AMD Athlon (TM) XP >>>>>>>2800+ (CPU MHz: 2250) >>>>>>>186.crafty 76.3
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.