AMD Hits The Wall, Too

Add Your Comments

Some leaked data (no doubt from within some mobo company) reveals some important news about AMD’s future processors.

Easily the most noteworthy of these tidbits is the statement that the 90nm chips will chew up 104 watts at speeds of less than 3GHz (2.6GHz in the case of the FX, probably 2.8GHz in the case of an Athlon 64).

These wattage numbers are just as bad as Prescott.

Of course, these processors will be doing rather more work for the wattage than Prescott (perhaps around 20-25% more), so you’ll get rather more for your hell compared to Intel.

However, this does show the limitations of SOI. As we’ve suspected for a while, SOI basically buys AMD an extra generation of CPU-business-as-usual, and that’s that.

What we don’t know yet is what the power gradient is going to be when these SOI chips get pushed. A few months ago, a little birdie gave me some power consumption figures for 90nm PowerPCs. They ran quite coolly at around 2GHz, but pushed to 3GHz, the power figures skyrocketed.

We can no longer assume a more-or-less linear relationship in CPUs between speed and power when we overclock. What appears to be happening is that when you push 90nm chips beyond a certain point, they start hemorrhaging current, so that if you increase speed, say, 10%, you need not 10% but maybe 20% or more power to make it run.

Again, the situation is relatively better for AMD than Intel, but it’s pretty easy to see the same kind of heat wall stopping at least early heatsink-and-fan Hammer overclockers somewhere around 3GHz, and water-coolers at not much more than that.

Like Intel with Prescott, AMD hasn’t been promising a whole lot in speed increases from 90nm (3, maybe 3.2GHz eventually), but they may have their hands full delivering even that.

They’ll probably have an easier time getting to 3-3.2 than Intel will getting Prescott to 4GHz, but after all, we’re overclockers. If there’s a wall, we’ll hit in first.

Why The Sudden Love For x86-64?

Intel is now grudgingly going along with x86-64, by announcing that x86-64 compatible Prescott will become available for the same price as “regular” Prescotts.

Even Microsoft has been showing growing enthusiasm for the notion lately.

Why?

As we said last December:

“However, if you can’t ramp frequency as easily or by as much as you could previously, it’s not so easy to turn your nose up at the benefits of x86-64. You can laugh at a 2.8 or 3 GHz x86-64 when you can easily make a 5-6GHz PIV. You can’t when you’re struggling to make a 4GHz PIV, because that’s going to lose against the x86-64.”

This is turning out to be almost as true for AMD. Consider this: if x86-64 gets you an average 15-20% more performance, that’s just as much an improvement as going from 130 to 90nm.

In an era where CPU circuits create their own firewall when pushed too much, that’s nothing to sneeze at.

Which creates another problem for AMD . . . .

The House of Straw…

The House of Straw

For some reason, according to the article mentioned on page one, AMD is going to bunch a whole lot of (we think) different processors into its “Value” line.

Socket A processors will become “value” processors. Socket 754 processors will become that, too. There will even be (this is new) value processors for socket 939.

It’s a bit unclear whether this represents anything more than a name change for the socket A and 754 processors. It probably isn’t, because anything more than a name change could mean big trouble for the Hammer platform.

If, for instance, socket A processors started having built-in memory controllers (and you’d have to have grave doubts about how that could work with any current socket A boards), and they gave the kind of performance increase it gives Hammer, most of you reading this will buy that instead of any Hammer.

I suppose having a “value” processor for socket 939 is a good idea; it at least gives AMD the option of doing what they should have done in the first place, putting everything on one socket, and eventually dropping socket 754. After all, you can always have a single-channel socket 939 system just like you can still have a single-channel PIV system.

However, the big problem with the “value” line is that disables x86-64. We’ve always found that a stupid strategy because Intel could blow it up at any time but just enabling x86-64 support on its Prescott Celerons.

And that’s precisely what it looks like Intel is going to do. From here:

Mr. Otellini also said that Intel Corp. is in position to enable 64-bit registers across top-to-bottom desktop family of CPUs, including Celeron (our emphasis) and Pentium 4 products when Microsoft releases its operating system for such processors.

If you’re Joe Sixpack buying a computer at Best Buy, and you’re comparing a Celeron system with x86-64 compared with a “value” socket 754 without it, what are you going to lean towards.

If AMD doesn’t change its mind, we may see the spectacle of Intel beating up AMD in the shopping malls with AMD’s own stick in the one area (at least in the U.S.) where it is even between Intel and AMD (and this will be true to a lesser degree between PIVs and Hammer in the same arena).

How can AMD treat x86-64 capability as something tremendously valuable and worth paying a good deal more for if Intel ends up providing it in a $70 Celeron?

Yes, I know at least some of the reason for these castrated Hammers is OEM desire for them, but do you really think OEMs objected to having x86-64 built into the chip? I think not. What they almost certainly objected to was paying more for x86-64 enabled chips.

Yes, I also know that even castrated Hammers ought to beat enhanced Celerons, but the Joe at Best Buy doesn’t know that, and in any event, the performance gap will be much closer with AMD having one hand tied behind its back.

And yes, I suppose an x86-64 enabled value processor might prove a goldmine for overclockers, but what happens in the Best Buys or equivalents is a lot more important to AMD than what overclockers do at Newegg.

AMD’s strategy has been what I call the First Pig Strategy (as in the tale of The Three Little Pigs. They built their house of straw (aka x86-64) because it was the easiest thing to do and they charged high prices for it because they never thought the big bad wolf would show up.

Well the big bad Intel wolf has shown up, and by stooping to conquer, is now threatening to huff and puff and blow their house in.

It remains to be seen whether AMD realizes its house of straw no longer offers much protection against the wolf and now not worth very much, or Hector tells the wolf (and us) “Not by the hair of my chinny chin chin.” and get the wolf into some heavy breathing.

Unlike the little piggie in the story, though, AMD doesn’t have to rely on just the straw house. Hammer certainly has other strengths. It’s only if AMD tries to make a last stand at the straw house (by overvaluing x86-64) that the wolf will end up with a ham sandwich.

Ed

Leave a Reply

Your email address will not be published. Required fields are marked *