Author: Vincent Diepeveen
Date: 13:53:47 08/30/03
Go up one level in this thread
On August 30, 2003 at 01:45:53, Jeremiah Penery wrote: >On August 29, 2003 at 23:45:44, Eugene Nalimov wrote: > >>On August 29, 2003 at 23:18:43, Jeremiah Penery wrote: >> >>>On August 29, 2003 at 22:19:59, Eugene Nalimov wrote: >>> >>>>Sorry, from SPEC Rate I see only that *independent* processes scales well. Any >>>>OS with minimal NUMA knowledge will allocate data for each processor in its >>>>local memory, thus totally avoiding such problems. >>>> >>>>And are you sure that NUMA system that for independent processes "scales almost >>>>as well" as shared-bus shared-memory system is really good achievement? >>> >>>I used the words "nearly as well" without checking. In CINT Rate, Opteron (from >>>what I could tell) at 1.8GHz scales at somewhere over 95% from 1 to 4 >>>processors. Itanium2 scales at 97+% in IntRate. So Opteron is marginally worse >>>here. >>>In CFPRate, however, the situation is much different. Opteron scales (from >>>different sets of numbers) still around 90%, while Itanium2 drops horrifically >>>all the way to near 70% efficiency. >> >>Itanium2 started from much higher initial score, so it saturated available > >It started from higher initial score because its architecture is almost >singularly favorable to FP code. > >>bandwidth earlier. SpecFP applications like lot of bandwidtdh -- trust me... I > >That's really one of the drawbacks to SPECFP, IMO. It relies too much on high >bandwidth. > >>believe hypothetical (much) higher-clocked Opteron with FP performance >>approaching Itanium's will also have worse scaling. > >When adding more CPUs, Opteron's aggregate bandwidth increases linearly with the >number of processors, while Itanium's remains more or less constant (for the >size of systems we're talking about - 2/4 CPUs). > >>In any case, 4-CPU SpecFP rate Itanium2 scores are still much better than >>Opteron scores. And I don't see 16/32/64 Opteron scores at SPEC web site, >>probably you can explain why :-) ? > >Because Opteron is about 2 months old. There were not scores for 16/32/64 CPU >Itaniums until Itanium2 came out, which was at least a year after the original. > >I see 2CPU score on CFPRate of around 13 for first generation Itanium. Should >you base the performance of all Itanium(2) chips by that? > >>Please notice: I don't have anything against Opteron/Athlon64. Very good CPUs >>with solid performance. I just strongly react to all the hype... >> >>>Remember also that this is Itanium2 with good compiler support against first >>>generation Opteron without even decent AMD64 compiler (GCC is not horrible, but >>>ICC 32-bit is still much faster - PGO's compiler apparently blows goats for >>>compiling SPEC for AMD64, see >>>http://www.digit-life.com/articles2/insidespeccpu/insidespeccpu2000-opteron2.html >>>). >> >>Till recently everybody was saying "EPIC needs compiler breakthrough that cannot >>be achieved, while OOO CPUs don't need that breakthrough". And now it happens >>that EPIC compiler already there... > >I did not say that. But it did take a very large investment in new compiler >technology to make Itanium performance competitive. Even now, the 'magical' HP >IA64 compiler produces scores that are far greater than other Itanium compilers. >I still think it has a long way to go before maximizing the capability, too. >x86 compilers are several generations old and still improving, while IPF >compilers are more like 2 generations old. IPF compilers today are not bad, but >they can probably be much better. those ipf compilers get 5.2 gflops out of 1.3Ghz itanium2. I am sure that this is simply the limit it can deliver. At opteron and OoO there is much more to achieve than currently gets achieved, especially when using the same FPU tricks. Just do a simple calculation. itanium2 can execute 2 blocks of 3 instructions each clock. So the situation is very crystal clear. For x86-64 dependant upon the cpu, in this case we just have opteron from AMD and in 2005 (not my guess that it will be released in 2005 i have no opinion there at the moment i just know that if i would be an intel manager i would hurry it up quite a bit more and release a x86-64 cpu in end 2004) we have x86-64 from intel. So it is very interesting to see how the compilers will perform for that platform in 64 bits. But just calculate the potential number of gflops that actually gets achieved by GCC in 64 bits now for opteron and what is theoretic maximum for sure. Then you'll realize very clearly that there is not much to win in IPF when compared to opteron. Best regards, Vincent
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.