What The Clawhammer Delay Is And Isn’t
AMD has had tremendous problems developing Hammer. They still are. They can make the things, but they can’t make them run very quickly.
An article over at XBit Labs purports to have read an AMD response to the delays which says in part that SOI is the problem.
There’s no link to the original document, but that seems to be the most likely culprit (with the memory controller being Suspect Number Two).
It’s simply a matter of elimination. It’s not that AMD can’t make .13 micron chips that can run a good deal faster than 2GHz, they’re doing about 2.4GHz now with TBredBs, and at least the high-end Barton should be able to do the same.
How is a Clawhammer significantly different structurally than an Athlon? It uses SOI and has a built-in memory controller (the x86-64 circuitry is actually a minor change in comparison).
So it is very likely that one of the other has a problem.
If your PC won’t boot, you have a problem. What is causing the problem could be a whole number of things.
It’s the same thing here. You could have design problems. You could have manufacturing problems. You could have performance problems. You could have yield problems, period. You could have yield problems at high speeds.
That’s a lot of room for a company to hide in. They can say, “Oh, this isn’t the problem” without saying, “but that is.” For instance, AMD kept telling those who went on that Dresden tour that their Hammer “yields” were fine.
That’s nice, but that’s not the single and only question. Being able to make them is better than not being able to make them, but being able to make them doesn’t mean you can make them fast.
Saying that yields are fine doesn’t cover all problems; it just covers that problem. The question that should have been asked after hearing about yields was “What are your yields of Hammer able to do 2GHz like?”
If you don’t ask the right questions, you end up getting suckered like those visitors did.
To be fair, I suspect this strategy in real-life would end up being, “Make them lie to you.” which is morally better than “Don’t ask, don’t tell,” but probably not much better in the getting-a-real-answer category, and probably a lot worse in the being-invited-back category.
The Two-Headed Beast
At the beginning of 2002, what AMD wanted to do was very clear. They wanted Hammer to become the star, and what’s socket A?
When they found out Hammer couldn’t even come out, they went to Plan B, which was to tweak a little more out of the Athlons and revive Barton. They had problems doing that, but by the end of the year, they succeeded.
Again, it was a matter of solving one problem, not all problems.
AMD solved the problem of having nothing competitive to sell, but by solving that, they created another.
Instead of having nothing competitive to sell, now they had the prospect of two products to sell, with the “old” line close enough in performance to the “new” one to remove most of the reason to buy the new.
CPUs Don’t Age Like Wine. They Rot Like The Things In the Back of Your Refrigerator
CPUs just don’t get better with age. They just get old.
Had a 3400+ Clawhammer come out a year ago, it would have been a stellar chip, even at a high price, and a real problem for Intel.
Does 3400 at a high price look better now or even as good now than it would have a year ago? Do you think 3400 is going to look better six months from now?
Of course not.
If you’re behind, and you have a world-beater, you get it out right away. Period. Anything else is incompetent nonsense.
It’s only when you’re ahead, can stay ahead without the new product, and get significant benefits from delay that you wait a while.
That could well be the reason why Intel is delaying Prescott. They don’t need it until the end of the year because AMD isn’t pushing them. At the least, they can defer the big fab refurbishments for a while. At most, it gives them more time to work out the increasingly difficult task of mass producing processors with smaller and smaller circuitry.
This may prove to be a real problem for Intel, and could have been part of the reason for the Prescott delay, too. A lot of companies have had real problems handling .13 micron chips. Going to .09 micron will be even rougher.
Imagine you’re in a game. You’re getting beat. You stumble upon a weapon that’s a lot better than the ones your opponents are grinding you into the dust with.
Do you say, “I haven’t gotten my ass kicked enough, I’m going to wait a while,” or do you use it?
You use it unless you can’t, for a very, very good reason.
Hammer is like that game’s weapon. If AMD could have started firing their killer weapon a year ago, they would have. They didn’t because they couldn’t.
Now it’s not such a killer weapon anymore, and you’re not too sure it just won’t blow up in your face if you use it now. In the meantime, you’ve come up with cheaper weapons that work well enough to keep you in the fight.
The best thing to do is keep working on the killer weapon until it catches up and surpasses your enemy’s stuff, but that’s not an optimal solution; it’s just the least bad.
This keeps you afloat for a while, but that’s about it. It certainly is no swift, crafty move on AMD’s part. It buys some time, that’s all.
Nor is emphasizing server chips over desktop chips some stroke of genius, either. You never choose one over the other if you can make both. It’s when you can’t that you make such a choice.
What are damaging speed disadvantages for the desktop aren’t as much so in the server market, and Sledgehammer certainly has other non-speed advantages.
But server sales can’t pay AMD’s bills by themselves. There’s just not enough of them, and AMD will just be breaking into the business.
It’s certainly nice-to-have, and can be a great supplement to desktop sales and profits. It’s just not a replacement for desktop sales.
For now and the foreseeable future, desktop sales will have to keep AMD afloat.
A Dual Challenge
If the XBitLabs article cited above is correct, SOI is AMD’s big problem now, and that’s why they’ve partnered with IBM.
IBM certainly has a lot of SOI experience, but from what they’ve put out, they don’t seem to have the answers AMD is looking for ready, either. AMD needs to be able to make desktop SOI chips that are fast and cheap to make. IBM doesn’t seem to have the second part down.
It’s not like there’s anybody better around, but this partnership has its work cut out for it.
Within the next eighteen months, AMD is going to have to do both the following:
If AMD can’t do both, unless Intel stumbles badly in .09 micron migration, they’ll be right back to having no answer to Intel on the desktop.
And I don’t think they can survive if 2004 becomes a repeat of 2002.
All copyrights belongs to their respective owners.