Author: Robert Hyatt
Date: 09:24:28 04/09/03
Go up one level in this thread
On April 08, 2003 at 01:54:55, Vincent Diepeveen wrote: >On April 07, 2003 at 22:17:44, Charles Worthington wrote: > >charles > > a) measuring bandwidth the AMD duals get a bit more than the Xeon duals. > b) theoretic bandwidth of the xeon mobo's is bigger but that's theory. i > can tell you a lot about theory and computers. example. the teras machine >at which DIEP runs at is having 1 terabyte bandwidth a second. I will not say a >word more. read the specs yourself and let me know if they confuse you. > >1024 processors of course and such... > >I know other machines as well at which the banwdith is also in the tens of >gigabytes a second up to a hundred gigabytes a second and that's a bit more than >it has too. > >no one checks ever. > >in theory latency of systems is like 0.5 usec and so on. Even bob claimshis >network is 0.5 usec a message. No Bob doesn't. The only .5 microsecond network I know of is the cLAN hardware I have here on my nine-node cluster. 1.25 gigabits per second, full-duplex, with .5 usec latency and this has been verified. TCP/IP thru a normal ethernet switch is about 1,000 times slower in terms of raw latency, being in the .5 msec range. > >All bla bla. > >that's all manufacturer bla bla. > >Brutus is not going to run on the paderborn machine because the practical >latency of it is like 8 us. > >theoretic latency at teras machine with 25 ns routers is 435 ns. however SGI >told me it has 50 ns and i count on a latency of 1200 ns to get a word from >other side of machine. You are now not talking "network" but "interprocessor communication" on a NUMA type machine. > >Still if it is 1500 ns that's a factor 6 faster nearly than what the paderborn >machine has where brutus was supposed to run at. > >It took 2 days of discussing about parallellism with me, ulf lorenz and chrilly >odnninger before i started to realize that i could talk optimistic till the end >of the year and still they wouldn't get much chessprograms to run parallel at >that paderborner cluster machine. 8us sucks ass for a supercomputer as latency. If you really mean 8 microseconds, then yes that is bad. Very bad. But it means that a different parallel approach _could_ work reasonably well, but it had better not depend on low latency/high bandwidth. As I do, for example. > >Note that machine is in top500.org even and not too low placed either. > >This where the superb SGI latency, theoretically like 435 ns. is just factor 3 >off what i feel it is. > >on paper that paderborn machine is 0.5 us latency .... ...but it is 8.0 us if >you measure a ping on it!! You are back to real networking now since you mention ICMP packet transmission round-trip times? 8usec seems incredibly fast for a true network using TCP/IP. The giganet (cLAN) stuff I have is faster, but I have not tested anything else that approaches such speeds for real networking. > >In that respect the x86 dudes are no better. in fact we talk about even more >money with regard to x86 cpu's than when we talk about these clusters or 0.5 >billion dollar supercomputers. > >For computerchess K7 rocks and it will be until the P4 moves to 0.09 or prescott >gets released, that it is faster for the average chessprogram. > >Your measurement of 35% is off by a large margin. > >If i measure diep then a 4 processes 2.8Ghz Xeon == 2 processor K7. in NODES A >SECOND. not in actual search depth. there the k7 wins. > >I calculated a 3.06Ghz dual Xeon is equal to dual K7 2.2Ghz. > >Now there is 166Mhz bus K7s. Cool. Cool. Let's run them dual they for about 1/3 >of the price of a 1000 euro Xeon 3.06Ghz here is ;) > >>On April 07, 2003 at 21:49:17, Aaron Gordon wrote: >> >>>On April 07, 2003 at 19:24:08, Charles Worthington wrote: >>> >>>>On April 07, 2003 at 19:13:50, Anthony Cozzie wrote: >>>> >>>>>Maybe you've never seen a real AMD system. I have 2 Athlons with HyperThreading >>>>>enabled (yes, Athlons *do* have it, its not documented and I had to do a little >>>>>soldering). With my liquid nitrogen cooling, they reach 2.93 GHZ stably. In >>>>>this position >>>>> >>>>>[D]rn3rk1/2p2pp1/bn2p2p/3qP1B1/1bpPN3/5N2/1PQ1BPPP/R4RK1 w - - 0 15 >>>>> >>>>>my system reaches 4727 knps. I'd like to see your Xeon do the same, and I notice >>>>>you haven't posted any screenshots. Stop living in the past Mr. Worthington. >>>>>Your "new" Xeon is only good for watching TV, serving porn, and heating your >>>>>room. Thats the problem with Intel technology. It becomes obsolete so quickly. >>>>> >>>>>Regards, >>>>>anthony >>>> >>>> >>>>I just emailed you the actual screenshot. I do not know how to post a photograph >>>>here. Sure i can hit 5777kN/s but not sustained and you cannot sustain 4727kN/s >>>>with deep fritz 7. Post the fritzmark result here using deep fritz 7. And if you >>>>could sustain 4727kN/s with deep fritz then i suspect you would be the first to >>>>come to the server for that ten game match. Talk is always cheaper than action. >>>> >>>>Sincerely, Charles >>> >>>He's messing with you. Athlons don't have hyperthreading (and I'm glad they >>>don't), and any soldering he'd have to do would be replacing mosfets and >>>capacitors on the motherboard to support the wattage of two chips near 3GHz. >>>Highest I've seen a dual Athlon is 2.6GHz. Yes, this would smoke any Xeon if >>>you're wondering. >>> >>>When you consider two Xeon 2.8's get about 2.1 million nodes/second in Crafty >>>and my single 2.5GHz AthlonXP gets about 1.8 million, you can easily guess what >>>a dual XP-2.6GHz would do to even a dual Xeon 3.06. >> >>Yes Aaron and that was my point. We can overclock all day. AMD's are not the >>only overclockable processors. Xeons can be overclocked as well by one who is >>knowledgable enough to do so. I am discussing factory, out of the box systems. >>It seems unreasonable at best to compare a factory machine to one which is >>overclocked unless you overclock them both for a comparison. I am trying to deal >>with the figures I see daily and not some of the fantasy figures I have seen >>posted here. The ability to build such a system and actually building one are >>two entirely different things. In practice they are not often what they seem >>they will be on the drawing board. Certainly I am no computer expert but i am >>fairly certain that the speedup resulting from overclocking is limited to the >>available bandwidth on the motherboard. To my knowledge AMD does not have a >>board with 4.1GHz bandwidth (of course I may be wrong). So I am thinking that >>there may be some bottlenecks that will slow that system down a bit. The bottom >>line is that they are both fine processors. I love AMD machines and I lose many >>games to them on the server so I know they can perform. Yes my nps are higher by >>some 35% than the 2600MP but I still lose games to them. I would not hesitate >>one bit to buy any processor made by amd. I know you like to overclock and I >>repect that. Personally it is not something that I care for but that doesn't >>mean that I am right. It's more a matter of preference than right or wrong I >>think. You and I have had some nice sparring matches over this in the past :-))) >>and I think we should just agree to disagree on that one issue. But please never >>think that I have anything against AMD. Had that been an overclocked Xeon the >>post would have read Xeon instead of AMD. I was thinking that it was a glitch in >>the cpu that caused the blunder but apparently it was shredder himself moving >>too fast in a bit of time trouble. >>And, yes, I gave my apologies to the cpu. I apologize if I have seemed >>confrontational at times because i certainly do not intend to be. I am just >>opinionated I guess. :-) have a nice evening Aaron. >> >>Sincerely, Charles
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.