Author: Vincent Diepeveen
Date: 19:40:26 07/08/03
Go up one level in this thread
On July 08, 2003 at 21:13:58, Tom Kerrigan wrote: >On July 08, 2003 at 16:16:00, Rajen Gupta wrote: > >>On July 08, 2003 at 15:37:39, Russell Reagan wrote: >> >>>On July 08, 2003 at 15:18:55, Rajen Gupta wrote: >>> >>>>http://www.amdzone.com/articleview.cfm?articleid=1304 >>>> >>>>looks like a tie with the pentium 3.2: hardly a pentium killer! my heart bleeds >>>>for the poor opteron when intel releases a prescott! >>> >>>Please post the link to the page where the have the results of the chess program >>>benchmarking that they did. >> >>they didn't: and i am talking of the chip's overall performance as most >>end-users dont buy a chip for its chess performance only: (even on this forum of >>die-hard computer chess fanatics its quite probable that there are more intel >>based systems being used) if this so-called pentium killer does not measure up >>(and i have a horrible feeling it won't), it will be the end of AMD as we know >>it; for the last 2 yrs they have been telling us: >> >>"never mind the athlon; it is only stop gap; wait for the hammer; it is going >>to smash intel;" and finally after endless delays it does show up it appears to >>be almost geting a smashing itself! >> >>if AMD disappears, we will then be subject to intel's monopoly which will be a >>horrible thing for us all: hence my disappointment with AMD: always too little >>and too late! > >One thing you have to realize is that Opteron is at the beginning of its life on >a new process. There may still be low-hanging fruit in terms of optimizing the >design or the process, and AMD might be holding back on clocking the chip higher >(because they don't have a reason to). I believe AMD has said the Opteron/Athlon >64 will reach 2.6GHz with the current 0.13um process, so if the 2GHz chip is >beating a 3.2GHz P4, the 2.6GHz chip would beat a 4.2GHz P4. Of course, to get >above 3.2GHz, Intel has to switch to their 90nm process, so AMD is quite clearly >beating Intel on the same process. Of course, Intel is moving to a new process >faster than AMD, so that has to be taken into account, and Prescott will be >significantly more "efficient" than Northwood, but IMO it seems like AMD has >enough headroom to at least stay even with Intel until they can switch to a >newer process. >Oh, of course this entire post assumes 32 bit code. 64 bit code will give the >Opteron a boost (although new HT will give the Pentium a boost). >-Tom Intel will have felt the heat long time before opteron was buyable from it. So i feel that with the prescott they really must have been working drugged to keep going day and night to improve the thing a lot compared to northwood. Basically the potence of Northwood is very big if improved well. Just speaking for 32 bits of course. Because when i talk about opteron as being a major winner i basically consider non-32 bits software from which chess is perhaps the worst example. 64 bits can speed me up only a few % and that still is a hell of effort to do. Trivially biggest speedups are the extra registers and the bigger branch prediction table and a lot of those seemingly small changes that work out great. However if we look from northwood to prescott. Let's start with the trivial changes. That's L1 cache. It gets doubled. Trace cache +33%. already those 2 impacts will be *major* and i'm just talking single threading then, because unless they really put more onto the chip the SMT will be not giving a drastic better speedup other than from the increase in L1 cache. What really is wondering me is whether the new core will do 2 x 2 integers a cycle now. The trace cache simply can't give more than 3 microops a clock which means that the execution of integers on steroids never happens. If they managed to improve *this* and some experts tend to believe that this is possible now to do because the trace cache is modified such that they could only explain it by it being capable of feeding 4 instructions a cycle to be executed. Trivially they need bigger logics for the branch target buffer too. But if that 4 instructions a cycle is possible in prescott then it will be a killer cpu for 32 bits software. Now it has of course a huge disadvantage now to make up for. It's like 1 to 1.6 now. So 1.6Ghz of a northwood equals a K7. AMD employees predicted they would get it to 2.8Ghz that opteron end of life span 0.13, and they said this about opteron, NOT AMD64. Start of 2004 then it would get 0.09. We know the 0.09 is already working there *now*. So they might be this time not be delayed 6 months or so. If that's the case then most likely prescott in 0.09 won't be delayed either. It's the same companies delivering those 0.09 machines... Yet the IPC increase of the prescott will be *huge* with 4 instructions a cycle, especially in the combination of all those changes. If they have done their job right this time, it'll be 30-50% faster than northwood is for me. less for your TSCP i guess. It means that prescott has about 6 months to be sold until 0.09 opteron will kill it for sure, depending upon at which clock 0.09 will be introduced. The story that AMD could easily get Opteron clocked higher is not the entire truth i bet. Probably they had some difficulties with the 0.13 technology. That should be fixed by now. So Q4 2003 that will be spitting out high clocked 0.13s by the millions. Both opterons and AMD64s and K7s. If they manage to release 2.6Ghz opterons or so Q4 then that will kill of course the first few 0.09 Prescotts for 32 bits chess software even (when compiled to 64 bits, 32 bits is meant in this context as: "not profitting from the 64 bits, but it does profit of course from extra registers etc"). Yet the most important fact we must consider is how it will do in the high end. Just when they have managed to put itanium2-madisons at 1.5Ghz for $$$$ with option to go to 1.8Ghz start of 2004, a 2.6Ghz opteron will completely kill it away of course. Even fantastic tricks of HP compiler team won't manage to make up for 0.8Ghz trivially. The itanium just has too many NOPs in each bundle. It can in advance never get therefore a much higher IPC than opteron. 20% more is really the maximum. The more i get confronted with itanic the less i believe in it. Yes it is a good cpu. No if you consider the cost of it and the bugs and the sad future it will have.
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.