In the last week, AMD began to produce CPUs at their second Dresden fab.
They had a little gathering to talk about it, and as we greatly suspected, said nothing more about 65nm conversion than they have in the past. Indeed, if anything they indicated a bit of a delay, now saying that 65nm conversion will be “substantially” rather than “completely” converted to 65nm by the middle of 2007.
The word “substantially” leaves a lot more wiggle room than “completely.”
Well, we believe we’re not going to see a substantial number of 65nm chips from AMD until sometime in 2007, and they said nothing this week that would change our minds.
But that’s not what we’re going to dwell on today. Instead, we’re going to talk about the capacity of Fab 36.
1) AMD says it’s going to make, one way or the other, 100 million CPUs by 2008. A previous roadmap showed the following:
2) AMD says that by 2008, Fab 36 will be making 20,000 300mm wafer starts a month. AMD refused to indicate how quickly Fab 36 would ramp up.
3) AMD says that Fab 30 is now making 30,000 200mm wafer starts a month.
4) Fab 30 is now making CPUs at a pace of about 50 million/year.
4) Do the math, and you’ll find that the total area of silicon to be fabricated by Fab 36 by 2008 is only 50% more than than the area of silicon fabricated by Fab 30 today.
5) AMD has given no indication that Fab 30 will be converted to 65nm production.
AMD boasted that Fab 36 would give a big boost to overall production. This will be true, but only so long as AMD makes a lot of 90nm chips.
As AMD phases out 90nm production, they phase out Fab 30, so they phase out the number of chips Fab 30 makes.
Under normal circumstances, a shrink to 65nm would allow a fully running Fab 36 to make 100 million CPUs, but we are not in normal times. We are shifting to dual-cores, and shifting to dual-cores eats up all the production increase you get from the shrink.
In other words, roughly the same amount of silicon space is taken up by one single 90nm CPU/a dual-core 65nm.
Do the math, and you’ll see that Fab 36 at full throttle should make about 75 million 65nm dual-core CPUs a year, which is just about what last year’s roadmap shown above says.
Two unknown factors will affect this number. First, it’s possible, if not likely that AMD CPUs will end up with more cache. If so, this would reduce the number of chips made (and may well explain why AMD is suddenly interested in L3 cache again).
More recent roadmaps indicate that AMD has quietly dropped dual-core Semprons. This would increase the number of chips that could be made.
In all likelihood, the two will more-or-less cancel each other out.
These numbers explain why AMD is going to need production from Chartered the next few years. Fab 36 alone will not be able to make 100 million CPUs a year until a 45nm shrink.
What they don’t explain is why AMD outsourced production to Chartered rather than convert Fab 30 to 65nm. Obviously, Chartered isn’t doing this for nothing.
The only conclusion one can reasonably draw from this is that AMD couldn’t afford to build Fab 36 and convert Fab 30, so they’ll pay Chartered extra for some extra capacity.
If that’s the case, if money is what is driving behavior, here is what is likely to happen over the next year:
Fab 36 won’t ramp up very much so long as it’s at 90nm. It will probably be used as a safety valve to take the pressure off Fab 30, which seems to be a rather overclocked fab.
The main effort at Fab 36 will be to get 65nm up and running. So long as Intel is still primarily selling Prescotts (which should hold true for the rest of 2006), AMD shouldn’t be hurt too much by the lack of 65nm.
The real crunch occurs in early 2007. Conroe should reach crossover (i.e., 50% of desktop chips) sometime in Q1, and 90nm Hammers will start looking decidedly second-rate. However, just when AMD will need lots of 65nm chips, they’ll have to phase out Fab 30 and lose that capacity.
Can Fab 36 and Chartered pick up the slack? It’s going to be close.