Author: Robert Hyatt
Date: 07:06:31 11/13/03
Go up one level in this thread
On November 12, 2003 at 23:55:53, Aaron Gordon wrote: >On November 12, 2003 at 13:30:11, Robert Hyatt wrote: > >>On November 11, 2003 at 22:41:08, Aaron Gordon wrote: >> >>>On November 11, 2003 at 21:05:31, Eugene Nalimov wrote: >>> >>>>On November 11, 2003 at 19:55:22, Aaron Gordon wrote: >>>> >>>>>On November 11, 2003 at 17:38:38, Eugene Nalimov wrote: >>>>> >>>>>>On November 11, 2003 at 08:50:53, Aaron Gordon wrote: >>>>>> >>>>>>>It would be better had they used a quad or 8-way Opteron running 2GHz or more. >>>>>>>From some testing I've done in the past you can figure a single Opteron 2GHz == >>>>>>>a P4-3.6GHz in Fritz 8 (32bit mode). So, a Quad Opteron 2.0 == Quad P4-3.6. >>>>>>>Almost 30% faster, plus the memory bandwidth available would probably push it a >>>>>>>bit over that with large hash table sizes. 8-way Opteron 2.0 would of course be >>>>>>>like 8 p4-3.6's (however with some 40gb/s+ memory bandwidth available depending >>>>>>>on bus speed). >>>>>>> >>>>>>>Why not use the best hardware? Seems like if you'd want to promote your new >>>>>>>'awesome' chess program you'd want to give it the best chance of winning. >>>>>> >>>>>>I am not so sure that for SMP program that is not NUMA-aware quad Opteron will >>>>>>be faster than quad Xeon, even if single-CPU Opteron is faster than single-CPU >>>>>>Xeon. >>>>>> >>>>>>At least it was so for Crafty. Before we modified it to be NUMA-aware, 1.8GHz >>>>>>Opteron was faster than 1.5GHz Itanium, but quad 1.4GHz Itanium was faster than >>>>>>quad 1.5GHz Opteron. Actually, Itanium was slightly faster even on 2 CPUs. >>>>>> >>>>>>Thanks, >>>>>>Eugene >>>>> >>>>>Fritz doesn't run on the Itanium platform that I'm aware of.. not without >>>>>emulation. All of the Itanium emulation I've seen runs like a 486.. so.. back to >>>>>the 4 to 8-way Opteron arguement I go. :) >>>> >>>>I gave numbers for Crafty. Not for Fritz, Chess Tiger, Hiarcs, etc. I definitely >>>>told so in my post. >>>> >>>>Ok, trying to re-phrase it: >>>> >>>>You told: "For program 'F' single-CPU system 'O' is 30% faster than system 'P', >>>>so quad system 'O' is definitely faster than quad system 'P'". >>>> >>>>I answered: "Not necessary. For (another) program 'C' (prior to my NUMA >>>>modifications) single-CPU system 'O' is faster than system 'I', bud quad system >>>>'O' is noticeable slower than quad system 'I'". >>>> >>>>Yes, I did not have the exact numbers for Fritz. But the numbers I have suggest >>>>that quad system 'P' can in fact be faster than system 'O'. It's hard to say >>>>without measuring... >>>> >>>>Thanks, >>>>Eugene >>> >>>I was just stating at this point it would be easier, and the best bet, to use a >>>quad Opteron now rather than spend countless hours redoing the entire engine >>>before an important match. You can get a good speed boost without optimizing. >>>Just install it on the quad and let it rip. Even if they were going to recode >>>the engine I'd bet the Opteron would require much less work than completely >>>porting over to the Itanium. >>> >>>About your Itanium performance comment. I have yet to see anything of the sort. >>>Mostly I hear of 'internal' benchmarks from Intel... then when the 'regular' >>>people try those systems the benchmarks show a different story. I believe there >>>was an article about this some time ago.. if I find it again I'll definitely >>>post the link here. >>> >>>Also, crafty has some SMP problems with some systems. How are you certain that >>>the particular problem in question did not come up? The Opteron has a memory >>>channel per cpu resulting in an excellent speedups. On the dual Athlon system I >>>saw speedups from 1.2x to 1.6x in crafty. Some people get 1.9x speedups in >>>crafty.. my dual Celeron 400@552 gets 1.9 or so. It could have been a Crafty >>>problem, not an Opteron problem. >>> >> >>The point being that this "problem" has been found and fixed. It was a >>result of how AMD handles cache coherency vs Intel. And about what Intel >>changed in the PIV cache coherency as well. The current version of >>Crafty is back to scaling reasonably on my PIV xeon. Before it was not >>doing well if you recall. We also had some ugly results from the Opteron >>due to the NUMA problem. Much of that has been fixed and Crafty seems to >>scale just fine there also. I'd hope that it now does pretty well on all >>boxes, but it will take time to find out... > >Ah, thats good that it has been fixed. Any ideas when that particular version >will be available for download? I can send it to you for testing if you want to try it on your AMD hardware. I've not finished the final changes Eugene sent to me. We were working on the idea of trying to uniformly distribute the hash table across all processors, to reduce "hot spots". I believe Eugene got this to working and sent me fixes, but I have not yet had time to get them into Crafty around some changes I had made on my end. But you are welcome to try it if you want to see if it scales better. It certainly seems to do much better on my dual xeon now. > >>>All of the other chess programs (such as fritz, deep junior, etc) had speedups >>>of 1.85x or more on the dual Athlon however. I'd try some more reliable testing >>>on the Itanium vs Opteron situation if I were you.
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.