Author: Aaron Gordon
Date: 12:53:12 06/15/04
Go up one level in this thread
On June 15, 2004 at 06:33:57, David Mitchell wrote: >On June 15, 2004 at 02:34:57, Aaron Gordon wrote: > >>On June 15, 2004 at 01:47:02, David Mitchell wrote: >> >>>On June 14, 2004 at 21:25:48, Aaron Gordon wrote: >>> >>>>On June 14, 2004 at 21:07:03, David Mitchell wrote: >>>> >>>>>On June 14, 2004 at 20:43:04, Jim wrote: >>>>> >>>>>>In your opinion what is the best processor to have for chess programs? >>>>>>I have noticed on the SSDF rating list that the Athlon 1200 is used >>>>>>for the higher rated chess programs. >>>>>>I also read at one time on this site that the Pentium processor's >>>>>>do not perform as well with chess program's. >>>>>>Your opinion is greatly appreciated. >>>>>>Jim >>>>> >>>>>The "best" CPU for chess programs will depend on the program, but in general: >>>>> >>>>>1) Opteron >>>>>2) Xeon >>>>>3) Itanium >>>>>4) Centrino >>>>>5) Athlon >>>>>6) Pentium III >>>>>7) Pentium 4 >>>>> >>>>>The above assumes ** equal speed ** of the processor (which is never the case), >>>>>and the program being optimized for that processor. Even within a single CPU, >>>>>different versions have different sizes of cache, etc., again changing their >>>>>capability. >>>>> >>>>>After AMD's strong showing with their new Opteron, you know that Intel is >>>>>working hard on a new 64 bit processor. When it is released, the list will >>>>>certainly change. >>>>> >>>>>The above is my opinion, and certainly not the result of some exhaustive tests. >>>>> >>>>>Dave >>>> >>>>Current Xeon chips are P4's with more L2 cache and multiprocessor support. Their >>>>performance in chess programs are identical to Pentium 4 chips. The older Xeon >>>>with the P3 core is identical to the Pentium 3 in chess speed, etc. >>> >>>Respectfully disagree, Aaron. A processor with identical chip, but a larger L2 >>>cache, should definitely be a faster CPU for most chess programs, given the same >>>clock speed, etc.. >>> >>>Anytime you can limit time consuming fetches from main memory, you're speeding >>>things up. >>> >>>Thanks for the info on the CPU cores, however. >>> >>>Dave >> >>I've run tests with various L2 cache sizes, Dave. The highest speedup I observed >>was between the Tbred & Barton, which was less than 1%. Test it for yourself if >>you'd like to see. >> >>There is absolutely no way adding a little bit of L2/L3 cache it is going to >>magically knock the performance up over 60% in a chess engine, because basically >>thats what it'll take for a Xeon (p4 core) to overtake an 'old' Athlon XP/MP. My >>dual Athlon MP 2.5GHz is 40-50% faster than a dual Xeon 2.8GHz in crafty, for >>example. >> >>With optimized versions of Crafty (one binary for the P4, one for the XP) I show >>the Athlon XP/MP is about 60% faster MHz for MHz vs a Xeon/P4. The Athlon XP >>2800+ (Tbred core, 166/333fsb) is 2250MHz. 2250*1.6=3600. So, an Athlon XP 2800+ >>== a theoretical Xeon 3.6 >> >>In 32bit chess programs an Athlon FX is 20-30% faster than an Athlon XP in >>chess, and the Athlon FX-53 is 2.4GHz. Which would be like running a >>2.88-3.12GHz Athlon XP.. which would be like running a P4/Xeon at 4.6-5.0GHz. >>Now, throw it in 64bit mode and it's even more of an annihilation. :) As Hyatt >>reported, I believe a single Opteron 2.2 in 64bit mode at crafty was faster than >>his dual Xeon 2.8GHz box. >> >>So, in reality, the list would be like this for a MHz for MHz comparison: >> >>#1: Athlon FX 939 pin (low latency DDR, non-registered/non-ECC memory) >>#2: Athlon 64 754 pin (same ram as 939) >>#3: Opteron / Athlon FX 940 pin (registered+ECC, higher latency) >>#4: Centrino (souped up Pentium 3) >>#5: Athlon XP / MP (Tbred/Barton cores) >>#6: Pentium 3 (Coppermine/Tualatin) / Xeon (P3 core) >>#7: Pentium 4 (Northwood) / Xeon (P4 core) >>#8: Pentium 4 (Prescott, runs about 20% slower than Northwood at chess per MHz) > >When you compared chess program speed with same CPU except for a different L2 >cache size, can you tell me what the L2 sizes were, and what chess program you >used for the comparison? > >Crafty is so cache-stingy, it might not make a big difference, but I'd be amazed >if a program like TSCP, wouldn't see a sizeable difference if the L2 cache size >was doubled. > >Of course, I've been amazed before, and nothing beats a little testing - all >theories aside, you see the darndest results! > >Dave The only time I did see a speedup was when testing an 8K L1 cache 486 vs a 16K L1 cache 486 cpu as well as an old P54C box with slow L2 cache (6-8ns), disabling and enabling that. Those actually did produce a very large increase in Crafty. Perhaps there is some level as to which everything thats important gets loaded into the L1/L2 and no further speedups are noticed. Modern cpus of which I tested resulted in absolutely no increase however. Why this is, I can only speculate. I did test the Duron (128K L1, 64K L2) vs the Athlon Tbird (128K L1, 256K L2), both of which have the exact same core, in Crafty, a version of Fritz, Shredder, and Junior (it's been so long I forget which versions exactly). I had two different people run the tests when I did the Xeon vs P4 test. The P4 was actually slightly faster, but the result was close enough to be dismissed as margin of error. Probably due to the registered/ecc memory difference and chipset. They are the same core however, just the Xeon has added cache. Just recently I tested the new mobile Athlon XP (same core as the normal XP, just a lower voltage) with the Barton core (128K L1, 512K L2) vs my current Athlon XP Tbred (128K L1, 256K L2). Programs tested were Deep Fritz 8, Shredder 8, Hiarcs 9 and Crafty. I didn't test any further programs as all 4 resulted in the exact same nodes/second between the two processors. As far as Celerons go, they can be a mixed bag. The P4 core Celeron (starting at 1.7GHz) is abysmally slow in just about everything. In games and such a Celeron 2.4GHz is somewhere on the order of the speed of a Duron 600MHz to 1GHz depending on the game tested. Chess is a different story however for the Celeron(p4). Due to chess programs not being affected in such a manner they're actually identical to P4 chips in nodes/second. So, you can get a Celeron 2.8GHz and overclock it well over 3GHz.. resulting in performance greater than the much more expensive P4 3.4GHz EE if you can get the Celerons clock speed over 3.4GHz. However, don't expect much performance in anything BUT chess programs (and other programs that do not rely heavly on cache size/speed). Note: There are speedups noticed from the extra L2 cache in various programs.. such as some 3D Games, encoding dealing with large data sets, Photoshop, etc. However I'm only referring to chess programs at the moment so please keep that in mind.
This page took 0.01 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.