Author: Vincent Diepeveen
Date: 07:57:29 08/30/03
Go up one level in this thread
On August 29, 2003 at 16:05:32, Eugene Nalimov wrote:
>Ok, so if we are talking about what is available *right now*, not about some
>future products:
>
>* Itanium2 has absoulute best FP performance, as well as absolut best SpecFP
>rate.
Answer= YEs.
However that is with a special HP compiler which most itanium2 systems do not
get delivered with. Further that takes for programmers 24 hours to compile with
all kind of freak options they don't know.
A default compile at itanium2 is very slow.
Additionally most of the itanium2s are idle, that's why i can run DIEP on such
cpu's a lot.
Majority of libraries needed for that floating point software is not ready yet.
Additionally the compilers achieving high speed for itanium2 are not following
ansi-C conventions which programmers would expect they would follow.
Good example is that for intel c++ at the itanium you have to know that you must
give option -mp to not lose floating point precision:
bash-2.05$ man ecc
NAME
Intel(R) C++ Compiler 7.1
-mp Maintain floating point precision (disables
some optimizations)
So if your research is important and you use floating point, then your results
will be useless without using that -mp option. For sure you have no garantuee
that your program runs bugfree.
If you compile specint again WITH that -mp option with intel c++ then the
floating point achievements of the itanium are NOT so impressive. Most software
is not faster than than it is with other compilers creating correct code.
Note that options from intel c++ at the itanium platform are not the same like
they are on x86 platforms.
Then further i must see the first scientist find this out this -mp option.
Secondly i must find the first scientist on the planet earth who has time to use
PGO with the itanium2.
If you just use pgo for 5 minutes it doesn't speed me up a lot. If i use it
however for 24 hours it will speedup DIEP around 2 times at the itanium2.
Only THEN the itanium2-madison with intel c++ 7.1 is equal to a 2Ghz K7, which
is a great achievement of course.
So your statements about itanium2 are heavily based upon a few assumptions and
the big L3 cache of the itanium2 for spec*.
>* Xeon has very competive price / FP performance for the uniprocessor system. I
>am not sure if Xeon or Opteron is better this week, because all depends on exact
>prices and discounts.
Opteron is 2 times faster for databases, that's for sure. The latency to normal
memory is 2 times faster. You can easily test this at quad xeons versus quad
opterons with Dieter's program or mine. Gives same result.
2 times faster random latency to memory for quad opteron.
Then secondly it is 64 bits, also a big advantage for professional software as
we all know.
>* Opteron and Itanium2 probably have comparable price / FP performance for the
Not at all.
opteron 2Ghz 4 way is $2100.
Itanium2 1.5Ghz is $8000 or something.
The opteron 2Ghz you can buy now. The itanium2 1.5Ghz can't get delivered yet,
it is already very hard to get the 1.3Ghz versions currently.
But let's take a dual opteron versus a dual itanium2. I'm using pricewatch.com
Dual opteron. Motherboard $398 for ARIMA (NUMA capable) rioworks dual
opteorn board. Great board. cpu's $818 for 2.0Ghz opterons. So for $3000 you
have a great system.
NOW a DUAL 1.5GHZ itanium2. Let's estimate it at $20000 shall we?
So opteron is more than a factor 6 cheaper for a dual system.
Also please tell me where i can get that 1.5Ghz itanium2-madison TODAY.
*nowhere* unless i am having test versions.
>multi-CPU systems (SPEC FP rate for the single-CPU Itanium2 is the same as for
>the dual-CPU Opteron, so single-CPU Itanium2 system can be 2x more expensive
>than dual Opteron system).
>For integer performance story is different, but that thread was about FP
>performance, right? :-)
>
>Thanks,
>Eugene
>
>On August 29, 2003 at 14:41:37, Jay Urbanski wrote:
>
>>On August 28, 2003 at 11:37:13, Eugene Nalimov wrote:
>>
>>>On August 28, 2003 at 00:43:34, Jay Urbanski wrote:
>>>
>>>>On August 26, 2003 at 20:02:19, Eugene Nalimov wrote:
>>>>
>>>>>On August 26, 2003 at 19:01:30, Dan Andersson wrote:
>>>>>
>>>>>> Uck! A real bummer since the AMD x87 is a stellar performer.
>>>>>>
>>>>>>MvH Dan Andersson
>>>>>
>>>>>I just looked at the Spec2k, and I did not notice AMD x87 is stellar performer:
>>>>>
>>>>>http://www.spec.org/cpu2000/results/res2003q3/cpu2000-20030728-02428.html
>>>>>http://www.spec.org/cpu2000/results/res2003q3/cpu2000-20030630-02310.html
>>>>>http://www.spec.org/cpu2000/results/res2003q2/cpu2000-20030505-02137.html
>>>>>
>>>>>Thanks,
>>>>>Eugene
>>>>
>>>>Right. Now calculate the price per SPECFP for each of those processors.
>>>>Opteron has *stellar* FP price / performance.
>>>
>>>That's probably true for the single CPU, but if you add system cost (i.e.
>>>motherboard/memory/peripherals) situation changes. Opteron can still win, but I
>>>am not sure.
>>
>>Opteron still does win, by a wide margin. The only way IPF gets close is if you
>>use the new 1.4Ghz 1.5MB cache chip Intel just released as a response to
>>Opteron. The price of a p690 is so high, don't bother.
>>
>>>But if we started talking about "price/performance" why not look at the ordinary
>>>Xeon or Pentium4? They have performance comparable to Opteron (within ~10%), and
>>>very competive price...
>>>
>>
>>For SMP server systems:
>>
>>http://www.spec.org/cpu2000/results/res2003q3/cpu2000-20030714-02406.html
>>http://www.spec.org/cpu2000/results/res2003q3/cpu2000-20030728-02426.html
>>
>>Not even close.
>>
>>If you compare Opteron versus a uni-processor desktop, it is back to
>>competitive:
>>
>>http://www.spec.org/cpu2000/results/res2003q3/cpu2000-20030728-02428.html
>>http://www.spec.org/cpu2000/results/res2003q3/cpu2000-20030616-02266.html
>>
>>But I would argue that's not a fair comparison until AMD releases their Athlon64
>>product.
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.