Author: Vincent Diepeveen
Date: 06:30:18 06/30/03
Go up one level in this thread
On June 29, 2003 at 19:25:10, Peter Stayne wrote: >>the commercial software, scales dual very well in nodes a second. No big diff >>between dual K7 and dual Xeon there. >> >>So don't use stupid excuses. The K7 is simply *way* faster for the vaste >>*majority* of chessprograms. > >It is well known that the 760MP chipset should have been improved upon by AMD as >they got faster in terms of the memory management. But, before you get even more >annoyed read further... Where i find the K7 a good design, it is public knowledge that the MP chipsets (also the mpx) are not exactly good ones. They do the job they are designed for, but that's about it. There is many problems with them. So much that most likely AMD overguessed when the dual opterons would get available. As it appears now most of the delay of the opteron is caused by court cases from intel and threats from intel against the different manufacturers of mainboards of them. That happened before with K7. it was delayed a year or so, simply by intel threatening manufacturers that would produce mainboards, that they would not get soon any intel chipset anymore (and because intel is such a big factor to consider, as they have 90% of the market or so, you can't afford to stop producing mainboards with intel chipsets). AMD should have been smarter than that. On the other hand it seems that the K7 where it can get clocked a lot higher, will be outpowered by the opteron so much, that it makes no sense to keep clocking the chip higher perhaps. It is AMD probably who realized this the first. This where intel has to improve chipsets of them in order to let the new P4s run better. So where we could possibly understand the choices from AMD's viewpoint, from my viewpoint the MP chipsets suck ass. >>Just try 20 programs at it and you'll see. > >I don't have a basis of comparison for these others. I am a big fan of >non-commercial engines and run tournaments frequently between many of these >programs. It is very easy. Look at what hardware they usually run at tournaments. K7 K7 K7 K7 K7. Only those who get for free their computer run at a P4, like chessbase. That is a major handicap for shredder for example. >>So wake up. open your eyes. there is another 200 engines or so. > >First part of that is unnecessary, I merely did not have a basis of comparison >for the results I was getting with my box. Would be great to have a webpage >devoted to chess benchmarks. And, for the record, I have around 260 engines on >my HD at the moment :) (a lot of them different versions of the same engine) If we count different versions and if i would exclude DIEP from that (because trivially i release sometimes within several days several versions) then i still would have more than 260 :) >>And those are *not* on par. >> >>Here is the problem of fritz. It is such a tiny program that it can't even take >>advantage of the big features the K7 offers for positional chess knowledge. >>example is its 128KB L1 cache, versus the 8KB + 12K (that's microops, not even >>instructions) L1 cache of the P4. > >Interesting. I did see the Diep benchmarks you had posted for Ace's Hardware a >ways back for the older chips. Also interesting. those were very old benchmarks. p4 is better now than it is looking there. that was with the RDRAM. It was in those days 70% slower the P4 than a K7. That huge diff is no longer the case. I still do not understand why intel went for RDRAM. It was trivially going to lose from DDR ram in the long run because of its price, no matter its good or bad performance. Perhaps because a monopolist likes to do business with a monopolist :) >>Of course the marketing department never told you fritz doesn't know the diff >>between a bishop and a knight. Perhaps you as USER should demand that it knows >>what a strong bishop is! >> >>Guess why it fits in such a tiny processor? >> >>It just doesn't have all that knowledge! >> >>You just look to 1 very tiny in assembly written program. Now please take a look >>to something that can run fast at the opteron in 64 bits too. > >What engines are written in x86-64 other than Crafty? I honestly would like to >know as I will be buying a dual when the 2GHz chips come out. All engines that you recompile. This is a 5 minute job at most. All you need is a x86-64 compiler. However engines like fritz and genius and rebel which are written in assembly won't be profitting too much from it. When i have a x86-64 compiler in my hands i can compile diep within perhaps 1 minute for x86-64 ? that's all. and then it takes advantage of the 16 registers instead of the tiny 8 we have at the x86 now. that will give *major* boosts in speed together with all the other advantages that x86-64 offers. 64 bits for DIEP is less interesting than it is for crafty. Still in the long run it will give some speedup. It is all the other advantages that make it a real pleasure. Note that i love in the future to get for example 8GB hashtable. No way at x86 it goes till 2GB more or less. >>I'm not sure Frans is going to do effort to rewrite his entire assembly to >>x86-64 :) >> >>Where others can efficiently work on the knowledge or search or whatever, he has >>to recode for a full year his engine in order to run at such a new architecture. >> >>All i need is someone else (Nalimov?) sweatting creating a 64 bits x86-64 visual >>c++ compiler. If not then i'll go for 64 bits GCC at the opteron. > >I'm confused, the original discussion was about MP's versus Xeons. I >wholeheartedly agree that Opterons, even with 32-bit software will blow >everything else out of the water. we should not compare 32 bits software compiled for 32 bits platform at the opteron with the Xeon. this is the gigantic mistake made. we must compare 64 bits software compiled for 64 bits platform with Xeon. if i compile diep 64 bits at 64 bits platform then it can take advantage of the extra registers etc, which is not there at the x86. Note that Xeon MP ===> $5000 a cpu or so and only 2.8Ghz nowadays. Now opteron will make that price more competative. >>And if that ain't enough then i'll take a bit older compiler (mipspro) and just >>get 500 cpu's :) > >>Frans on the other hand is really outdated busy. Half his life he is just >>porting his code to a new assembly language and a new processor. No time to >>improve its knowledge! >> >>Best regards, >>Vincent > >The Playchess server lets you play with any engine you want. That includes Bloody Nonsense. It only runs the fritz interface. I do not recollect that the chesspartner interface or the winboard can logon to that server. My own GUI won't connect to it either. i do not have the protocol to start with! >Fritz, Shredder (which I've been using a lot lately), DJ, and any other CB, UCI, >or WB (via wb2uci of course) engine you want. Not sure how you got kicked into >this Fritz-bashing session :). > >Anyways, objective information is far more important to me than patriotism to a >company (which was why I went Xeon in the first place despite my greater respect >for AMD, because it did what I did buy this machine for better, which was not >chess). But, like I said before, I simply don't have anyone with an MP system Well patriottism has nothing to do with it. AMD is american. Intel is american. Intel gets produced basically in Malaysia and such republics and has some factories in ireland as well and everywhere on the planet. AMD has a 0.13 factory in germany if i remember well and now they deal with IBM to make a 0.09 factory, and it is hard to deny that IBM isn't world wide either. patriottism has nothing to do with the choice. the owners are all americans. >or >any webpage to go to when I run, say, Ruffian, or Deep Sjeng, or whatever to >compare it to. i have comparisions of deepsjeng as i happen to have quite some versions of sjeng. it kicks butt on K7 when compared to P4. Sjeng is no exception there. All the software is fast at K7 simply. Please keep in mind that those compilers tested do not even optimize very well for the K7. This where we WILL see very good optimizing compilers for x86-64. I hope you see that the only advantage of the P4 is that it has SSE2 which the K7 has not. For some software that is a big advantage because they require 64 bits and the P4 offers a very slow form of it in SSE2 (up to 128 bits actually) and the K7 doesn't. So it is the true 32 bits software which is not fitting within the trace cache too well OR is having a bunch of branches which always will be faster at the K7 than the P4 by *definition*. It is just some SSE2 software and software that does NOT need a processor, just a big bandwidth of the cache, and P4 sure delivers a big cache bandwidth thanks to 128 bytes cache line length versus K7 64 bytes. Therefore the K7 is again superior for chessprograms because a smaller cache length usually means a quicker latency. And chess just needs latency. Now every bad thing from the K7, and it sure has little from computerchess viewpoint, it improved in the opteron. Some strong points, like a huge BTB is increased even factor 8 times. to 16384 entries !!!! In short that opteron when software is compiled for it, will kick butt. Now perhaps you can ask Nalimov when microsoft will release a visual c++ compiler for the opteron or some service pack that will do the job :) That would answer the question when even the most lazy programmer will be fast at the opteron. >The observations that I had made were based off of the only means of comparison >I did have. And, TBH, they took me by surprise. For bandwidth it is easy math. What is faster: x*128 or y*64 Usually y is a bit bigger than x, but not factor 2. x and y represent the number of times a second that one can transport a cache line. It is trivial then that y which is bigger than x, is better for chess. >If you have a resource that'll allow me to do this, I'd be very interested in >using it. I know you've been in this game for a long time, so I respect what you >say on this matter. > >Ran across this link that evidences your viewpoint: > >http://www.tecchannel.de/hardware/1164/9.html > >Also, offtopic for this discussion, but I would like to run an SMP tournament >with the same conditions as the RK2003 tournament. I would like to include Diep >in this, is this privateware? freeware? commercial? The links I've seen pointing >to a Diep homepage are dead. Answer is no. Note i do not know what RK2003 is. I just am interested in joining world champs 2003 basically. All the other tournaments are commercially not important. Yes all my homepages are died. Perhaps i start homepage in september 2003 again or januari 2004. Not sure yet. Not before that for sure.
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.