What makes Xeons so expensive and is the performance worth it?

[AluminumStudios]

I've having trouble understanding Intel Xeon CPUs (PIII based and the new P4 based ones.)

I know they have more cache that runs at full speed. In the PIII Xeon's case this was external cache. But the caches don't seem large enought to justify the extremely high prices. I've seen PIII based Xeons with 512 KB, 1 meg, and 2 megs of cache. I don't think I've seen these much above 900 MHz. How would these stack up against a faster CPU, such as the new PIII's running at 1.26 or 1.4 GHz that have 512 kb of cache. Or even and Athlon?

What differentiates a P4 based Xeon from a "normal" P4? What performance characteristics to they offer? I read a comparison on some tech site between a dual Athlon and a 1.7 GHz dual Xeon setup and the Athlon was a pretty clear winner ;-)

Any light that someone could shed on the topic would be appreciated.

 

[AndyHui]

Pentium III Katmai class XEON processors, had discrete external cache in sizes of up to 2MB. Discrete cache chips of the speeds and sizes required by these processors are prohibitively expensive.

Pentium III Cascades class XEON processors feature large on-die cache running at full speed. The yield on such parts is much lower than the standard Coppermine. A premium part, with a premium cost.

Bear in mind that Coppermine XEONs exist: these are simply standard Coppermine Pentium III processors with 256KB of L2 cache in a Slot 2 form factor. These processors are barely more expensive than their standard Slot 1 and Socket 370 FC-PGA counterparts.

Also note that Slot 2 processors do have one or two other things extra, such as management features, an EEPROM with a unique serial number and a management ASIC.

The standard Willamette and Northwood processors in the Socket 423/478 format are not SMP capable. Pentium 4 class XEONs, the Foster and Prestonia respectively, come in Socket 603 format, obviously with SMP enabled, and more power input pins.

As far as the Willamette and Foster are concerned, these are the same processors....there is not much price difference between the two.

The Prestonia is a little different from the Northwood, in that the Prestonia has Jackson SMT Technology enabled, allowing the operating system to effectively use a singal physical processor as two logical processors.

I would like to point out that generally these days, the XEON processors are not much more expensive that the standard desktop line of processors. It is the motherboard that has sky high prices.

Under standard desktop applications, there is little or no performance increase to justify the price difference. Where you will see the difference is applications that need massive IO bandwidth such as databases and servers. Here the XEONs do very well, and seem to be able to fully justify the cost.

 

[AluminumStudios]

Informative, thanks!

I think it's kind of silly though for Intel to put a Coppermin in a slot2 package and call it a Xeon.

 

[dullard]

<< How would these stack up against a faster CPU, such as the new PIII's running at 1.26 or 1.4 GHz that have 512 kb of cache. Or even and Athlon?

What differentiates a P4 based Xeon from a "normal" P4? What performance characteristics to they offer? I read a comparison on some tech site between a dual Athlon and a 1.7 GHz dual Xeon setup and the Athlon was a pretty clear winner ;-) >>



AndyHui answered pretty well on the technical differences. I'm a Xeon user and I'll answer more on the practical differences.

Some computer uses need as much cache as possible. One good example is with servers. Cache is extremely expensive, so you don't want more than you need. However if you get a larger cache, I've seen some programs running drastically faster. A 900 MHz PIII with a large cache may easily be faster than a 1.26 GHz PIII in these few programs.

More importantly, the Xeons can be used in multiprocessor machines. The PIII Xeons can have up to 8 processors running in a single machine. When you use that many processors, you often start to have inefficiencies - so the 8 processor machine runs roughly 6 times as fast as an equivalent one processor machine. The 6 is a rough number as it depends drastically on the specific application. Thus a computer with eight 900 MHz PIII Xeons runs about equivalent to a 6*900 = 4.5 GHz PIII machine! That will easily outpace ANY single processor machine currently available on programs that can use multiple processors.

Currently the P4 version of the Xeon (Intel Xeon) can only use two processors. Intel is supposed to release a 4 or 8 processor version of the Intel Xeon any day now. This summer I will buy an 8 processor 2.5 GHz Xeon machine. In my applications I will expect this to run at a speed equivalent to a 6*2.5= 15 GHz P4! Obviously, the P4 and Intel Xeon do fewer operations per cycle than an Athlon, so this will be roughly equivalent to a 10 GHz Athlon. Imagine the speed of a 10 GHz Athlon...

Now the practical use. Suppose a company uses engineers that are paid $80,000 per year doing only computer research. That company can buy six $2000 single processor machines and hire 6 engineers to do their work - total price in 1 year is 6*2000 + 6*80000 = $492,000. Now suppose that same company buys a $30,000 computer with eight processors. That company then only needs 1 engineer. Total cost in 1 year: 30000 + 80000 = $110,000. That company saves over $300,000 per year by purchasing these "expensive" processors. To be realistic, few engineers work solely on computer research, so the savings will be slightly less.

As a side note: the scientific programs that require such speed often cost $20,000+ per year. Thus no normal place (like Anandtech) ever benchmarks these programs. In my case, a P4 happens to run circles around the Athlons. I use Fluent. This obviously isn't the same with many other programs - the Athlons may run circles around a P4 in them. So take a benchmark of the fastest dual Athlon vs a dual computer with one of the slowest Xeons (1.7 GHz) with a grain of salt.

 

[dullard]

Here is a typical benchmark comparing single processor machines to multiple processor machines. This benchmark is for the aerodynamics of a car (similar to the work I do).

Note: some of these are machines with multiple processors, but most are multiple machines in a cluster. So the efficency will be different for the clusters since there is a delay when transmitting data from machine to machine.

Here is one example from that page for the lazier Anandtech readers:

# processors / Speed in multiples of a single processor
1 / 1.0
2 / 1.9
4 / 3.5
8 / 7.2
16 / 14.5
32 / 28.4
64 / 46.4

In this case, using 8 processors was equivalent to 7.2 individual processors. This is slightly different from the 6 I used in my previous post.

 

[AluminumStudios]

Thanks for your explanation dullard. You gave it the perspective that I was looking for.

I am a system administrator at a large university. I worked in the Neurobiology dept. now I'm in a Bioinformatics department. I certainly understands the rigerous demands of high-end software. After seeing how Suns and other platforms come with 2,4, 8 megs of cache, I was doubtful how much performance 512kb to 2 megs would actually gain you. I guess it does have a noticeable effect since you vouch for them and people continue to buy them.

One thing that urks me though is that support for SMP is really an artificial thing. Celerons weren't supposed to work in multiprocessor configuration, but they did (I think on the Asus BP6 if I'm not mistaken.) I'm have a feeling that the only reason you don't see quad or 8 way PIII or P4 systems is because of artificial limitations that Intel has placed on their chips to control how the different product lines are used.

I still have one question. Aside from the CPU packaging, what differentiats one of the new PIII chips @ 1.26 or 1.4 GHz and 512 kb of cache from the lower end PIII Xeons that had 512 kb?

 

[dullard]

<< I guess it does have a noticeable effect since you vouch for them and people continue to buy them. >>


Again, it is a noticeable effect for the people who need it. For the vast majority, it is a waste.



<< One thing that urks me though is that support for SMP is really an artificial thing. >>


Yes it is artificial. However it is about reducing expenses and maximizing profits. The more a company can make a product specific to their target customers, the less the product costs to make and support and the more the company can charge for the item. Thus Intel and AMD get to charge more for selling Xeons and MPs yet they spend less to make each processor.

Think about it in terms of cars. There are low end cars, regular cars, and luxuary cars. If the car manufacturers only made one model, then they would have to put leather in every car. That leather is an extra expense that the people who would buy low end cars don't even want. Also the manufacturers would miss the opportunity to double the car price and call it a luxury for the people who want to spend lavishly and have something better than the rest. With just one model the manufactuers have extra expenses and less revenue.

Sure it sucks for those people who need the leather and we have to pay more. But the manufacturers maximize profit that way, and consumers get the benefit of a product that matches what we want.

 

[Diable]

<< I still have one question. Aside from the CPU packaging, what differentiats one of the new PIII chips @ 1.26 or 1.4 GHz and 512 kb of cache from the lower end PIII Xeons that had 512 kb? >>



The only P3 Xeons that had 512K of cache are the 500 and 550mhz versions, the 600mhz to 1GHz P3 Xeons had either 256K, 1mb or 2mb of cache. The 500-550mhz Xeon's were 0.25 micron chips and the cache ran at full cpu speed unlike the desktop P3's at the time who's cache ran at half the cpu's speed. The chips also had a core voltage of 2.0v, ran on a 100 fsb and were the first Xeons to support SSE. Like Andy pointed out earlier the P3 Xeons are 550mhz are just desktop P3's in a Slot 2 case.

The new 1.13-1.4GHz P3-S(there new name)differ from the P3 Xeons in a couple of ways. Their only available in a socket 370 form factor and the only chipset from Intel is the i815E. You can get dual proc motherboards that support the P3-S' but they use either a Severworks or VIA chipset no Intel. These chips are also 0.13 micron, have a core voltage of 1.45v, they can be run without a fan, have hardware pre fetch and they overclock like crazy.