Anand: "Intel has performance crown in all measurable categories"

[GL]

Replace the terms P4 with Pentium and Athlon with 486 and this could be a discussion straight out of 1993. The 486 sure was impressive towards the end of its life span because software wasn't available to take advantage of the Pentium platform. But the industry moves on. What has impressed me is the rapidity with which Intel has gained the performance crown yet again with the P4. They did it in at least half the time that it took them to make the Pentium more attractive than the 486. At this point, you can't go wrong with either an Athlon or a P4 unless, as somebody has pointed out, you run Linux where things just haven't been optimized for the P4.

Yes I'm a P4 owner but only out of monetary reasons. I was deadset on getting a new AthlonXP when I became aware that going with a P4 1.6A + DDR platform was cheaper than going with an AthlonXP 1800+ + DDR. As it stands, I'm at 2208MHz with my P4 and it's just faster than an Athlon XP 1800+ in the things I do (lots of video encoding, mp3 encoding, simulation games) but was $100 cheaper.

It is a well documented fact that the biggest bottleneck with business applications is generally found in the I/O subsystem. If you want big gains in that area, make sure your hard disk and CD-ROMs aren't the bottleneck. Moreover, increase the amount of RAM and try to get the fastest RAM your system supports. Additionally, if you're using XP, avoid ATI cards if you want a general "snappy" feel as they don't have as good 2D acceleration as Nvidia cards do (being an ATI owner I can verify this unfortunate fact).

 

[Jhhnn]

At stock clock speeds, AMD definitely has the price/performance advantage. A quick jump to pricewatch shows the AMD XP 2.1 at $236, the Intel 2.4 Northwood at $588. So if you're willing to pay 2-1/2 times as much for the processor, you might be able to detect some real-world performance difference. Personally, I have a lot of other things to do with the extra $350, and I suspect that most everybody else does too, but to each their own.....

 

[ssanches]

<< Apparently you haven't understood the crux of the statement.

I understood perfectly. My message was seeking superior alternatives to benchmarks, ideally, forms of evaluation that needn't be discounted.
>>



To be honest, you won't find any alternative anytime soon. Even the latest breed of benchmarks have their own shortcomings. That's why we need to understand the limitations of today's benchmarks themselves. Add the fact that desktop systems are so modular and can scale from a simple mail server to a large Beowulf cluster. A couple years down the line, I don't think Mhz (Ghz) shall be an accurate rating of a processor's performance (in fact even today it isn't). One would be tempted to hope some sort of ISO "standard" is made; probably on similar lines as AMD's PR rating, but one that is more logical and which is based on benchmarks truly neutral to all platforms.

IMHO AMD's PR rating is a good start but it has the shortcoming of not being truly neutral and not based on an independent organisation's performance rating. Also AMD's rating starts with a PR of 1.125 PR points for every Mhz (the AXP1500+ CPU). It then assumes a delta of 100PR points for every 66.67 Mhz increase (i.e. 1.5 PR points for every Mhz increase). This is definitely illogical; their CPUs don't scale that way.

 

[dullard]

<< IMHO AMD's PR rating is a good start but it has the shortcoming of not being truly neutral and not based on an independent organisation's performance rating. Also AMD's rating starts with a PR of 1.125 PR points for every Mhz (the AXP1500+ CPU). It then assumes a delta of 100PR points for every 66.67 Mhz increase (i.e. 1.5 PR points for every Mhz increase). This is definitely illogical; their CPUs don't scale that way. >>



I always see this formula stated in some really confusing way like you did. I think it is much more clear in this format:
PR rating = 1.5 * XPspeed - 500
(try it, it works perfectly with every XP produced so far.)

But there is a really interesting error in this formula. An underclocked Athlon XP at 1 GHz has a PR rating of 1000. Thus an Athlon XP at 1 GHz performs identical to a T-bird at 1 GHz. We all know there are architectural improvements in the XP (one of the reasons that a PR rating was used) - so the XP at 1 GHz should be faster than the T-bird. The formula is dead wrong in this simple case. In fact the formula predicts an Athlon XP underclocked below 1 GHz is slower than an identical clocked T-bird.

The formula has a problem in the long run too. As the XP speed increases toward infinity, you eventually get:
PR rating ~= 1.5 * XP speed
Sure there are architectural improvements, but the XP is not a full 50% faster than the T-bird at identical speeds.

Now lets look at the numbers in the formula: 1.5 and 500. Don't these look quite pretty? If they were actual benchmark numbers they would look more like this: 1.487 and 562.4. But no they are perfectly pretty numbers of 1.5 and 500. I think these weren't ever made from a real benchmark - or we have a major coincidence.

So we have a PR rating that fails at low speeds, looks fishy at moderate speeds, and fails at high speeds. That is the problem. How can anyone trust those numbers?

 

[CraigRT]

Performance aside, they both perform well.. but Intel charges too much for their CPU's.. I remember with the 2.0 P4 (willamette) came out around here in Canada, it was like $900... WTF?
I could get a 1.4 GHz T-Bird for $200 with equal performance... That is the only thing that irks me about Intel... oh, and the obvious whole SCAMbus idea... being forced to buy RDRAM in the P4 initial stages was just plain gay. they both do what they are designed to do, and both do it well.. but it boils down to price.. if they were the same price, i'd have one of each in my house... now that they are beginning to become close in price, i am considering a Northwood.. but before then... Intel was a COMPLETE RIP OFF... That's the only reason I am die-hard AMD, but maybe not for all that long if prices continue to drop..

 

[Red Dawn]

<< But there is a really interesting error in this formula. An underclocked Athlon XP at 1 GHz has a PR rating of 1000. Thus an Athlon XP at 1 GHz performs identical to a T-bird at 1 GHz. >>

Just for the sake of adding a little fire to this debate (as opposed to a flame) How would a P4 underclcoked to 1000 mhz fare against a P3 1000eb Coppermine? If it isn't as fast as the Coppermine then even though the clock speed for the P4 is true how can you trust it in regards to performance?

Maybe it would be more accurate if AMD were to use a PR to compare their Processors with the P4's.

 

[dullard]

<< How would a P4 underclcoked to 1000 mhz fare against a P3 1000eb Coppermine? If it isn't as fast as the Coppermine then even though the clock speed for the P4 is true how can you trust it in regards to performance? >>


The P3 would blow the P4 away. Everyone who knows anything about computers would agree. Intel doesn't have a performance rating that says a P4 will beat an equally clocked P3. Intel doesn't have any performance rating at all. Intel SHOULD have a performance rating, but they don't (Here I mean an independant rating that is the same basis as AMD, Apple, Sun, etc). Intel just hopes that less knowledgeable people blindly buy based on MHz - and this is misleading.



<< Maybe it would be more accurate if AMD were to use a PR to compare their Processors with the P4's. >>


It would be the better in my opinion. When a new computer user goes to buy a computer they ASSUME the PR rating matchs a P4 speed. My grandmother doesn't go into a computer store to buy her first computer wondering how this XP machine compares to a T-bird! AMD's PR rating is very misleading for new buyers since it doesn't compare to a P4 and looks awfully similar to the P4 speeds.

So both companies mislead. Intel misleads by not giving a performance rating at all hoping that new computer buyers compare the P4 MHz to the P3 MHz. AMD misleads by giving a performance rating compared to a processor that no one cares about hoping that new computer buyers compare the XP rating to the P4 MHz.

The least AMD could have done with a PR rating is to use real benchmarks (not the fishy numbers from my post above) or made it match existing processors that are actually selling. If they took the trouble to make a performance rating, at least make it correctly match the T-bird or make it easy to use by matching other currently selling processors. Making a fishy rating that is based on a processor that isn't selling is not the way to go.

 

[Rob9874]

<< Performance aside, they both perform well.. but Intel charges too much for their CPU's.. I remember with the 2.0 P4 (willamette) came out around here in Canada, it was like $900... WTF?
I could get a 1.4 GHz T-Bird for $200 with equal performance... That is the only thing that irks me about Intel. >>



But Intel's making a profit (kind of important to those of us who rely on Intel to put bread on the table). AMD can't make a profitbale year. How long can a company stay in business losing $61M like AMD did in 2001? Sooner or later, they're going to have to charge more for their product. You usually see a dramatic increase in price when a revolutionary product comes out, in which the consumer will pay whatever it takes to get it. Like the Pentium in 1994, or the Hammer. AMD is going to raise prices, believe it or not. They're not in the business of losing money just to keep their customers happy. People are "irked" at Intel for charging more. Well, why do you think AMD doesn't? Do you really think they're being altruistic?

 

[dullard]

<< How long can a company stay in business losing $61M like AMD did in 2001? >>


Quite long, remember AMD had larger losses in 1999, 1998, 1997...



<< Sooner or later, they're going to have to charge more for their product. >>


They have been quietly. Here are the processor release prices from AMD in 1000 lots in US dollars. These are not street prices.
Feb 2000: 850 MHz, $849
Mar 2000: 1000 MHz, $1299
Aug 2000: 1100 MHz, Sorry I lost the price
Oct 2000: 1200 MHz, $612
Mar 2001: 1333 MHz, $350
June 2001: 1400 MHz, $253
Oct 2001: 1533 MHz, $252
Nov 2001: 1600 MHz, $269
Jan 2002: 1666 MHz, $339
Mar 2002: 1733 MHz, $420

Looks like AMD prices are going up since Oct 2001.

Intel top processor prices are always a consumer rip-off. Never buy the top speed P4. But with two price cuts in 6 weeks totalling up to 57% off, then the 2.4 GHz will be affordable.

 

[ssanches]

My humble comments annexed below...



<<
I always see this formula stated in some really confusing way like you did. I think it is much more clear in this format:
PR rating = 1.5 * XPspeed - 500
(try it, it works perfectly with every XP produced so far.) >>


The thing I wanted to highlight was the performance delta AMD is trying to portray. An AXP1500 will beat a P4-1.5 by a larger margin than an AXP2000 would beat the P4-2.0



<< In fact the formula predicts an Athlon XP underclocked below 1 GHz is slower than an identical clocked T-bird. >>


True. But AMD has never targeted selling a 1.0Ghz palomino anytime soon



<< The formula has a problem in the long run too. As the XP speed increases toward infinity, you eventually get:
PR rating ~= 1.5 * XP speed >>


Yeah, I too feel the AXP will (say) perform at par with a P4 at PR-2600 v/s 2.6Ghz for the P4 on most benchmarks and below par on a few. After that I doubt AMD can justify the PR points using the present formula



<< Now lets look at the numbers in the formula: 1.5 and 500. Don't these look quite pretty? If they were actual benchmark numbers they would look more like this: 1.487 and 562.4. But no they are perfectly pretty numbers of 1.5 and 500. I think these weren't ever made from a real benchmark - or we have a major coincidence. >>


AMD had to make them look pretty. Marketing an AthlonXP as PR:2162.5+ would seem downright stupid



<< So we have a PR rating that fails at low speeds, looks fishy at moderate speeds, and fails at high speeds. That is the problem. How can anyone trust those numbers? >>


Dullard, can we also trust Mhz/Ghz? IMHO AMD has really made a good start with this performance rating but it does have shortcomings. It's much more objective than the PR used for the K5 and Cyrix chips. A couple yrs down the line, it'll be downright impossible to use Mhz as the metric for measuring CPU performance! This problem will only get accented once we see the ClawHammer and then the n00bs shall be wondering how a CH @ 2Ghz outperforms a P4 @ over 3.5Ghz....

 

[ssanches]

<<

<< But there is a really interesting error in this formula. An underclocked Athlon XP at 1 GHz has a PR rating of 1000. Thus an Athlon XP at 1 GHz performs identical to a T-bird at 1 GHz. >>

Just for the sake of adding a little fire to this debate (as opposed to a flame) How would a P4 underclcoked to 1000 mhz fare against a P3 1000eb Coppermine? If it isn't as fast as the Coppermine then even though the clock speed for the P4 is true how can you trust it in regards to performance? >>



That's an excellent point you've put out there RedDawn. Maybe the folks at Aberdeen should try it out I know this has been said many times, but Mhz ain't the metric for measuring CPU performance

 

[gregor7777]

I know this has been said many times, but Mhz ain't the metric for measuring CPU performance


Then what is?

And I don't mean this as a jab at anyone, I mean it seriously. What is the measuring stick that is common to all CPU's that should be used in your opinion?

 

[dullard]

"True. But AMD has never targeted selling a 1.0Ghz palomino anytime soon "
They also aren't selling a 1.733 GHz T-bird. So why base a Palomino on a T-bird? A base on a P4 would make more sense.

"Yeah, I too feel the AXP will (say) perform at par with a P4 at PR-2600 v/s 2.6Ghz for the P4 on most benchmarks and below par on a few. After that I doubt AMD can justify the PR points using the present formula"
Luckily for AMD the Athlon XP won't be reaching the extreme limit since it will be replaced by other processors. But I think the formula needs some serious work.

"AMD had to make them look pretty. Marketing an AthlonXP as PR:2162.5+ would seem downright stupid "
They can sell them to rounded numbers - I'm fine with that. But the linear formula doesn't work at all when you extrapolate in either direction. If AMD made a 1.8GHz XP it might be better to label it as 2175+ rather than 2200+, eventually they need to stray from the formula.

"Dullard, can we also trust Mhz/Ghz?"
We can trust MHz/MHz when comparing the same processor line. The 1.5 GHz P4 outperforms the 1.3 GHz P4 in basically everything (unless there is some odd program that is completely CPU independant). But it doesn't apply accross different types of processors. The 1.6A GHz P4 outperforms the 1.6 GHz P4 even though it is the same exact frequency. Look at some of my posts, I strongly urge 3rd party benchmarks as better than what we have from Intel or AMD.

I collected computer magazines about 10 years ago. But every computer in those magazines gave benchmark performance in order to enhance sales (some motherboards were a few percent faster than others). This was quite useful since they usually used the same benchmark. We need industry standard agreements on benchmarks something like these:
1) A gaming benchmark based on the top 25 currently selling games - each weighted evenly.
2) An office benchmark on the top selling office programs.
3) A graphics benchmark on the top used professional graphics software.
etc.
Every computer maker agrees to the standard and every performance mark is based on the scores.

Intel does nothing, and that is misleading. AMD uses a PR rating that isn't based on any actual benchmark - but rounded to nice numbers for good marketing.

The key is 3rd party benchmarks. I could easily choose 5 popular games that makes AMD shine, I could easily choose 5 popular games that makes Intel seem far better. Thus AMD and Intel should have no input in the games chosen (I would let the customers choose with their wallet - the most popular games are included). A 3rd party will take the motherboards and chips that are provided and bench them using the same equipment (no false benchmarks with one company using IDE and another using SCSI drives).

Benchmarks are more a focus on how the program was compiled. Thus we cannot use synthetic benchmarks since the person doing the compilation may be heavily biased - only use the most popular selling (already compiled) versions.

 

[ssanches]

<< I know this has been said many times, but Mhz ain't the metric for measuring CPU performance


Then what is?
And I don't mean this as a jab at anyone, I mean it seriously. What is the measuring stick that is common to all CPU's that should be used in your opinion? >>


IMHO a set of benchmarks neutral to diff platforms selected by a neutral third party (who understands the limitations while interpretating benchmarks) and a logical application of it....(discussed in earlier posts and dullard too stated this above)

"Yeah, I too feel the AXP will (say) perform at par with a P4 at PR-2600 v/s 2.6Ghz for the P4 on most benchmarks and below par on a few. After that I doubt AMD can justify the PR points using the present formula" - Luckily for AMD the Athlon XP won't be reaching the extreme limit since it will be replaced by other processors. But I think the formula needs some serious work.

Actually I was thinking about the T-bred that will be .13micron and similar to the AXP in all respects. Got any info if AMD shall be revising the formula, cause the T-breds shall definitely scale above 2600+ levels? Or shall AMD be taking the slowest T-bred as the base and then extrapolate?

 

[Howard]

SPEC... interpreted as 3DMark-type scores... Rounded off and divided for ease of use...

 

[BMdoobieW]

As many have already said, I too am happy with my 1.4 T-Bird and will be keeping it for a while (the upgrade from a P2-266 laptop was breathtaking!)