SpecInt/SpecFP - Intel vs AMD

[Raystonn]

These results speak for themselves:
http://www.spec.org/osg/cpu2000/results/cpu2000.html

For those who don't feel like searching the table:

SpecInt (Integer performance test):
Company Name System Name Base Peak
Intel Corporation Intel D850GB motherboard(1.5 GHz, Pe 524 536
Advanced Micro Devic Gigabyte GA-7DX Motherboard, 1.33GHz 482 539
Intel Corporation Intel D850GB motherboard(1.3 GHz, Pe 473 483
Advanced Micro Devic ASUS A7V Motherboard, 1.3GHz Athlon 438 491

SpecFP (Floating-Point performance test):
Company Name System Name Base Peak
Intel Corporation Intel D850GB motherboard(1.5 GHz, Pe 549 558
Intel Corporation Intel D850GB motherboard(1.3 GHz, Pe 503 511
Advanced Micro Devic Gigabyte GA-7DX Motherboard, 1.33GHz 414 445
Advanced Micro Devic ASUS A7V Motherboard, 1.3GHz Athlon 348 374

---------------

The Athlon's floating point performance is looking rather pathetic. The 1.3GHz P4 beats the 1.33GHz Athlon!

The Athlon's integer performance is more robust. Averaging the base and peaks together, the 1.3GHz P4 still beats the 1.3GHz Athlon!

Be aware that this is an industry accepted benchmarking application. Spec is the standard by which all CPUs are judged. There is no bias. AMD gets to run their own benchmarks on their own equipment and submit the results.

---------------

The only barrier to owning a P4 at this point is price. And that barrier will be coming down before the end of this month. I look forward to it.

-Raystonn


-- The center of your digital world --

 

[kurokaze]

Hey I've seen these before..

THIS IS NOT A BASH AGAINST INTEL, I repeat THIS IS NOT A BASH. But I just want to point out Spec's disclaimer:

The results published by SPEC have been reviewed by the SPEC organization prior to publication. However, these are submissions by member companies and the contents of any SPEC reporting page are the submittor's responsibility. SPEC makes no warranties about the accuracy or veracity of this data.

As always, come to your own conclusions.

 

[Raystonn]

This is a highly respected, industry accepted benchmark. Are you trying to make the libelous statement that Intel fudged the numbers? I don't think so. Look at the numbers submitted by other companies using the same CPUs, they are close to the same numbers. Are you now saying that they _all_ fudged their numbers? I wouldn't bet on that.

These are real scores, not fakes. If I were to pick and choose which benchmarks were likely to contain mistakes, it would be those done by third parties.

In addition, you'll see legal disclaimers on _all_ benchmarks indicating that the numbers don't necessarily mean anything. It's similar to what you see on late night television... "For entertainment purposes only." This is only because noone wants to be sued.

-Raystonn


-- The center of your digital world --

 

[kurokaze]

Why is it that simply making people aware of the disclaimer is immediately translated into an attack??

Holy F*cking Sh*t.....

 

[Kelledin]

Nice...if SSE2 catches on, then AMD will be back where it was with the K6-2. Unless, of course, AMD gets their own SSE2 implementation working by that time...

In any event, SSE2 will definitely be an improvement over the old stack-based FPU.

Kelledin
<font color=red>"Step away from the gimp suit and put your hands on top of your head."</font color=red>

 

[Raystonn]

Because you happened to point out the disclaimer on this one benchmark without reminding everyone that a similar disclaimer is on every other benchmark they'll ever read. You insinuated that this disclaimer somehow affects the validity of the results of this benchmark. If this wasn't your intention then I apologize, but ask why you pointed it out. It's no different than the disclaimers you'll read on all the other benchmarks.

-Raystonn



-- The center of your digital world --

 

[Grizely1]

Oh yes all high Raystone, you have brainwashed us and we will now go and buy a P4 to run SPEC benchmarks all day and see those beautiful fake number! All hail Raystone!!

-----------------

All your RAMBUS are belong t............ ahh screw it

 

[AmdMELTDOWN]

griz, can you post a link to back up your claim that the spec bencmarks are fakes?

"Amd cpu...Gone in 2 secs flat, it truly is a fast chip!"

 

[kurokaze]

because of the fact I myself am familiar with the benchmark and know its acclaim. I am merely trying to make sure people aware of the fact that it too is not perfect.
(in fact, I used these very same spec results in determining my next CPU, P3 or Athlon)

Also, the reason I don't quote disclaimers for other benchmarks is that everyone should be taking those
numbers with a grain of salt anyway.

<P ID="edit"><FONT SIZE=-1><EM>Edited by kurokaze on 04/12/01 11:58 PM.</EM></FONT></P>

 

[Raystonn]

That's a nice troll there, Grizely1. At least I hope that was a troll. I don't see how anyone could take your claims seriously considering you can't even spell my name correctly. These are not fake numbers. I assure you they are quite real. If you'd like to test it yourself, feel free to download the spec benchmark and run it on a P4. Until you show the numbers to be off, we can safely assume that the numbesr submitted by multiple companies with P4 machines are right on.

-Raystonn

-- The center of your digital world --

 

[Grizely1]

Jeez louise you people have a bad sense of humor LOL! I make one little comment and you go bezerk protection your beloved P4 shroud from evil

-----------------

All your RAMBUS are belong t............ ahh screw it

 

[Guest]

The one question I've had on these is if SSE2 optimizations were used, or if you can just take an off-the-shelf compiler and get these numbers. Nothing wrong with SSE2, except that I don't know how to code it personally :-(

I welcome the price cuts also, although I would wait for the new socket.

In theory, there is no difference between theory and practice.
In practice, there is.

 

[Raystonn]

I'm sorry for jumping all over you if that was a joke. There are way too many flamers on this board and you can't tell who's being serious. I just want a serious discussion on the merits of the CPUs, that's all.

-Raystonn

-- The center of your digital world --

 

[Raystonn]

These results were obtained compiling with the Intel C/C++ Compiler 5.0 and Microsoft Visual C++ 6.0 (for libraries). These are both commercially available compilers. You don't have to do anything special to use SSE as the Intel compiler will automatically build it with SSE if you enable this option..

The common method of CPU optimizations is to use function pointers or a class hierarchy that allows you to duplicate the important code into multiple source files, and then build each source files targetting a different CPU. You then link them together into the resulting executable and you choose which function/class hierarchy to use at runtime based on the current CPU.

This allows you to target every CPU fully, simply recompiling the important core module once for each CPU. A simple compile of the one .cpp file with the Intel compiler will optimize it for you, adding SSE if you desire, for the Intel CPU you select.

An interesting note is that binaries (programs) built with the Intel compiler are better optimized for AMD cpus as well (with SSE turned off). This is shown by the fact that AMD used the Intel compiler for their benchmarking as well. (Go to the page listing the results and click on the 'html' link to the right of the score.)

-Raystonn

-- The center of your digital world --

 

[Guest]

My understanding is that compilers can do some basic SSE optimizations, but for more complicated code you generally have to get down into assembly language by hand. That's my question really. Is this straight from an off-the-shelf compiler, with no in-line assembly.

I've seen header files abstracting some 3DNow assembly, but it's still pretty ugly. I hope to learn more now that I'm not working on an antique cpu

Besides, Intel compilers aren't available on my platform of choice.


In theory, there is no difference between theory and practice.
In practice, there is.

 

[Raystonn]

As opposed to 3DNow, this is straight off the shelf C/C++ code with no inline assembly required. The reason for the inline assembly with the AMD toolkit is AMD doesn't make a compiler. Using MMX/SSE/SSE2 is much easier. It's all natively built into the compiler.

By chance is your platform of choice Linux? The Intel C/C++ Compiler for Linux is going into Beta in May of this year. All the linux coders are going to want this one. "gcc" has really crappy optimizations thus far for anything beyond the Pentium.

-Raystonn

-- The center of your digital world --

 

[Guest]

You got me interested, so I went poking around Intel's site last night. One thing I'll say for Intel, they have some excellent development support wrt AMD. Now if they would work on the price/performance ratio a little ... but I digress.

I did find the announcement for the Intel C/C++/Fortran compiler beta for Linux! Signed up also. Very cool. What's the price range on Intel compilers? Gnu is very good technically in my experience (well, not the fortran), and its free, but its certainly not the fastest thing out there. We've been looking at Portland Group's compilers.

>As opposed to 3DNow, this is straight off the shelf C/C++
>code with no inline assembly required. The reason for the
>inline assembly with the AMD toolkit is AMD doesn't make a
>compiler. Using MMX/SSE/SSE2 is much easier. It's all
>natively built into the compiler.

Well, I'll have to disagree with you somewhat here. I downloaded the SSE2 tutorial from Intel (http://developer.intel.com/software/products/itc/sse2/sse2down.htm). I haven't read through the whole thing yet, but it definitely seems that you need to code specifically for SSE2, either in assembly, or through some of Intel's C/C++ abstractions. I'm sure the compiler optimizations take some advantage of SSE2, but the tutorial definitely emphasizes hand coding.

This is also in line with one of Tom's early P4 reviews with the whole Flak mpeg stuff:
http://www4.tomshardware.com/cpu/00q4/001125/p4-02.html
http://www4.tomshardware.com/cpu/00q4/001125/p4-07.html

While re-compiling certainly helped, the big gains were made by re-coding the algorithm with Intel's SSE optimized vector class.

So, you may not have to learn assembly, but its more then a simple re-compile. Very interesting overall.

I also found something on Intel's site called the approximate math library (http://developer.intel.com/design/pentium4/devtools/) that I might look into.

In theory, there is no difference between theory and practice.
In practice, there is.

 

[Raystonn]

"I haven't read through the whole thing yet, but it definitely seems that you need to code specifically for SSE2, either in assembly, or through some of Intel's C/C++ abstractions. I'm sure the compiler optimizations take some advantage of SSE2, but the tutorial definitely emphasizes hand coding."

Using certain techniques to help point the compiler at optimizations does give you extra performance, but it's by no means required. The Spec benchmarks were not modified at all, just recompiled.

"While re-compiling certainly helped, the big gains were made by re-coding the algorithm with Intel's SSE optimized vector class."

There are already several implementations of the STL classes out there. People constantly compare and swap out different versions to find out which is best. This is just another case of a better implementation for Intel processors. It should easily plug in like any other STL implementation should you decide to use it.

-Raystonn

-- The center of your digital world --

 

[Guest]

>Using certain techniques to help point the compiler at
>optimizations does give you extra performance, but it's by
>no means required.

Well, Intel's tutorial definitely paints a different picture. They are very much emphasizing assembly & their abstraction mechanisms. The tomshardware mpeg article I referenced previously shows significant improvements for the recoded algorithm.

>There are already several implementations of the STL
>classes out there. People constantly compare and swap out
>different versions to find out which is best. This is just
>another case of a better implementation for Intel
>processors. It should easily plug in like any other STL
>implementation should you decide to use it.

Nope, this isn't an STL implementation, it's more like a native vector class. The Tomshardware article on Flask MPEG specifically mentioned re-coding with this vector class, not just swapping STL implementations.

Actually, I'm not sure how you could do anything useful with SSE in the STL unless you made special cases for when STL objects are instantiated with native int or double types. If they have an SSE optimized STL, they don't mention it in the tutorial.


In theory, there is no difference between theory and practice.
In practice, there is.

 

[ksoth]

That's good news for Pentium 4. Northwood will do even better. The main thing, though, is that P4 is way overpriced for the performance benefit it provides. 1.33 gHz Athlon = $218, 1.5 P4 = $470 (without RAM). That is a 116% price premium for 4% better performance Integer and 22% FPU (based on adding the two numbers together). I mean, to me, that simply isn't worth it, and considering P4 motherboards and RAM even further the cost of it... Nevertheless, it does show that P4 does have promise, and hopefully the price cuts rumored for this month are true and we'll see $220 1.5 gHz P4s soon. However, if Intel cuts their prices, you can be guaranteed AMD will follow up. I wouldn't be too surprised if we see sub-$100 1.0-1.2 gHz chips from AMD, especially on Palominos debut. I could be wrong though, as that would be the best of curcumstances.

"We put the <i>fun</i> back into fundamentalist dogma!"

 

[Raystonn]

"The main thing, though, is that P4 is way overpriced for the performance benefit it provides."

This is true, but only until the end of this month. Prices will drop dramatically to less than half of where they currently are. AMD cannot make price drops that drastic as their profit margins are not as large as Intel's. (Intel has a profit margin 50% higher than AMD.) Additionally, the new models are coming out and AMD currently has no released answer to the P4.

When is the Palomino scheduled to debut? Is it designed to be the answer to the P4?

-Raystonn

-- The center of your digital world --

 

[Kelledin]

Actually, the current 1.33GHz T-bird is AMD's answer to the current P4. AMD hasn't been releasing new speed grades because what they have competes with the P4 quite nicely.

Kelledin
<font color=red>"Step away from the gimp suit and put your hands on top of your head."</font color=red>

 

[Raystonn]

"Actually, the current 1.33GHz T-bird is AMD's answer to the current P4. AMD hasn't been releasing new speed grades because what they have competes with the P4 quite nicely."

I'm discussing the P4 as a whole. This is only its beginning range of speeds. The P4 will scale up to at least 5GHz. The Athlon is not the answer to the P4 except for the very short term.

-Raystonn

-- The center of your digital world --

 

[Kelledin]

Do we have any idea what the Athlon will scale up to? We know it will scale up to 1.7GHz; it will probably scale even farther once it drops to .15u and finally to .13u. Has anyone said what the limit is for the Athlon architecture?

And beyond that, there's the Hammer family, which will probably scale even farther. Even though it's a 64-bit chip, there will probably still be a model geared towards the desktop/workstation.

Kelledin
<font color=red>"Step away from the gimp suit and put your hands on top of your head."</font color=red>

 

[ksoth]

AMD can make the price cuts. Sure, their profit margins aren't *as* high as Intel's, but they know that they have to keep their prices lower than Intels to even compete, simpy because the market sees Intel when they think CPUs.

The Palomino is scheduled to debut the 2nd half of 2001, at a speed grade of 1.533 gHz. Current Thunderbird Athlons will scale up to 1.5 gHz, possibly higher considering what overclocking heights have been reached with AXIA coded Athlons. The Palomino is simply another "enhanced" version of the Athlon. The true answer to the Pentium 4 will be the Hammer line of processors, with the Clawhammer being the consumer chip and the Sledgehammer being the server chip. However, the Athlon core will also see another die shrink to .13 micron with the Thoroughbred version, which will scale to even higher speeds and perform even better. Both the Thoroughbred (Athlon)/Apaloossa (Duron) and the Hammers are scheduled for release in the first half of 2002, not too long after Pentium 4 Northwood's release late 2001.

"We put the <i>fun</i> back into fundamentalist dogma!"