Intel's heroic v11.1 compiler, and the impact on AMD

[flippin_waffles]

Found an interesting post over at Real World Tech. It looks like this de-optimising effort will never, ever end.

There was something very strange about the SPEC results, however. I knew that Intel's v11.1 compiler was producing 2.5x higher results on 456.hmmer compared to their v11.0 compiler, which suggested something naughty was going on. However, what shocked me is how poorly Phenom was doing on the same benchmark. Specifically, a 3.0GHz Phenom scores 5.90 on the benchmark, while a 2.93GHz Core-i3 scores 40.6.

While everyone is aware that libquantum is a broken benchmark, Intel's new-found 2.5x speed-up on 456.hmmer suggests that they have broken another CPU2006 test. I thought it would be illuminating to re-calculate the INT2006 averages while removing these questionable benchmarks from the results. These calculations are listed as "Subset1" and "Subset2" in the table above.

http://realworldtech.com/forums/inde...08017&roomid=2

[kl0012]

Found an interesting post over at Real World Tech. It looks like this de-optimising effort will never, ever end.






http://realworldtech.com/forums/inde...08017&roomid=2

What will never end, it's blaming Intel by some fans which even don't completely understand what they talking about.
Firstly i wold not put too much attentions in results of amd processors which were submited by intel. (in fact amd did the same by submiting intel's unoptimized results).
Second the result by self still makes sence since libquantum uses RSA crypt and such programs may get big bust on 64-bit systems. So no surprises here because phenom results were compiled for 32-bit os and i3/xeon results - for 64 bit os.

[Serpentarius]

old news ... Intel compiler scandal again .. FTC already wrestling the issue. It'll take time for the court to decide. It always takes time

http://www.xtremesystems.org/forums/...d.php?t=242272

[informal]

Priceless :

A different story applies to hmmer. While Phenom scores an abysmal 5.90 using icc 11.0, the 3 year old Opteron 2222 scores 14.9 using the PathScale compiler, over 2.5x better. For reference, a Core i7-920 scores 15.8 using icc v11.0 on the same benchmark. In other words, AMD is getting hammered on hmmer because icc is producing awful code for Phenom. I suspect the same thing is happening with mcf.

So the 3 year old K8 using PathScale compiler is performing the test 2.5x better than 2nd gen 10h using intel's v11 compiler(scoring 5.9 vs 14.9 K8 is getting)... This is just lame,plain and simple.

[josh1980]

If you don't like it, but an Intel CPU. Then it'll be optimized. /sarcasm

[alfaunits]

Boooring

[Piotrsama]

The guy benchmarking had some flaws, he got some answers, like:

You can't directly compare these results. The Phenom results are on Windows and use the MS libraries. All the other scores are Linux-based scores. Need to have the same environment to draw any conclusions. You would be amazed how much OS & libraries can have an effect in some cases.

And BTW, Intel must (or should) start avoiding this kind of things on their new compilers, after the agreement with AMD.

[kl0012]

And BTW, Intel must (or should) start avoiding this kind of things on their new compilers, after the agreement with AMD.

But before you demand it from Intel, you first need to prove that the flaw is exists. Don't you think so?

[oohms]

So what does a Phenom 2 get when using the PathScale compiler?

[informal]

But before you demand it from Intel, you first need to prove that the flaw is exists. Don't you think so?

Situation normal,all f*ed up? Or SNAFU in military terms

[kl0012]

Situation normal,all f*ed up? Or SNAFU in military terms

I wonder why are you so surprised. 64-bit binares (PGI) on opteron faster then 32-bit binares (intel c++) on Phenom, so what? Seems pretty normal if 64-bit arithmetics is used. Also this is integer apps where k10.5 is not much different from k8 from performance view.

[informal]

2.5x faster on 2 gen older core?You gotta be joking me. 64bit binaries can help a bit,maybe 10-15%,but 2.5x is just not right. Also, 10h is a solid 15% jump in int performance (using the same compiler and available opt. flags).

[Piotrsama]

But before you demand it from Intel, you first need to prove that the flaw is exists. Don't you think so?

Intel agreed not to resort to that kind of tactics when they settled with AMD. That was one of the conditions besides the 1.25 Billion payment.

So then again, Intel should avoid that, because if AMD could prove that new Intel's compilers still do "strange" stuff.... they can go to court again. (but keep the money).

I hope it's clear what I tried to say.

[kl0012]

2.5x faster on 2 gen older core?You gotta be joking me. 64bit binaries can help a bit,maybe 10-15%,but 2.5x is just not right.

You may get even bigger diff where 64-bit integer arithmetics is used. You can see 4-5x perf jump in encryption/decryption code (long integers) and probably in some other algorithms as well. Also Intel used SSE4 compiler option for it's cpus and if you would look at SSE4 specs you will see that packed 64-bit integer instructions is one of the features which SSE4 brings to the table (even for 32-bit programs).

If you look into the results of Phenom 920 (2.8GHz, 32-bit intel c++) and Opteron 2389 (2.9GHz, 64-bit PGI c++) you will see that they pretty similar except two components which probably uses large integers.
http://www.spec.org/cpu2006/results/...608-07784.html
http://www.spec.org/cpu2006/results/...427-07223.html

[saaya]

But before you demand it from Intel, you first need to prove that the flaw is exists. Don't you think so?

no, intel shouldnt upload flawed results, isnt that obvious?
following your logic i could travel over to whatever city you live in and spread photochopped pics of your sucking some guys c0ck and telling everybody what a flamin queen you are, and only stop when you somehow proved that im wrong and the pics are fake. without any consequences for me whatsoever... so then id start spreading pics of you engaging a sheep etc etc... thats pretty much what intel is doing...

intel shouldnt upload amd results, amd shouldnt upload intel results, let each upload whatever they want of their own stuff, to show off where their hardware shines. and second, especially after the recent agreement between amd and intel where intel admitted they were wrong and paid amd, errrrr i mean where intel PROVED they didnt do anything wrong but gave amd a billion bugs cause they are so nice chaps and feel sorry for amd, im surprised that they continue with their compiler bs...

[kl0012]

no, intel shouldnt upload flawed results, isnt that obvious?
following your logic i could travel over to whatever city you live in and spread photochopped pics of your sucking some guys c0ck and telling everybody what a flamin queen you are, and only stopping when you somehow proved that im wrong and the pics are fake.

intel shouldnt upload amd results, amd shouldnt upload intel results, let each upload whatever they want of their own stuff, to show off where their hardware shines. and second, especially after the recent agreement between amd and intel where intel admitted they were wrong and paid amd, errrrr i mean where intel PROVED they didnt do anything wrong but gave amd a billion bugs cause they are so nice chaps and feel sorry for amd, im surprised that they continue with their compiler bs...

1. What makes you think that those results are flawed?
2. I don't remeber that there was too much attention when amd submited suboptimal benchmarks of intel cpus.
http://www.formortals.com/amd-submit...rks-for-intel/

All other stuff about amd/intel legal batle has nothing to do with the subject.

[saaya]

1. What makes you think that those results are flawed?

the article and the comments in here. its beyond my experience and knowledge to judge if the results make sense or not...
i KNOW that they did some ridiculous sabotage acts against both amd and via in the past, and if there is an article about this with most people agreeing and some saying the results are MAYBE ok, it sounds to me like theyve done it again...

2. I don't remeber that there was too much attention when amd submited suboptimal benchmarks of intel cpus.
http://www.formortals.com/amd-submit...rks-for-intel/

no idea... they did that a few times i think and i always thought it was a terrible idea... even IF the numbers are right, whos gonna believe it if it was submitted by the competition? all it does is create drama and hurt the credibility of BOTH amd and intel...

All other stuff about amd/intel legal batle has nothing to do with the subject.

how so? didnt intel promise to "follow several points amd laid out to ensure fair competition"?

they basically said sorry for breakin amds leg, wont happen again, paid some of amds medical bills... and then this is intel nudging amd on the day it leaves the hospital...

[BrowncoatGR]

2. I don't remeber that there was too much attention when amd submited suboptimal benchmarks of intel cpus.
http://www.formortals.com/amd-submit...rks-for-intel/

All other stuff about amd/intel legal batle has nothing to do with the subject.

1. George Ou is a troll
2. I see a lot of speculation and very few facts in that article. AMD clearly did not show the best possible Intel system, but they did show a valid system.
3. It was PR and any company will try to show itself in the best possible light and anyone reading it should always keep that in mind. If you take any PR from any company at face value it's your own damn fault

[saaya]

As for the AMD-Intel agreement iirc JF mentioned it does not include any clauses relating to the compiler

wow... thx for the headsup, i didnt expect that...
thats odd... so its all just about sales?

[Nedjo]

As for the AMD-Intel agreement iirc JF mentioned it does not include any clauses relating to the compiler

actually it does:

Intel shall not include any Artificial Performance Impairment in any Intel product or require any Third Party to include an Artificial
Performance Impairment in the Third Party’s product.

As used in this Section 2.3, “ Artificial Performance Impairment ” means an affirmative engineering or design action by Intel (but not a failure to act) that (i) degrades the performance or operation of a Specified AMD product, (ii) is not a consequence of an Intel Product Benefit and (iii) is made intentionally to degrade the performance or operation of a Specified AMD
Product.

For purposes of this Section 2.3, “ Product Benefit ” shall mean any benefit, advantage, or improvement in terms of performance, operation, price, cost, manufacturability, reliability, compatibility, or ability to operate or enhance the operation of another product

Intel's product that artificially degrades AMD's product performance in this case is 11.1 compiler!

[Particle]

1. George Ou is a troll

But the fact that AMD could have turned in even worse performance numbers for the Intel system is totally irrelevant. If anything, submitting plausible performance numbers on the Intel system is far more insidious because it is a more effective deceit.

It certainly looks that way. "The fact that the results are somewhat fair just makes it that much worse."

Oh please.

The difference, since it seems to escape people like him, is that we rage when somebody goes out of their way to make their competitor's performance look worse. If they just don't put in much effort to improve their competitor's results, that's (usually) fine. It's the difference between active sabotage and inaction.

[JF-AMD]

As for the AMD-Intel agreement iirc JF mentioned it does not include any clauses relating to the compiler

I don't believe that I ever said that, just that I cannot discuss any of the details. However there are legally available copies of the agreement online, not all of the detail, but possibly some of the things that you might find of interest.

[BrowncoatGR]

I don't believe that I ever said that, just that I cannot discuss any of the details. However there are legally available copies of the agreement online, not all of the detail, but possibly some of the things that you might find of interest.

I'm sorry, my memory isn't always reliable.

[zalbard]

The guy benchmarking had some flaws

This.
Seriously, you can't test that way...

[kl0012]

the article and the comments in here. its beyond my experience and knowledge to judge if the results make sense or not...
i KNOW that they did some ridiculous sabotage acts against both amd and via in the past, and if there is an article about this with most people agreeing and some saying the results are MAYBE ok, it sounds to me like theyve done it again...

This is the problem! Intel was blamed again for using "deoptimization" in compiler for amd system while no one actually tried to analize the results.
Yet i agree with you that not Intel, nor AMD should submit test result for competitors systems.

2. I see a lot of speculation and very few facts in that article. AMD clearly did not show the best possible Intel system, but they did show a valid system.

Well, this can be said about the subject article as well. It is clear that results submited by Intel are not "best optimized", but this doesn't means that they are flawed or moreover - "deoptimized".