Time for an Intel Itanium recall

Hi Mike,

Well, Intel engineers are at this very moment sending out to HP, NEC, etc the program to test the Itanium 2 processors for the newly discovered bug - details of which are right now unknown, except for the fact that it´s a hardware bug and it cannot be fixed through any microcode update or patch. The present fix is to slow down the CPU to 800MHz. And an Itanium 2 running 32-bit software at 800MHz is slooooooooow. Even running 64-bit software, it´s slower than the slowest Opteron chip (the 240 at 1.4GHz) in most, if not all applications. I doubt many users are going to want to stick with their defective CPU.

Conclusion: this bug/fault/errata will lead to a recall.

Remember the Pentium FDIV bug in 1995? Intel provisioned $475 million for that one. But this time around there are at most 5,000 processors to be recalled, so even at $3,000 a pop, it´s only going to cost Intel around $15 million.

The problem, as pointed out already, is that many clients are going to delay any purchase of Itanium 2 systems until Madison comes out in July. So, if sales of Itanium 2 systems were slow up to now, Intel can count on approximately zero sales for the next two months.

Not good. The ¨marketing opportunity cost¨ for these two months at this strategic moment is certainly going to cost Intel much more than $15 million.

Time to bring that Yamhill ¨plan B¨ thingy out of the Oregonian closet, I would guess. But that won´t happen either. IIRC, Yamhill was buried some time ago.

So it´s Madison or bust for Intel 64-bit.

Ahem, may I add that this time around AMD seems to have a good product, a good strategy, and not-too-bad timing?

Regards

Name, email address supplied, as usual

re: http://www.theinquirer.net/?article=9467 "3700 in 2002"

FWIW - I don't know about 2001, but I have seen the number for 2002 referred to as "over 3000" (implying less than 4000). So you're within 700 at the worst.

The Itanium wasn't really shipping until 2003 though. Even now, it's in fits and starts, with this latest erratum being a potential millstone for a segment that counts on near-100% reliability.

Corporate users name, email, supplied

Mr. Magee,

It is true that there are applications and benchmarks that perform poorer with HT enabled because both threads must share the same very small L1 cache. In apps that cause frequent cache misses or don't lend themselves well to proper branch prediction that L1 sharing becomes a bottleneck creating extra overhead that does not exist when HT is disabled. AMD obviously, purposefully chose a benchmark suite that caused that type of bottlenecking to show off the Athlon's strength under those conditions.

And, yes, I'm quite aware that I don't understand what I'm talking about but am merely repeating what others have said. Hopefully those others DO have a proper education and experience.

Take care,

Scott, email address supplied

Read your column in 'the inq' ..

Dear me, it reads like old hollywood, where stars had to suck up to the reviewers, if they wanted to stay popular..

These days, most people rely on personal feedback, or someone who has good experience in the industry - barry norman etc.

Most people treat benchmarks as only a guide, but with the 'net, there is now a lot of choice! I prefer anandtech and tomshardware, but it would be good to see your experts do some benchmarking of basic stuff. (hard discs, mobos, etc)

cheers, bob.

Email address supplied

I'm not surprised, because Windows 2000, XP and 2003 all have special versions of some components for single-processor systems. These components allow the OS to take shortcuts (for example, disabling interrupts on the processor or changing the Interrupt Request Level [IRQL]) on single-processor systems that aren't safe on multi-processor systems. On Windows 2000, the files Kernel32.dll, NTDLL.DLL, WinSrv.dll and Win32k.sys have differing implementations. There are also separate builds of the NT kernel (NTOSKRNL.EXE, NTKRNLPA.EXE) and the various hardware abstraction layers (e.g. use of halaacpi.dll rather than halmacpi.dll) for uniprocessors and multiprocessors.

With the P4 in HT mode, it pretends to the OS that it is a multiprocessor. However, it isn't. It can only execute some operations in parallel due to limitations on the number and type of execution units. IIRC, in non-HT mode, all execution units are available to the single instruction stream. Also, I believe there were no new execution resources added to the HT P4s over the non-HT parts. Code optimised for P4 without HT which uses all available resources is liable to stall on HT because some resources are being used by the second instruction stream.

Nevertheless, for safety, the OS must still use the multiprocessor versions of these components. These code paths are significantly slower.

As a result, you may gain a certain amount of parallel processing, but you lose it all again in system overhead.

It _may_ be possible, with newer operating systems coded explicitly with HT support, to use lower-cost code paths if it is known that there is a single HT processor in the system.

Most of the time you don't get a lot of benefit with desktop apps anyway, because the tasks performed in the background are normally things like saving, printing, sending messages etc which are mostly blocked on I/O resources and actually have very little computation to be done.

My personal opinion is that HT is pointless. If you need parallel processing, use multiple processors.

Of course, if you want your SETI@home or Cancer Research project to complete quicker, go right ahead ;)

Mike D
Email address supplied

Don't quote me, but this is frequently true. It applies if - and only if - the application being benchmarked does all its work in a single thread. However, this is pretty damn common with today's software.

If an app spreads work over several threads (and some do, even if not designed for SMP systems) or it causes the OS to generate a number of threads, then HyperThreading can help performance.

As an Intel engineer, who I'm afraid I can't name, said to me last year: "With HyperThreading on, you effectively get two processors, each of which is less powerful than the same chip with HyperThreading off".

This means that HyperThreading is quite useful for file/print/ web serving, but it's often not helpful for the kind of heavyweight calculations that cause apps to be deliberately written for SMP, and it's actually useless - and sometimes counterproductive - for single-threaded games and the like.

This isn't quite what Intel marketing say. Should it be a surprise that a marketing department selling something complex and technical is over-simplifying?

Name, email, supplied

Hello there.

I invented the iLoo, the tablet PC and the smart display.

This iLoo inventor is a joke, can someone tell him to quit whining. Probably everyone in the UK invented an iLoo at one stage in their life.

He has provided great entertainment with his whining though

Jaz
Every corporate email supplied

Hmm, when I clicked on "flame editor" I expected to see something like this:

Dear Mr. Magee,

I think your web site is terribly biased against [Intel|Microsoft|Big business|Masons|Black people|Star Trek fans]. You are a [Bad journalist|Poopy head] and I feel that it wouldn't be unreasonable if you were to [Check you facts thoroughly|start telling the truth about how great Windows is|start telling the truth about how great Pentium 4s are|start telling the truth about your fluency in Klingon|fornicate with a stick full of rusty nails]. That would make the world a whole lot better and I could go back to downloading MP3s and putting the American recording industry out of business.

Yours sincerely,

I. Punctuate

P.S. Microsoft Word is the best application ever. Most other office suites wouldn't even come close to filling one CD! And backups are for losers. I've never lost any data to any kind of cr

Hi Mikee!

Fudo's reporting has been bang straight out accurate on the FX 5900 specs. Checked on www.XBitlabs.com and the goddamn guys have actually done a full fledged review on it and no they didnt fry any eggs on the chip ;-). But if you check the readings out properly keeping in mind that it was 3D Mark 2003 that was used & which was derided by Graphzilla sometime back, I wont be surprised if there was some good ol "tweaking" of drivers to the point of crafty "detonation". Anyway the ATI underdogs held up well and now Im just waiting for the final retail prices. Its always nice to extrapolate the pricing factor with the performance that you are going to get.

If Geforce FX 5900 at $ 499 vs ATI Radeon 9800 Pro at $ 399 vs ATI Radeon 9700 Pro $ 299, then the ATI Radeon 9700 Pro is your best cost to performance ratio bet. Rest assured, for the simple reason that as we will have more pipeline thwacking texture rich cum poly rich games, at one point says about 2 years hence the difference in performance between these chips will boil down to a coupla frames thats it. So where u gonna lay your greens out pardner?

Hmmm now what was that talk on the optimisations that Graphzilla has incoprorated into its newest kid on the block?

Regards

Aaron Mathias, email supplied

Dear Mike,

I recently paid a premium for a Barton processor (2500+) and an ASUS A7N8X deluxe motherboard. I kept my lowley Geforce TI-4200 video card and invested in a 52X cdrom drive($20) and a 80Gig WD 8MB cache hardrive (80 bucks after rebate), I still have to install those parts. My take on the Barton with a Gig of memory in an Asus A7N8X (Deluxe) motherboard under Win2K (98SE can't cut the mustard) is blindingly fast. I cannot describe how good the Barton is under a multitude of loads and performance conditions.

Looks like AMD has kept its promise in delivering a world class processor.

Just an opinion from a long time reader, I am very happy to take this machine into a lan-fest two weeks from now.

Hey if I sound shrill, those KT400 boards are looking rather lame in their first incarnation.

Nforce2 with 333 memory is damn good right now, very fast and cheap.

Doom III may change all of that, however I feel that the most I will have to do is change my video card.

Regards as always,

Name, address, supplied

http://www.theinquirer.net/?article=9463

What about the fact that the graph starts at 80% ? That makes the differance seem 5 times what it actually is unless you look at the numbers.

3200 is 6.67% more than 3000, however the XP3200+ only got 3.9% more than the XP3000+. Are AMD accepting the idea the PR rating is wrong ? hehe

Now let's talk about the poor people who actually buy these CPUs. You can get an XP2400 or P4 2400 for about a third of the cost of one of these high end chips, i f not less, and still get around 80% of the preformance. It's true you need to talk about the total cost of the system and not just the CPU, but still I find it hard to believe people actually pay 500$ for a CPU when in fact in the avarage user's usage the CPU has about the least noticable impacts (after the VGA card, memry, motherboard and HDD).

Let's get back to the fact that 3200 is 6.67% more than 3000. The price differance between a P4 3000 and 3200 is about the same as the one between an XP3000+ and and XP3200+ which is about 150$. That's 30% more for 6.67% more preformance. It is normal for this situation to happen in the high end sector and very understandable. However, 6.67% is too small a differance, some motherboards come with over 2% factory overclock for instance so 6.67% is really getting too small. If they keep it up this way then by the time they reach 4GHz they will be selling CPUs with 5% differance in speed.

One upon a time CPUs would be speeded in half's of a multiplier. But since then we have been introduced to double and quad pumped FSB's which allow for these tiny speed steps.

I say it's about time they went back to the old speed grading scheme in intel's case, and AMD's lower FSB calls for a whole multiplier at a time.

Everything supplied

Dear Mike,

Who is this guy? His credibility vanishes in one sentence - the one where he cites the 8088 as a 16-bit processor with 8-bit addressing. It was 20-bit addressing by pages to use 16 segments of 64k. Etc. The article is rambling, confused, vague, and has numerous strange statements that indicate a severe lack of familiarity with computers or understanding of common industry terminology.

ND, email supplied