P4 SpecCPU2k results

[Guest]

SpecCpu2k results on intel web page (http://www.intel.com/procs/perf/Pentium4/brief/docs/pentium4.pdf) seem to be unbelieveable high. Still I believe they are for real, because Intel just cannot afford posting fabricated SpecCpu benchmark results of P4. If you take time to compare these results with other results in the SpecCpu results database(http://www.spec.org/osg/cpu2000/results/cpu2000.html), you can get some kind of picture, what P4 really is capable of. Only high-end Alpha processors are able to compete and surpass P4 in both integer and FP performance.

In my opinion SpecCpu is respected and reliable meter of system level real application performance. So, why P4's performance is so good in Spec bench and so bad in 'legacy' benchmarks, which we have seen on every tech web site last days. One big reason is of course the compiler. Spec bench is compiled with P4 optimizing compiler and we can see huge potential performance gain when optimizing is used. So a big question remains, when will we begin to see P4 optimized software in large scale. At easiest, optimizing can be just compiling source code with the new compiler, which should not be too big task to do for new software developers. Also researchers who need great computational power and are developing their own applications and algorithms can readily take advantage of the optimizing compilers making P4 very attractive computational workstation.

P4 has great potential, which waits to be unleashed by software optimizations. For the time being I cannot see any reason why home user should upgrade to P4, mainly because of high price and non-optimized applications. This will change in next two years, and by then also the prices have gone down quite a bit. Some specialised tasks (for example scientific computations) may allready justify the high price of P4 (which is not that high after all if you compare to Alpha/SUN workstations).

<P ID="edit"><FONT SIZE=-1><EM>Edited by Zam on 11/23/00 08:45 AM.</EM></FONT></P>

 

[Guest]

.. I thought I saw Spec 2000 graphs for PIV, PIII, and Athlon on one of the independent review sites, now I can't remember which one, any one recall where they are?

.... The high SpecFP score for PIV is caused by SSE(2) in the drivers/compilers and is NOT representative of real world performance, that's why PIV falls on its face anytime you try to just run an app. at random on it. Its just a trick to fool prospective buyers.

.... Scientific apps. also run faster on Athlons, unless tricked up for PIV. If you want a PC that you can just slap any ol' app. in and get MAXIMUM performance, get an Athlon, and save a bunch of $$$$$, to boot.

 

[Guest]

Not only that, but x87 instructions are 80 bits vs 64 bits for SSE-2, meaning that in calculations that need a lot of precision, you're better off with an athlon.

 

[Guest]

I include reply to 2 above messages.
First reply to BuzzLOL:

>>.. I thought I saw Spec 2000 graphs for PIV, PIII,
>> and Athlon on one of the independent review sites, now
>> I can't remember which one, any one recall where they
>> are?

It would be nice to see those benchmarks. But I still believe that results on Intel www site are not fake. And believe me those results are just wonderfull, just go and compare to Alpha processor results on Spec database.

>>.... The high SpecFP score for PIV is caused by SSE(2) in
>>the drivers/compilers and is NOT representative of real
>>world performance, that's why PIV falls on its face
>>anytime you try to just run an app. at random on it. Its
>>just a trick to fool prospective buyers.

Of course better FP performance is due to SSE2, but this does not mean it is not real world performance. Of course we have the problem that we must have optimized software to get SSE2 benefit. SpecCPU is good real world system level berformance indicator (go and ask your local computer engineering professors).


>>.... Scientific apps. also run faster on Athlons, unless
>>tricked up for PIV. If you want a PC that you can just
>>slap any ol' app. in and get MAXIMUM performance, get an
>>Athlon, and save a bunch of $$$$$, to boot.

The performance/price ratio is definitely favourable for Athlon, and I myself would go for ~1GHz Athlon. P4 must prove it's worthy when SSE2 optimized applications appear and Intel should be aggressive in price reductions to make P4 competitive. Of course there is problem, because P4 has bigger die size and is more costly to manufacture.


Reply to Seoman:

>>Not only that, but x87 instructions are 80 bits vs 64
>>bits for SSE-2, meaning that in calculations that need a
>>lot of precision, you're better off with an athlon.

Both int and fp SSE2 arithmetic are 128 bit SIMD (single instruction, multiple data). Where did you get your information?



<P ID="edit"><FONT SIZE=-1><EM>Edited by Zam on 11/25/00 05:49 AM.</EM></FONT></P>

 

[Guest]

Here is the explanation for your question

<A HREF="http://www.geocities.com/tnaw_xtennis/Pentium4-4.htm" target="_new">http://www.geocities.com/tnaw_xtennis/Pentium4-4.htm</A>

 

[Guest]

I heartily disagree with the "SpecCpu is respected and reliable meter of system level real application performance"... It is an indicator of CPU/compiler/cache-memory subsystem performance. I/O is completely out of the picture. Most high-end Unix servers spank the IBM S80 on SPECintRate, but nobody is close on TPC-C, which more closely models an actual database application.

Also, I wonder why the results haven't appeared on spec.org yet. I'm wondering if maybe Intel is reporting figures for the SSE system that wouldn't pass muster if formally submitted to SPEC. Does SSE use fully IEEE compatable floating point with all IEEE exceptions, etc?

 

[Guest]

hello, zam,

do you have any Athlon/Alpha optimized SPEC data? i am not saying the data Intel provides on their website is fabricated, but surely they've been optimized, right? however, the data on SPEC's website were from plain benchmarks without any oriented optimization. this means that you pick any Alpha workstation on the market and it will give you that performance right out of the box without any new compiler or optimization applied.

if P4 could really beat Alpha in real life performance, Intel should have stopped "wasting" money on their 64bit projects, and then Geee, Intel should rule.



-----------------------------
Some are ignorantly happy,
While some, happily ignorant.

 

[Guest]

>>I heartily disagree with the "SpecCpu is respected and
>>reliable meter of system level real application
>>performance"... It is an indicator of CPU/compiler/cache-
>>memory subsystem performance. I/O is completely out of
>>the picture.

You are right about IO subsystem performance not being important part of SpecCPU benchmark. And specCPU not being specifically system level performance indicator. Of course IO system also contributes to the results, but it is not stressed to the max by SpecCPU tests. The same goes with networking and graphics.

Compilers seems to play very important role, and it seems that Intel is making great job with optimizing compilers, leaving one question open: is average software developer Joe able to benefit from these high end Intel compilers (I remember hearing some bad words about them). I have read articles about specCPU using mostly large datasets(which would give P4 edge over other systems with less memory bandwith). However this would suggest P4 being good choise for large dataset problems (scientific ones, maybe).
It seems SpecCPU2000 contains more exotic scientific applications than normal word processing and gaming.

I wellcome everyone to find out more about SpecCPU2000 by following this link:
http://www.spec.org/osg/cpu2000/press/faq.html

 

[Guest]

Hello Frankel,

>>do you have any Athlon/Alpha optimized SPEC data? i am
>>not saying the data Intel provides on their website is
>>fabricated, but surely they've been optimized, right?

No I do not have Thunderbird SpecCPU2k results, but I expect them to be great, maybe not to the extent of P4 results, but definitely better price/performance. I personally like Thudernbird and current AMD processors pretty much, and don't have any Intel/AMD love/hate attitide. I just think P4 SpecCPU2k results are very impressive and it makes also P4 very interesting processor. Overall PC processors (both AMD and intel) have evolved pretty fast lately bypassing some 'bigger' workstation systems in performance. They seem to gaining more ground from higher end systems in workstation and server space, just due to competitive pricing and huge performance gains last few years.

>>however, the data on SPEC's website were from plain
>>benchmarks without any oriented optimization. this means
>>that you pick any Alpha workstation on the market and it
>>will give you that performance right out of the box
>>without any new compiler or optimization applied.

You got valid point. It seems to be P4's problem now, that it needs optimized software to really show its strenght. It remains to see if Intel can 'force' developers to use P4 optimizing compilers, personally I think they can.

>>if P4 could really beat Alpha in real life performance,
>>Intel should have stopped "wasting" money on their 64bit
>>projects, and then Geee, Intel should rule.

P4 is far a away from beating Alpha processors. Current situation just seems (when judging by SpecCPU2k) that Intel is closing the performance gap pretty fast. Of course Alpha crew is not resting, but they are pushing for more performance and new platform and architectures, and they REALLY have great potential with Alpha architecture. P4 has also big platform and chipset improvements over the P3 systems, for example the quad pumped bus, which can really make difference in certain applications. Alpha has very mature processor and platform architechture and Intel will have hard time competing Alpha. But if they can, it is just consumers advantage, because Intel can maybe lower the prices of current state-of-the-art-performance systems.





<P ID="edit"><FONT SIZE=-1><EM>Edited by Zam on 12/03/00 11:53 AM.</EM></FONT></P>

 

[Guest]

I have been looking at the SPEC numbers for Dell at the spec website, comparing a 1ghz P3 machine with 1.5 ghz P4
SpecInt 1ghz P3 = 418, P4 = 493 (18 % increase)
SpecFP 1ghz P3 = 292, P4 = 543 (86 % increase)
The Alpha (833 Mhz) has a specint of 511 and 571 in specfp. The P4's specfp scores are very close to the high end Alpha machine, which had 8 M.B of secondary cache, compared to 256kb in P4. It looks like Intel focused on floating-point (SSE2) in the design. I am thinking P4 may be aimed at the non-ia workstation market. the specint improvements could possibly just be due to the move to bigger bus and from 4.5 to 5.0 compiler. Anyway the floating point scores look impressive.

sv1104

 

[Guest]

Is same FP precision used, comparing Alpha, PIII and P4? If thats unknown, the test is worthless.

---
Engage!

 

[Guest]

Generally speaking most calculations use the double
data type which is 64-bits long. The 80-bit extended
type hurts performance a great deal and is rarely used
(from the source code I have encountered, that is).

In my opinion, SSE2 at 64-bit is perfectly valid for
high-end math unlike 3dnow (32-bit) which is good for games.

The only potential problem with SSE2 is that it has to
be implemented and we don't know how long it will take for
that.

Petros


Sum of intelligence constant. Population increasing.

 

[Guest]

Here is a link of an article that was published in the IEEE Computer magazine on spec2000 : http://www5.compaq.com/AlphaServer/download/ieee_spec2000_article.pdf

sv1104

 

[Guest]

The results are now on the web page and they kick butt. They are all 500+. Intel has posted their results with great numbers. To back it up, there are scores of dell workstations that are just as good. This processor looks very promising!

http://www.spec.org/osg/cpu2000/results/cpu2000.html

 

[Guest]

Good question, re: can everybody benefit from Intel's optimizations. I doubt it. Everybody cheats on benchmarks, but Intel is VERY good at it. They usually only report peak rather than base results and use feedback-directed optimization. Fine for compiling SPEC, not so easy on a full Linux kernel compile.

In any case, my previous question has been answered-- Intel now has published results for the P4 and they are very good

The below has benchmarks for many machines, not just x86
http://www.ideasinternational.com/benchmark/bench.html

This one has results for P4 on the specbench site
http://www.specbench.org/osg/cpu2000/results/res2000q4/

 

[cipher]

remember when the P3 came out, lots of software posted patches, like photoshop to use the extra cpu instuctions. What makes you think this wont happen with the P4s new instructions?