Monday, September 11, 2006

K8 crushes Conroe/Woodcrest on floating point performance

It's old news. But now SPEC.org published the SPECFp scores for Socket F and Socket AM2 Opterons. The 2P 4 core 2.6GHZ Sun Fire x2200 M2 got a SpecFp_rate2000 score of 117. A 3GHZ dual Woodcrest 5160 got 80. That's a 46% lead by AMD.

In other news, Dell delayed its 10Q filings as SEC probes its past reporting. I repeatedly told Michael Dell to be careful there since sometime last year. When your company keeps winning and growing, everyone sings praises. When your company starts downhill slide, everyone will step over you. Enron folks went to jail because Enron crashed, not because what they did was particularly worse than others.

The only way for DELL to avoid Enron scenario is to find massive growth in AMD market.

In futher news, EU has started investigating Intel. As I estimated, Intel owes AMD $15 billion minium in damages. At the end, I think those Intel FABs belong to AMD.

27 Comments:

Blogger "Mad Mod" Mike said...

The Operating Systems & Compilers are different. AMD64 CPU's always do great with the Solaris OS, so I have to say this is just proving the point that AMD CPU's work best in environments optimized to them, same as Intel CPU's. 64-Bit Solaris 10 is about the best OS to use on an AMD platform, you'll see HUGE gains in performance (as proven by SPEC).

4:34 PM, September 11, 2006  
Anonymous Anonymous said...

Congrats AMD, you win another synthetic besides memory bandwith:)

5:27 PM, September 11, 2006  
Anonymous enumae said...

I am wondering how you can compare two systems that have a different amount of ram and ram speed and claim a victory?

What about SPECint?

Conveniently left those out...

IBM System X 3655 (AMD Opteron (TM) 2218)

PRIMERGY TX300 S3, Intel Xeon 5140 processor, 2.33 GHz

In those test AMD has more ram.

So does this show you how selective you are about AMD?

5:28 PM, September 11, 2006  
Blogger "Mad Mod" Mike said...

The only tests you can compare are ones with SAME RAM, SAME OS, SAME COMPILER, and SAME APPLICATION. If those parameters are met, there is less that can go wrong or be disputed (not to say a moron didn't put the systems together).

6:04 PM, September 11, 2006  
Anonymous Anonymous said...

"At the time, they also said they did not have enough evidence to pursue an AMD complaint on microprocessors."


"AMD filed another complaint in 2004 that EU officials said they had no choice but to investigate - or risk AMD taking court action for negligence."

Little evidence and AMD being whiney finally lead to the investigation.

6:28 PM, September 11, 2006  
Anonymous Anonymous said...

"I repeatedly told Michael Dell to be careful there since sometime last year"

Um, yeah right. Like a man who built a business from nothing to the largest computer seller in the U.S. is going to listen to someone like you.

I firmly believe that you are either

a) insane
b) manic
c) on some serious drugs
d) 14 years old
e) All of the above

6:37 PM, September 11, 2006  
Blogger Sharikou, Ph. D said...

Like a man who built a business from nothing to the largest computer seller in the U.S. is going to listen to someone like you.


You can say the same about Jeff Skilling and Ken Lay. Their success was even more spectacular... You can also say the same for that WorldCom dude...

They are all in jail now.

6:48 PM, September 11, 2006  
Anonymous Anonymous said...

I am totally convinced that AMD socket-F with DCA (Direct Connect Architecture: Hypertransport + on-chip memory controller) will take over the super-computing market from Intel.

Until Intel introduces CSI.... We don't need to debate on this!

-Longan-

7:00 PM, September 11, 2006  
Anonymous Anonymous said...

Ken Lay is dead and Skilling is awaiting sentencing..

7:09 PM, September 11, 2006  
Anonymous Anonymous said...


The only tests you can compare are ones with SAME RAM, SAME OS, SAME COMPILER, and SAME APPLICATION. If those parameters are met, there is less that can go wrong or be disputed (not to say a moron didn't put the systems together).


If Spec favors someone this one is Intel, they have their own compiler wich by default create a very otimized code for Intel CPUs, Solaris runs well on Opteron? Yes, good for AMD, but ICC gives a very big advantage for Intel,

The amount of RAM don't make diference, the SPEC uses about 200-300MB the extra memory isn't there to increase performace due to bigger size, but to fill all channels (increasing the performace due to the otimal bandwidth...).

And finally... 117 for Cfp2000_rate is impressive, this 2P system beats 2P Woodcrest, 4P Tulsa and old socket 940 2P Opterons.

At least as impressive as Woodcrest Cint2000.

7:54 PM, September 11, 2006  
Anonymous Anonymous said...

Enron guys were not more successfull than DELL even at their peak. Dell is a multi billionare. Enron guys were not.

But something does seem fishy with Dells balance sheet. I wonder if this will turn into another Worldcom, Enron accounting scandal?

8:51 PM, September 11, 2006  
Anonymous Anonymous said...

YOu are misquating. For comparisons, base scores are supposed to be used.

http://www.spec.org/cpu2000/results/res2006q3/cpu2000-20060721-06538.html

So you should be comparing the Sun Fire's 101 to Woodcrest's 81.5. (The Dell system you picked is slower than Fugi's.) The reason why you pick base as a comparison is because peak scores use custom optimizations while base scores do not. Therefore if you want to compare anything you should use base rather than trying to skew things with peak vs. base.

8:57 PM, September 11, 2006  
Anonymous Anonymous said...

When you can't beat them sue them.

Good strategy Mr PhD..

AMD will be bankrupt so use the legal system to extort money from the company that kicked AMD's butt

What a company, what benchmarks, what a bunch of lawywers, what a bunch of crybabies AMDers are.

Just go out and beat INTEL. Let your benchmarks, sales, profits say it all. Who needs a blogger flogging a fantasy. Go out and do it AMD, show us them profits, and benchmarks we await..

8:58 PM, September 11, 2006  
Anonymous Anonymous said...

Anonymous said...

Ken Lay is dead and Skilling is awaiting sentencing..

AND... look where there past got them. lol

1:33 AM, September 12, 2006  
Anonymous Anonymous said...

"But something does seem fishy with Dells balance sheet. I wonder if this will turn into another Worldcom, Enron accounting scandal?

I seriously doubt it. Enron was a supplier of services, while Dell is a supplier of (mostly) physical goods. Enron were able to fiddle their accounts better, because it was harder to keep track of what they were or weren't supplying. If Dell tried to pull anything like that, they'd be found out straight away once it was realised that their sales figures and recorded shipments didn't match up.

2:06 AM, September 12, 2006  
Blogger Sharikou, Ph. D said...

Enron were able to fiddle their accounts better, because it was harder to keep track of what they were or weren't supplying.

Wrong. I analysed Enron BK in this blog and compared Enron to DELL. DELL is in a worse situation than Enron. Enron had $9 billion of net assets even after BK. DELL now has only $3 billion net assets. All Enron did was hiding $2 billion of losses, which caused the stock to crash and unable to borrow more...
DELL is in a more fragile situation.

2:17 AM, September 12, 2006  
Blogger Sharikou, Ph. D said...

Enron was a supplier of services, while Dell is a supplier of (mostly) physical goods.

You are on this too. Enron owns and operates a lot of energy companies. If Enron was only doing paper goods, then its risk would be small. Same applies to DELL. Imagine a situation DELL bought $2 billion Conroe CPUs, suddenly K8L floods, all those Conroes worth 0... BK

2:20 AM, September 12, 2006  
Anonymous Anonymous said...

"You are on this too. Enron owns and operates a lot of energy companies. If Enron was only doing paper goods, then its risk would be small. Same applies to DELL. Imagine a situation DELL bought $2 billion Conroe CPUs, suddenly K8L floods, all those Conroes worth 0... BK"

A guy living in a blog bubble doesn't know the world.

Just because a faster chip comes out, it doesn't mean the value of the slower chip is suddenly zero.

A moderate reduction in inventory value does not result in a BK.

What I find disappointing in this blog is that there is no intelligent discussion beyond "this chip is faster than that chip". It is stuck at the mentality of a 5 year old.

I find it laughable that with 4 cores, 4GB, and 2 16x12 displays, the dipshits at Google still make me type text into a box roughly 2" x 2.5". No matter how fast your chip goes, or how many you have, you are limited by the sheer stupidity of man.

The insurmountable idiocy that drives the tech business is far more interesting blog fodder than "my Daddy's chip is faster than your Daddy's chip! Waaaaahhh!".

2:39 AM, September 12, 2006  
Anonymous Chin Gim Leong said...

"YOu are misquating. For comparisons, base scores are supposed to be used.

http://www.spec.org/cpu2000/results/res2006q3/cpu2000-20060721-06538.html

So you should be comparing the Sun Fire's 101 to Woodcrest's 81.5. (The Dell system you picked is slower than Fugi's.) The reason why you pick base as a comparison is because peak scores use custom optimizations while base scores do not. Therefore if you want to compare anything you should use base rather than trying to skew things with peak vs. base."

Have you ever wondered why all the Intel SPEC scores are identical in base and peak? Is it even possible that there are no more possible optimizations to be made for the peak runs? And that the Intel compilers have so few additional options even for peak? Or is it that the compiler has already made all the possible optimizations at the so called "base" setting without the need to set explicit options? Actually, the SPEC "base" or peak setting is nothing more than a game where the number and type of compiler flags allowed differ. There is nothing in the rules that say that optimizations not explicitly specified by user cannot be quietly done by compiler. And I have suspicions that Intel compilers have some special "tuning" for SPEC benchmarks.


GL

4:10 AM, September 12, 2006  
Anonymous Anonymous said...

"Until Intel introduces CSI.... We don't need to debate on this!"

Are you serious? CSI won't be in Xeons until 2009. This is just Intel sdmitting that their architecture is antuquated. Intel hopes to get in 2009 what AMD has has since 2003.

5:25 AM, September 12, 2006  
Anonymous Anonymous said...

The "base vs peak" is a long discussion...
ICC does a good otimization for Intel CPUs, the Sun Studio used for Opteron is a generic compiler,
Comparing the 100 base points of Opteron with the 80 base points of Woodcrest is a fair comparison?
And if AMD buy Sun and turn the default flags of Sun Studio to help Opteron and Opteron gets, let say, 110 points, still fair? 10% improvenment and i only changed the owner of compiler...

For such comparisons i think the peak is more accurate... A simple rule, "What's the best your system can do?".

BTW, the 100 points of the 2.6GHz Opteron still bigger than 80 (ok 90 to make Intel happy...) of the 3.0GHz Woodcrest.

6:28 PM, September 12, 2006  
Anonymous Anonymous said...

Have you ever wondered why all the Intel SPEC scores are identical in base and peak?

You really should look at some Intel scores then. It's only when people don't bother optimizing further that the peak score equals the base.

http://www.spec.org/cpu2000/results/res2006q2/cpu2000-20060612-06165.html

For this Xeon 5160 system by HP the SPECfp_rate_base2000 is 81.1 and the peak is 85.9. A decent 6% increase. For reference the highest 5160 base seems to be 81.5 by Fuji which is higher than HP's base, but lower than their peak.

http://www.spec.org/cpu2000/results/res2003q4/cpu2000-20031020-02593.html

For this HP Itanium 2 the SPECfp_rate_base2000 is 126 and the peak is 142 or nearly a 13% increase.

It's true that generally Intel systems builders don't bother spending time to try to create optimized peaks. Why I'm not sure. HP seems to though, but even they don't seem to try very hard. It's really too bad that Fugi doesn't spend time trying to optimizie Intel peaks since their Intel base scores are usually the highest compared to other identical processors.

9:52 PM, September 12, 2006  
Anonymous Anonymous said...

SPEC.ORG introduced the new Spec2006 suite of benchmarks. A few benchies were already published on spec.org site. Most of them have Opteron, very few Xeons, and about to none Woodcrest.
Why? May be in reality WC is not able to compete with Opteron in real world benchmarks?
uf.

1:40 AM, September 13, 2006  
Blogger Kalle said...

"It's true that generally Intel systems builders don't bother spending time to try to create optimized peaks. Why I'm not sure."

Because they don't care of the results given by synthetic benchmarks? Nobody will ever be that stupid to recompile production software with the same compiler flags as the benchmark. Compilers tend to break stuff with lots of optimization flags.

Btw, does anyone know if those FP tests use the FPU stack or SIMD registers for computation? In 64bit, the FPU stack is depricated and generally shouldn't be used since SIMD ones are considerably faster. I have a feeling that as this benchmark is several years old it doesn't use the optimal units for doing its calculations.

4:36 AM, September 13, 2006  
Blogger Kalle said...

"Have you ever wondered why all the Intel SPEC scores are identical in base and peak? Is it even possible that there are no more possible optimizations to be made for the peak runs? And that the Intel compilers have so few additional options even for peak? Or is it that the compiler has already made all the possible optimizations at the so called "base" setting without the need to set explicit options?"

I missed that part last time.
Intel compiler has lots of tuning flags. I've seen FP intensive SIMD code to get ~20% faster just by switching from -O2 to -fast. The latter is basically -O3 with some extra flags. Of cource such huge improvements are rare.

ICC's base flags are rather agressive. I've used 9.0 and it defaulted to generating -O2 P4 specific code with alternative codepaths for other CPU's. At that time it was quite a bit faster in 32bit than GCC. But in 64bit it is mostly somewhat slower than newer versions of GCC, especially if you use the prealphas from 4.2 tree. MSVC is by far the slowest of the three.

Too bad that test costs lots of money. I'd try to use Acovea on it to find out the best possible compiler flag setup :)

Also I'd like to know when will we see more results from the new SPEC tests. The 2000 versions will be depricated soon

6:24 AM, September 13, 2006  
Blogger Kalle said...

"Why? May be in reality WC is not able to compete with Opteron in real world benchmarks?"

There are no Socet F Opterons too. That must mean it can't compete with Netbursts, Sparcs and old DDR1 Opterons.
[/irony]

10:02 AM, September 13, 2006  
Anonymous Chin Gim Leong said...

"You really should look at some Intel scores then. It's only when people don't bother optimizing further that the peak score equals the base."

OK, I did not look at the scores of all Intel systems, but the throughput "rates" scores submitted by Intel are either identical or almost identical for base and peak. Surely Intel knows best how to optimize for their own compiler and have good motivation to do so?

As for IA64, the scores submitted by SGI for their Altix are identical for base and peak, if they bother to do for peak. As a supercomputing company that makes its own MIPS and compilers, SGI folks of all people definitely know how to optimize their systems to get higher scores. Yet they chose not to do so for all the more recent IA64 CPUs.

Thanks Kalle, for your information on the agressive default behaviour of Intel compilers.


GL

11:24 PM, September 13, 2006  

Post a Comment

Links to this post:

Create a Link

<< Home