Bye, Bye, Cache . . .

Add Your Comments

OK, it’s not an unofficial phaseout of 1Mb AMD desktop chips, it’s an official one. If you want a desktop with 1Mb cache from AMD, it will have to be an Opteron (which will not be getting any known price cuts).

And, of course, the choir of fanboys sing, “What a good move this is! Less cache is better!”

Well, while this is probably not going to have much impact on this audience, it’s not a good move, simply because it’s a move made out of weakness.

First, it was another last-minute move on the part of AMD. Think about the 4000+. How long a life did it have, a month?

Second, and more importantly, the Semprons that have a big whopping 256K cache are being aborted, too, in favor of those with 128K.

For those who say, “Cache doesn’t matter with Hammers,” let me ask you this: “If it doesn’t matter, why did AMD put them out in the first place? If getting rid of excessively cached chips is a good idea now, it must have been a blunder to build them to begin with and then keep building them.

Finally, if less cache is so good, why aren’t the Opterons being downsized, too?

The Real Reason: Capacity

Whenever AMD (or for that matter Intel) does something that doesn’t make a lot of sense, rest assured that they’re being forced by circumstances into doing it.

For the next year or so, AMD is going to be faced with a series of capacity crises. Yes, I know they’ve been saying the exact opposite, but hear me out:

For the next few years, AMD is trying to roughly double the number of CPUs it makes while the CPUs themselves will be doubling in cores (and size) a couple times. That’s a double whammy on production capacity, since a doubling of cores on a CPU effectively eats the die space you save on a process shrink.

The first crisis is being caused by a combination of success and lack of fab capacity due to the mainstreaming of dual-core processors and AMD’s inability to make 65nm processors yet. Yes, Fab 36 is making a few 90nm chips, but I would bet dollars to doughnuts the number of chips being made is nominal, after all, they have to make 65nm chips shortly. Yes, Chartered is making a few, but they’ll have to convert to 65nm sooner rather than later, too.

A shift of even 20% in the production mix towards dual-core probably will eat up any additional capacity provided by Fab 36 and Chartered.

By early next year, Fab 36 ought to be making serious quantities of 65nm chips, but it will take a few quarters to ramp up, and in the meantime, Fab 30 will be phased out. AMD will have the same problem in 2007 with its 65nm chips that Intel will have with Conroe, etc. in 2006: the bulk of their capacity will remain with older chips.

Even by the latter part of 2007 going into 2008, Fab 36 really isn’t all that terribly big, and if AMD really wants to make anything close to a hundred million chips a quarter, Chartered is going to have to be the ones who’ll have to really ramp up.

Meanwhile, so long as Intel can get to smaller process technologies long before AMD, they can continue to torment AMD with developments like quad-core.

I suspect this is part of the reason why AMD seems so hyped up about L3 caches for future chips, they’re looking for anything that will help keep CPU sizes down.

Reducing the size of mainstream chips is an immediate way to do the same thing, and that’s the main reason why this is happening (selling a few more expensive Opterons probably doesn’t hurt, but that’s not what is driving this).

Ed


Leave a Reply

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