To understand AMD’s dilemna the second half of the year a bit better, perhaps we should inject some numbers into the picture.
Until fairly recently, the mainstream A64 single-core chips averaged, very roughly around 90 sq.mm. in size (averaging out the 512K and 1MB sizes), and AMD probably didn’t get much less than $150 average for them; let’s say $140.
For the second half of the year, dual-core processors will become the norm. Even the 512K AM2 version of these chips will come in around 180 sq. mm, or double the space taken by the single core chips. Assuming additional price cuts don’t take place, maybe they’ll get $200 average for them.
To keep the math very simple, let’s take 9000 sq. mm. of die space, assume 100% yield, and see what happens under both situation.
In the first, AMD makes 100 processors (9000/90), and gets $14,000 for them (100*140).
In the second, AMD makes 50 processors (9000/180), and gets $10,000 for them (50*200).
Going to dual-core after price cuts is obviously not a good thing for AMD.
You may say, “But AMD is expanding its capacity.” This is true. Let’s see how much more capacity AMD needs to a)make the same amount of money from dual-cores and b)make the same numbers of CPUs from dual cores.
To make the same $14,000, AMD would have to make 70 dual-core processors. That would mean they would have to expand capacity 40% just to get the same revenues they used to get from single-cores.
Mind you, revenues are not the same thing as profits. In this situation, AMD would end up with less profits simply because they had to do 40% more fab work (and spend a good deal more doing it) than they did before.
AMD would also lose marketshare because even with a 40% increase in capacity, they’re making 30% fewer CPUs. To keep making the same number of CPUs, AMD would need to increase their capacity 100%, which they can’t do in the next six months.
Now reality won’t be quite that bad, simply because AMD isn’t going from 100% single-core to 100% dual-core in the next six months. Realistically, the shift won’t be 100%, but more like 40%. But even then, you can see how making more dual-cores can easily chew up any additional capacity from Chartered and Fab 36 and why AMD so desperately wants to shrink die sizes.
And yes, I’m leaving out a lot of minor factors to make this as clear as possible, but don’t miss the forest for the trees: this is the major factor driving AMD decisions at the moment.
You might ask, “But doesn’t Intel have the same problem?” and the answer is essentially “No” because all their dual-core production is at 65nm. Roughly, a dual-core 65nm chip takes up no more space than a single-core 90nm. The Conroes can be packed with 4Mb cache and still be somewhat smaller than AMD’s 90nm 1Mb cache X2s.
This is why AMD so desperately needs to get to 65nm, and not being able to do it until long after Intel is a major failure on their part. AMD can whine about relative yields and the like, but a process shrink covers a lot of faults. What matters isn’t what Intel got a year ago, but what they can do now, and now they can pump out the chips and raise the bar knowing that AMD will have to struggle enormously just to keep up. Intel didn’t have to wait until IBM got 65nm SOI down right; they could afford to do the heavy lifting on their own.
You might ask, “But doesn’t this problem go away when AMD finally gets to 65nm?” The answer to that is “Yes and No.” It’s “Yes” in that a process shrink will certainly double the capacity of those fabs using 65nm technology, which would mean Fab36 and (hopefully) Chartered, but Fab30 will have to go offline to be rebuilt, and at best, AMD will lose most of that capacity for about a year.
Overall capacity at 65nm will go up somewhat, but by nothing huge.
It’s also going to be interesting to see how AMD is going to pay for the Fab30 rehab because the one problem with these price cuts (for both AMD and Intel) is that any price cut is essentially a profit cut. If you make a processor for $50 and sell it for $150, you make $100. Start selling it for $100, and it still costs you $50 to make it, so you now make only $50 from it.
Very, very roughly, here’s some back-of-the-envelope calculations on how much AMD gets and where the money goes per CPU:
AMD gets roughly $100 a CPU (keep in mind all those sub-$100 Semprons and big OEM deals). It costs them about $40 to make it, which leaves AMD with %60. Out of that $60, they spend about $20 for R&D, and another $20 to run the company, leaving them about $20 before taxes and interest expense.
So AMD really “makes” about $20 a CPU. If they start getting an average $20 less per CPU, well, no more profits.
Intel, on the other hand, gets a little less than $150 per CPU, and after all the expenses are accounted for, makes maybe $50 a CPU. So they have more leeway in cutting profits than AMD, and if worst came to worst, could sustain a few pretty big losses better than AMD.
You might ask, “Isn’t Intel going to pay a price for all this?” and the answer is “Yes, but it’s one they can take.” Intel is going to have a couple relatively horrible financial quarters the rest of the year, much worse than they already have. But we’re talking about slashed profits, not losses, and investors know there’s light at the end of the tunnel.
In any event, AMD will be OK for a couple quarters until Intel gets to 45nm, and probably starts mainstreaming quadcore by early 2008, and the same process shrink advantage cycle starts all over again.
But this time, Intel won’t have the anchor of Pentium Ds hanging around their necks.
The point of all this is not to say that AMD is doomed, but to point out that this is a company that’s going to have a hell of a time staying where they are the next couple years mostly because Intel has money and they don’t, and can make AMD go places they don’t want to go, both technologically (dual core) and financially (big price cuts), because of their financial and market strength.