Interesting SPEC comparison

[Tea]

I nicked this from The Inquirer. Interesting.

SPECint 2000
400 Itanium 1 800MHz 96KL2 4MB L3
458 IBM RS64 IV 750MHz 8MB L2
604 HP PA 8700 750MHz 0L2 0L3
610 Sun UltraSPARC III Cu 1.05GHz 8MB L2
648 Intel Xeon MP 1.6GHz 1MB L3
679 Compaq Alpha 21264C 1GHz 8MB L2
749 AMD Athlon XP 2100 1.667GHz 256K L2
760 Itanium 2 1GHz 3MB L3
839 IBM Power 4 1.3GHz 128MB L3

SPECfp2000
410 IBM RS64 IV 750MHz 8MB L2
581 HP PA 8700 750MHz 0L2 0L3
642 AMD Athlon XP 2100 1.667GHz 256K L2
671 Intel Xeon MP 1.6GHz 1MB L3
701 Itanium 1 800MHz 96KL2 4MB L3
827 Sun UltraSPARC III Cu 1.05GHz 8MB L2
960 Alpha 21264C 1GHz 8MB L2
1266 IBM Power 4 1.3GHz 128MB L3
1,350 Itanium 2 1GHz 3MB L3

 

[timwhit]

Is the point that the Itanium 2 is a good CPU? Or is it that the Itanium 2 isn't worth the extra cost over consumer CPUs?

 

[Tea]

Well, it has superb floting point performance, but what is the point of a fast FPU in a server CPU? For scientifc and engineering workstations where cost is no object, sure. But to the extent that we can trust SPEC as a guideline (which is not all that much, I daresay) it needs 8MB L2 to more-or-less level peg with a 1/4MB Athlon.

Not exactly a thrilling performer. Still, it's no longer absolute bottom of the pile, which the Itanium 1 was. It wil be very interesting to see what the 64-bit Athlon does. We can write off the Alpha now, PA-RISC is in its final generation, Sun sort of go their own sweet way and don't really worry too much about the numbers (I gather), so I guess it's down to a three-way shootout: Intel vs AMD vs IBM.

Notice that all three are mainstream consumer-level CPU manufacturers. The days of the specialist big-iron maker have been numbered for a good while now, but it's starting to look like the end game.

 

[Prof.Wizard]

I nicked this from The Inquirer. Interesting.

SPECint 2000
400 Itanium 1 800MHz 96KL2 4MB L3
458 IBM RS64 IV 750MHz 8MB L2
604 HP PA 8700 750MHz 0L2 0L3
610 Sun UltraSPARC III Cu 1.05GHz 8MB L2
648 Intel Xeon MP 1.6GHz 1MB L3
679 Compaq Alpha 21264C 1GHz 8MB L2
749 AMD Athlon XP 2100 1.667GHz 256K L2
760 Itanium 2 1GHz 3MB L3
839 IBM Power 4 1.3GHz 128MB L3

SPECfp 2000
410 IBM RS64 IV 750MHz 8MB L2
581 HP PA 8700 750MHz 0L2 0L3
642 AMD Athlon XP 2100 1.667GHz 256K L2
671 Intel Xeon MP 1.6GHz 1MB L3
701 Itanium 1 800MHz 96KL2 4MB L3
827 Sun UltraSPARC III Cu 1.05GHz 8MB L2
960 Alpha 21264C 1GHz 8MB L2
1266 IBM Power 4 1.3GHz 128MB L3
1,350 Itanium 2 1GHz 3MB L3

Is it only me or Athlon XP fairs great in SPECint 2000?

 

[cas]

Is it only me or Athlon XP fairs great in SPECint 2000?

It does. If you really want to get a sense of how poorly today’s multi-thousand dollar 64bit cpus perform on integer apps (which includes most business computing), consider that a Pentium 4 @ 2.53GHz scores 922 SPECint2000.

More than any other processor, ever.

 

[flagreen]

Well for some reason they did not include the Xeon cpu's. Only the Xeon MPs. Here are the results from Ace's with the Xeon's and P4 included.

 

[Buck]

This data is good to have. How do these CPUs compare when arranged in a dual processor setup? Additionally, SPECint2000 is being related in peaks on the graph that flagreen posted. Are the numbers from the INQ peaks as well? Is it worth measuring a sustained or average rate?

 

[flagreen]

Here's a link to Ace's SPEC2000 page where you can compare all the SPEC tests by all sorts of criteria. Basically the order of results doesn't change much.
http://www.aceshardware.com/SPECmine/

 

[flagreen]

P.S. You can also compare results with 2 processors. Again the xeons lead the pack.

 

[Buck]

Thanks flagreen.

 

[alpha754293]

Is it only me or Athlon XP fairs great in SPECint 2000?

It does. If you really want to get a sense of how poorly today’s multi-thousand dollar 64bit cpus perform on integer apps (which includes most business computing), consider that a Pentium 4 @ 2.53GHz scores 922 SPECint2000.

More than any other processor, ever.

Those multi-thousand computers that you talk of pays for support, to make sure that it works nicely together, and that if it does, someone to bitch at.

You try hooking up 106 AMD Athlon MP processors and getting it all to play nicely, as well as having a high speed, high bandwidth system interconnect, management system, stability and compatability that you get with the big name servers. Besides, if you screw up on hooking it up, the only person you can really bitch at is yourself...(or whoever it was that put it together for you).

AND THEN....ESPECIALLY if you're a financial institution, now try getting your month end reports all done in about 8 hours (over night) and have them ready to be sent out the next morning. Which means that you gotta calculate all the accounts, and then print off the reports so that they can be sent and processed.

Granted, I realize that not everybody has their own financial institution or a need for 106 processors, but that's what those multi-thousand dollar computers pay for.

Which would also explain why the processors are the way they are. Or if you connect 4 SGI Origin 3000 servers together to tackle an 8-hour rendering job, granted the Athlons DO render significantly faster, but it's the whole interconnect bit, along with stability and hardware and software support.

Obviously, if you're a home user, there isn't much need for a IBM RS/6000 server in your room. And if there is, I'd be VERY curious what in the world you'd be doing (and how much your power bill is monthly.) Not to ment the HP PA-8700/8800 series, and the Sun, etc etc etc.

Considering that pretty much the ONLY people that are just stepping into the archaic 64-bit arena now are AMD and Intel, it won't be surprising for the servers to be going to 128-bits within the next year. After all, the DEC Alpha along with it's Tru64 UNIX has been around since AT LEAST 1994.

Practically comparing apples and oranges here. AMD/Intel/IA-32/64/x86 - common desktop usage (for the most part). HP/Sun/IBM/SGI VERY VERY VERY specific applications.

 

[cas]

My comment was about processors, not complete machines. Regardless of the processors used, enterprise level servers will always ship with RAS and other features demanded by enterprise customers.

Like the mainframe vendors before them, high end server and workstation vendors work very hard to raise the veil of inscrutability you have repeated for them. I hear it, but I don’t buy it.

Companies like Sun and IBM continue to develop their own CPUs, not because they are better. They continue to develop them because they are different. This is how they maintain their hefty margins.

Despite what they would tell you, and what you seem to believe, you can run a Fortune 50 company just fine, without a single one of their machines.

You might want to reread this thread, in which you participated superficially a few months ago.

 

[Pradeep]

Itanic 2 sets sail without suppliers

"No systems spotted on the horizon"

http://www.theinquirer.net/?article=4331

 

[The Grammar Police]

Hmm. One would have thought that The Inquirer would have known the difference between "discrete" and "discreet".

 

[alpha754293]

My comment was about processors, not complete machines. Regardless of the processors used, enterprise level servers will always ship with RAS and other features demanded by enterprise customers.

Like the mainframe vendors before them, high end server and workstation vendors work very hard to raise the veil of inscrutability you have repeated for them. I hear it, but I don’t buy it.

Companies like Sun and IBM continue to develop their own CPUs, not because they are better. They continue to develop them because they are different. This is how they maintain their hefty margins.

Despite what they would tell you, and what you seem to believe, you can run a Fortune 50 company just fine, without a single one of their machines.

You might want to reread this thread, in which you participated superficially a few months ago.

strictly looking at processors, you're right. However a processor without a board to sit on or a system to run, doesn't nothing more than expensive decor. (In fact, I have a Pentium III 1 GHz sitting on my desk right now. Looks kinda shabby in the anti-static wrap.) But in either case though, I personally think that a processor without a system is not anymore useful than a system without a processor, and that's why I drew the conclusions for the prices as well as performance in that respect.

I guess that overall, my conclusion regarding the Itanium 2 processors would be: If you're using it for computer rendering or in a financial institution, it'll world great because of it's floating point capabilities. Besides, it's a nice fresh page that looks like Intel is trying to start on now, since the current P4s certainly don't give quite a "show" with regards to floating point performance. (Yes, I am aware that a dual 2.2GHz Xeon, with HyperThreading enabled gets just a tad shy under 9 GFLOPs, but as we all know very well, Xeon's are hardly aimed at the mass market.) [entering into vicious circle territory...lol]

 

[cas]

I see, Xeons don’t count because they are not aimed at the “mass market”, but Itanium 2s do? It’s probably best to leave it right there for now.

 

[alpha754293]

I see, Xeons don’t count because they are not aimed at the “mass market”, but Itanium 2s do? It’s probably best to leave it right there for now.

they do.....why do you think I put the whole vicious cycle remark. LOL.....(hence the abrupt stop when I realized that I'm about to rant...incessantly...pointless....and a one-person debate, minus the split-personality....lol)

(at least I kept it to being about processors....lol...it's an improvement, right?? sorta.....somewhat???

 

[Fushigi]

Those multi-thousand computers that you talk of pays for support, to make sure that it works nicely together, and that if it does, someone to bitch at.

The work nicely together part is a major factor here. Sun has been known to release Solaris for small systems 1st while waiting & doing more testing before certifying it for their largest systems. Since the architectures are rather complex, they take the time to do additional testing before releasing the code. IBM does the same this with the zSeries, pSeries, and iSeries. New OS releases go through a ton of testing before being released as does new hardware. And if it doesn't work, you can count on their engineers being at work on the issue until it is fixed. After all, Microsft has a reputation for mediocrity, which it lives up to quite well. IBM, Sun, etc. have reputations of providing mission critical capable systems. In both instances, the reputations have been earned.

You try hooking up 106 AMD Athlon MP processors and getting it all to play nicely, as well as having a high speed, high bandwidth system interconnect, management system, stability and compatability that you get with the big name servers. Besides, if you screw up on hooking it up, the only person you can really bitch at is yourself...(or whoever it was that put it together for you).

Agreed. The part about bandwith is especially true. PC-class systems do not offer the internal bandwidth that is currently available in high-end servers from Sun & IBM.

AND THEN....ESPECIALLY if you're a financial institution, now try getting your month end reports all done in about 8 hours (over night) and have them ready to be sent out the next morning. Which means that you gotta calculate all the accounts, and then print off the reports so that they can be sent and processed.

Although this can be handled differently. Our ERP does month-end cutoff in about a second. Reports are then run & sent to clients over the next 2 days.

One thing about financials .. they are very often not dependent on floating point. Sure, money has decimal values in it, but a common technique is to treat $123.45 as 12345 internally and just use a display mask to display it with the decimal point. Calculations are thus integer-based are faster than they would be if relying on FP performance.

Granted, I realize that not everybody has their own financial institution or a need for 106 processors, but that's what those multi-thousand dollar computers pay for.

Support, reliability, and availability of application software is what sells the big iron.

Considering that pretty much the ONLY people that are just stepping into the archaic 64-bit arena now are AMD and Intel, it won't be surprising for the servers to be going to 128-bits within the next year. After all, the DEC Alpha along with it's Tru64 UNIX has been around since AT LEAST 1994.

And the AS/400 (iSeries) has been 64 bit since 1995. But I honestly think we will stay 64 bit for server & general purpose computing for a few years. I think the trend will be towards more CPUs in SMP / NUMA type configs and clustering. Specialty processors like graphics chips will progress to 256 bit, maybe more as the desire for better rendering & displays will continue until games have near photo-realistic imagery.

- Fushigi

 

[alpha754293]

I think that I've seen banks that do their month end calculations in floating point. LOTS of floating point at that.....15...16...decimal places. Course..it COULD be possible that they're using the display mask that you're talking about, but still.....it's still a big job that slapping machines together just doesn't necessarily equal the power of a mainframe system.

What WOULD be neat is that if there was a company that actually takes off the shelf components, and put it together into a mainframe or closer to mainframe-like system. Like taking 40 1U dual AMD systems, and have the management console somewhere in there as a 1U rackmount. That....would be neat...especially if it's backed by AMD itself.

 

[Clocker]

From my perspective, I'm REALLY enthused about Itanium2. For my work, it will do a lot to decrease the compute time of my Dyna models. 10-11 hours to compute about 200 milliseconds of data for a 45,000 element door slam model is still too long for me....

My J6000 needs more speed. I wonder how fast my TigerMPX set-up is in comparison? Any ideas where I'd find the SPECfp numbers for a single or dual PA-8600 set-up like mine here ? : http://www.hp.com/workstations/products/unix/jclass/j6000/summary.html

C

 

[cas]

I tend to tear down misconceptions about supposed ‘big iron’ or ‘high end’ workstations with regularity, and I seem to remember doing the same with your machine. Since I can’t find the link just now, I will assume it was before the loss of SR’s archives.

HP’s PA-8600 @ 552 MHz should turn in a peak SPECfp2000 score of roughly 485. A 1.8GHz AthlonXP comes in around 671. Thanks to SSE2, the Pentium 4 @2.53GHz manages 901. Given that the HP system uses dual channel SDR SDRAM, I would be very surprised if a commodity single cpu Pentium 4 system did not outperform the j6000 workstation for all real loads.

That said, newer HP systems based on the Itanium2 will indeed be much faster. I have long found it curious that the Itanium line is often associated with servers. Presumably this is because servers have a clear need for 64 bits, and can better accommodate Itanium’s outrageous prices. This is curious because the Itanium2 offers record breaking fp performance, with the integer performance of a sub $100 cpu.

Despite the Itanium2’s strength, I can’t get around its cost structure. I would be very tempted to pitch HP-UX and their $23,000 workstations for faster $7,000 NT/XP workstations. Add this to the money they would save bypassing terminal services, and they might be able to afford a full 29 days off per year for employees who started after the late eighties.

 

[cas]

Even if your applications are not presently available for XP, I suspect splitting the $16,000 delta would offer a pretty decent incentive to port the code.

 

[Mercutio]

... or transition to 64bit Linux, which seems to be more in-line with HPaq's strategic posturing of late and would doubtless pose less difficulty in portability.

 

[Clocker]

I'm not sure how much we pay to use our J6000's but we lease rather than purchase all of our workstations. When you have hundreds and hundreds of them (if not thousands world-wide, I'm not sure), you gotta lease'em. Luckily, our upgrade cycle is 3 years...

C

 

[Buck]

So, how much time do you have left on the leased system Clocker?

 

[Clocker]

Scratch that...make that number about 4,000 world-wide (an educated guess this time)....

C

 

[timwhit]

Well I'm sure it costs more to lease a $23000 computer than a $6000 computer. Who is in charge of deciding what computers you use anyways?

 

[cas]

In fairness to Clocker’s employers, my first job in a frigid, raised floor room, was working with a Wang VS-100. Our maintenance contract, which included full-time, on site personnel, was exorbitant. If it wasn’t bad enough with our users enjoying Wang ‘office’ applications, it got much worse once they decided that Word for Windows was cooler. We implemented the reverse of Clocker uses, $2000 VS communications cards inside of brand new PCs.

If it wasn’t expensive enough to maintain the VS, now we were maintaining two parallel systems. Eventually, the critical software was ported, but the intervening period was very expensive.

Clocker had previously mentioned that they use terminal services for office apps. This is the reason I recommended NT/XP over Linux. I have managed mini-computers, Unix systems, and NT systems. Whichever you prefer, homogeny is always easier to deal with.

 

[Clocker]

Yeah, I agree the homogeneous Win2K route would probably be cheaper and better if everybody used it nad it was possible to roll it out everywhere. However, we are split up between the following types:

Somewhere around 100,000 people or so using (on average) P2-450 machines with Win98. THese people are being migrated to Win2K. THese people are supported by Company A. These people have always only used office apps.

About 4000-5000 people are using UNIX based Workstations for technical stuff. Before we had a terminal service type set-up, besides having the UNIX workstation for our technical work, we also shared a WinNT PC which we shared with one other person. So that is 1.5 computers per person. These people are supported by Company B (for UNIX) and Company C (For WinNT Workstations).

At least now, with the terminal service for Office stuff, we don't need to rely on Company C or the use of a separate PC for reporting and creating presentations etc.

Back a few years ago, I doubt the computing situation was the same as it is now (I could be wrong). I don't think your run of the mill P2 450 was in the same league as the RISC boxes we were using at the time (i.e HP J5000s). That being said, it would be nice if everyone could just switch over to Win2k (P4/Itanium/Athlon) boxes but to do everyone like that to create a homogeneous environment would probably be cost prohibitive. Hopefully, over the long haul, we will migrate to a more homogeneous and cost effective system. But, we're talking about corporate America here....so I doubt that will happen.

 

[Clocker]

Well I'm sure it costs more to lease a $23000 computer than a $6000 computer. Who is in charge of deciding what computers you use anyways?

As I stated earlier, we lease our machines. We could not afford to buy them all... Also, based on our volume, I'd expect we get a volume discount of some sort.

C

 

[James]

As I stated earlier, we lease our machines. We could not afford to buy them all... Also, based on our volume, I'd expect we get a volume discount of some sort.

The lease is much more likely to be because that way your workstations and PCs are opex rather than capex, a result of the blinkered view that capex = bad, opex = good that continues to blind analysts and company managers because they don't understand basic finance and accounting. (These are generally the same people that cling to the view that pro forma accounting and EBITDA actually offer any remotely useful information about a company's performance.)

It is also possible that it is more tax effective to lease than to buy the machines outright and then depreciate them over three years. It depends on how your company accounts for losses and depreciation.

 

[time]

Computer leases are a mug's game, yet there seems to be no shortage of punters eager to boost financial institution profits. Leases only really make sense where the goods retain some sort of value over a lengthy period. Clearly, that is not the case with computer hardware (or even more mindboggling, software).

Governments have taken to leasing everything that isn't nailed down, simply to hide debt. Leasing disappears into operating expenses, whereas borrowing the money to buy something seems to makes credit rating agencies waggle their fingers disapprovingly. Leasing is popular with state governments here, for example, because their total borrowings are restricted by the federal government. 'Sale and Leaseback' is a concept that defies any kind of rational scrutiny. :roll:

Struggling corporates also embrace leasing for the standard reasons of deferring reality: "Just worry about this year. Next year can take care of itself, and besides, we or the company may no longer be here."

Don't misunderstand, leasing of plant and equipment makes good sense when a company is expanding and needs to preserve cashflow. And rates will be good if the lease tracks the asset's realistic value (because the financier is always covered).

AFAIK the tax 'advantage' falls into the same category. If you can depreciate an asset over three years, its entire value plus interest is tax deductible over that period. So are lease payments. Anyone who believes they can gain an easy advantage gravely underestimates the taxation beaucracy.

There's a worrying trend in Oz for retailers to push lease (hire) agreements onto individuals and small businesses. For example, Dell is currently advertising a P4 1.8 with LCD monitor for AU$1999, or only $14.40 a week.

Unfortunately, the contract duration is four years, not three, so the hapless buyer ends up paying $2995, plus a whopping $99 delivery! And that is one of the more reasonable schemes I have seen.

The real killer is when you decide you want to upgrade after two years, as many such schemes 'offer'. If they're unbelievably generous, you may only owe them half the original purchase price plus a penalty. Of course, the computer will be worth a fraction of this by then, so you can't fund an escape by selling it.

In effect, you continue to make your lease payments as if nothing had happened, except you also have to find more than half the original value of the goods in cash.

But that doesn't happen in real life. Real lease plans are way more punitive than that. Lease payouts are often determined by the "Rule of 78", which effectively guarantees you will always pay the 'interest' for the entire lease, regardless. So if applied to the Dell example, after two years you would owe nearly $1500, even though the original price was just $1999, and the realizable value may be only $500.

 

[Tea]

Amen