Author: Robert Hyatt
Date: 18:42:40 07/18/04
Go up one level in this thread
On July 18, 2004 at 11:24:01, Sune Fischer wrote: >On July 18, 2004 at 10:52:52, Robert Hyatt wrote: > >>Correct me if I am wrong, but the issue was "how much faster would falcon have >>to go to eliminate the hardware advantage Crafty had?" > >Yes something like that I think. > >>I believe I answered that as clearly as possible. It was _not_ a factor of 4x >>faster. It was exactly as I stated... It _could_ have been a bit more, perhaps >>10%, but that didn't happen. > >The thing that bothers me is the way you judge some factors to be important >while other factors "aren't your problem". >And it's always in a way that seeks to downplay your hardware advantage. Where did I try to "downplay" my hardware advantage? I told you _exactly_ what it was. I didn't tell you "what it might have been, or what it could have been" but rather _exactly_ what it really was. I am _always_ consistently comparing things in this regard, by _only_ considering machine A with Crafty vs machine B with Crafty. Because I can verify those numbers easily enough. IE my comparison here is a 64 bit application on a quad NUMA box with a NUMA-ignorant OS kernel, vs a single-cpu machine again running 64 bit Crafty. The speedup there is trivial to compute. I have done it dozens of times here and we have 2.0ghz, 2.2ghz and 2.4ghz opteron numbers readily available from results I posted here over the past 6 months. I think I even had some 1.8ghz numbers prior to CCT6. So, when the question was asked "how much hardware advantage did Crafty have?" I gave as accurate an answer as possible. I don't care how my opponent compiled his program since I can't compare his NPS to my NPS and get any useful information from such a comparison. I _can_ compute the actual hardware advantage I had on the quad vs had I been running on a 2.0ghz single-cpu box. I don't see how that is trying to "downplay my hardware advantage" in any shape, form or fashion. What would you prefer? 4 X 2.4 * 1.47 to compute the speed of the opteron with a 64 bit program (1.47 X faster in 64 vs 32 bit for Crafty)? As opposed to a 2.0ghz opteron in 32 bit mode? What sense does _that_ make? If falcon was 32 bit on the 2.0 box, it would be 32 bit on the 2.4ghz box. And if his parallel search is as good (and no better), then he would get _exactly_ the same speed improvement I got. Which was about 3.1 * 2.2ghz. What more could I do to make the comparison correct? you can't mix apples and oranges and compare anything usefully... So I picked _one_ program, and varied the hardware and gave the most accurate comparison I could. That's all I can reasonably do. The other stuff (32 vs 64 bit and so forth) is all random, meaningless noise, because _I_ am a 64 bit application. > >>It seems (to me) irrelevant as to what falcon _might_ have been able to do. IE >>if it was a 32 bit application on a 64 bit machine, that was a choice made by >>the programmer since 64 bit operating systems are available for the Opteron. I >>simply responded to the above question... > >I understood at first that it was running 32 bit, now after reading the thread >again I'm not so sure. > >I think what is clear is that is was a quad opteron 2.4 GHz versus single amd64 >2.0 GHz. Correct. > >Now correct me if I'm wrong, but the Opterons has more cache too, so the most >primitive estimate here would probably be around a 5 times faster box. Again, how about not inventing numbers? I gave some _real_ numbers. We have had lots of FX51 and FX53 numbers posted here. The 53's have matched my 2.4ghz numbers very closely. Again I am _only_ considering Crafty on the 850 vs Crafty on the FX-53. That is all _I_ can compare. > >Then you lose something to inferior compilers and search overhead and Falcon >might do the same (was it 32 bit?). > Again, one more time, I'm not going down that road. I don't know how to get any factual data. I _can_ run crafty on an 850, and crafty on an FX-53 using the same O/S and Compiler, to see what the _hardware_ difference was for the WCCC event under discussion. I did that already. 32 bit vs 64 bit and so forth is not comparable. IE I can tell you crafty is 1.47X faster in 64 bit than in 32 bit, as published by the folks at AMD after _they_ did the measurements. Doesn't mean that Falcon would speed up that much. Or it might speed up more. I have no way to know and no way to test that. >Maybe not an easy estimate with the software and kernel issues, but it appears >closer to a factor 4 than factor 3 if you count everything, IMO. Whatever it may "appear" to be to you, if I run Crafty on the box I used, vs crafty on a single CPU 2.0ghz box, the number I gave is _very_ close. That's all I can compare or say about them. But for Crafty, I _have_ done the tests... > >-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.