There’s an article over here which purports to explain Why Hammers Have To Be Expensive. A number of people wrote to me to say that it proves what I’ve been saying about Hammers is incorrect.
In essence, it says that because Hammers (especially those with 1Mb cache) chew up a lot of die space, and because AMD only uses 200mm wafers rather than Intel’s latest 300nm, Hammers are doomed to be expensive and relatively rare until 90nm, and maybe even after that.
This article makes two major points. The first basically says, “Gallons of milks cost more than half-gallons of milk.” That’s fine and undeniable, but that’s not the question. The question is “How much more does the gallon of milk have to cost?”
The second point basically says, “You can only get half the number of gallons out of a cow than you can half-gallons.” Again, the point is self-evident, but that’s not the question. The question is “How much milk can the cow make, and are you milking it enough?”
Let’s examine these points:
In the case of the first question, why yes, a 190mm² CPU does cost more to make than a 110mm², but this is pretty useless information unless you know how much the 110mm² costs to make.
If all you’re concerned about is the wafer, the difference in cost, yields being equal (and AMD has said that’s the case with Hammers) the difference in marginal cost (that means how much extra you have to pay to make an additional processor, not counting fixed costs like the cost of the equipment) between the 110 and the 190 is roughly $10-15 per CPU.
Once you get the equipment, CPUs don’t cost a lot to make: a figure of $25 for each extra CPU about Barton size is roughly right. Matter of fact, that’s just about what it costs Intel to make a Northwood using a 200mm wafer process, and Northwoods are rather bigger than Bartons, about the same size as Newcastle, and 3/4 the size of Clawhammers. See for yourself.
Using marginal cost is probably not the best way to cost out these processors. A better rough rule of thumb (at least according to conventional cost accounting) is probably to figure out how much AMD needs to get for a 110nm processor, including all the costs allocated to the CPU division, direct or indirect, from wafers to equipment depreciation to Dresden salaries to Hector’s bonus. For a Barton, that probably comes in around $60. This would make the figure for a Clawhammer about $110, and about $85 dollars for a Newcastle.
Like a gallon of milk, the cost of which covers everything from the feed to the supermarket’s profit, it’s fair to assume that a gallon of milk, which is double a half-gallon, will cost roughly twice as much as a half-gallon.
Perhaps a better way to look upon this is to see what other Hammers cost. The lowest speed Opteron can be bought at Newegg for $158. Take out the free FedEX, and the cost is probably about $150. This is not much above our $110 figure, and given that both AMD and Newegg are undoubtably making some money out of that $150, it indicates that our $110 figure isn’t too far off.
The point of this is not to come up with an exact cost figure. The point is to show that while Hammers are somewhat more expensive to make, they’re not that much more expensive to make. A gallon of milk costs twice as much as a half-gallon, not five or ten times as much.
The article also throws a bunch of numbers which would give the impression that by going to 300mm wafers, Intel gets overwhelming cost advantages. Not so.
Yes, as a percent of cost, there are significant cost savings, 300mm fabs can crank out chips for about a third less than 200mm. However, as we’ve seen, Intel can make a Northwood the old, expensive way for just $25. The dollar cost savings per CPU is less than $10.
Of course, if you make 130 million CPUs a year, saving $10 a chip adds up and is very good for Intel’s profits, but it’s hardly an advantage Intel can use to crush or even hurt AMD very much. If the average selling price of CPUs were $40 or $50, yes, that would be huge, but CPUs cost a lot more than that.
The article implies that yields on Hammers are very low. Well, AMD has repeatedly stated in their conference calls that their yields on Hammers are as good or better than for XPs.
Of course “Hammers” are a broader category than “2.4 GHz or better Hammers.” I wouldn’t be terribly surprised if AMD doesn’t get a whole lot of 2.4GHz, and despite the article’s contention, it’s probably going to really rough to get a lot .13 micron chips running at 2.6GHz.
But again, that’s beside the point. The complaint is not that AMD is charging a lot for its FX or its two high-speed socket 939 Newcastles; the complaint is that these are the only choices if you want socket 939.
What happens to all those chips that work fine but don’t quite make the high speed grade? Does AMD throw them out? If they did, and only if they did, then one can talk about low yields driving AMD’s costs skyhigh legitimately.
But of course they don’t do that. Instead, it looks like they end up as other Hammers, most likely lower-speed socket 754 chips. Essentially, the only difference between a socket 939 and 754 CPU is whether or not AMD fries the circuitry for that second memory channel.
That’s fine and dandy, but there’s no reason for such chips not to be lower-speed socket 939 chips. Enabling or disabling dual-chanel memory has nothing to do with CPU speed, and it’s pretty silly to disable a function to make undisabled chip “worth” more when your competitor doesn’t do that even with its bargain CPUs.
The article points out that AMD has only one fab which also has to make a lot of Athlon XPs to pay the bill, plus convert over to 90nm production. Take all this into consideration, and AMD has only a limited capacity to make Hammers.
All of this is true. It could also be pointed out that it may well not make a lot of financial sense to convert more of Dresden over to 130nm Hammer production just to rip it down a few months later for 90nm.
Again, this is true, but besides the point.
AMD originally planned to ramp up Hammer considerably in 2004. As they said in their January conference call, they had the capacity ready to make “substantially” more than a million Hammers. This is unsurprising since they intended in late 2003 to make about 1.7 million Hammers last quarter, and over five million this quarter.
Later, they massively scaled back their expectations.
So AMD originally saw no problem with ramping up Hammer capacity (and scaling back XP production) by this point in time back in September. Something made them change their minds big time.
It’s conceivable that technical problems caused this pullback, but that’s hardly good now, is it? It’s also conceivable that the costs of conversion for just a short period of time played a role, too.
However, the most likely culprit was a crushing lack of demand for the product at the prices AMD wanted to charge OEMs.
To be fair, it’s quite possible OEMs were looking for hypercheap Hammers; that’s the only reason I see for Paris coming out. Lack of a 64-bit Windows OS probably didn’t help, either.
Whatever the reason, AMD appears to have rather grossly undersold the rather limited Hammer production capacity it does have. We estimated 350,000 last quarter, that could well have been on the low side, but even if the figure were 500,000, that’s less than half their capacity. We’re pretty sure AMD didn’t sell anything like a million Hammers last quarter, because if they did, their CPU revenues would have jumped up a lot more than it actually did.
Again, to be fair, the numbers will probably look rather better this quarter due to increased OEM sales of fairly reasonably priced socket 754 systems with (hint, hint) cheaper Newcastle processors. AMD may well meet its 750K production estimate this quarter.
AMD in all likelihood has had, and probably continues to have spare Hammer capacity. That doesn’t make the company money; it costs the company money. From the newer production numbers, it looks like they’re not going to substantially add to it until about a year from now.
Since OEM demand is not likely to explode upward anytime soon, selling mainstream-priced socket 939s to primarily enthusiasts will simply make AMD extra money from unused capacity.
That’s why we’ve complained about this so much. If AMD were truly selling every Hammer they could make at the prices they’re charging; we wouldn’t be complaining and demanding they ramp up Hammer production skyhigh immediately. Believe it or not, I’m a reasonable guy. 🙂 I can see legitimate reasons why it’s turned out that it might not be in AMD’s best interest to ramp up bigtime at 130nm. If this were the case, we’d say, “God bless them, and I guess we’ll just have to wait for 90nm.”
However, that doesn’t appear to be the case. It doesn’t look like AMD has been even come close to using the limited capacity they already have sitting around. If that’s wrong, AMD can say how many Hammers they’re making anytime they want. It’s not like Intel can’t figure it out; I’m sure the professional counters tracking PC sales have a pretty good idea what those figures are.
This appears to be a matter of letting capacity go idle and getting nothing for it rather than using it to make what would still be quite profitable chips in the hope of getting a relative handful of hyperprofitable sales.
Why Apologize For the Unapologetic?
Never forget that AMD is a multi-billion dollar multinational corporation. They’re big boys; they are perfectly capable of defending themselves and setting things straight. They could in an instant set people like me straight if what we’re saying about the company isn’t true.
True, it’s extremely unlikely Overclockers.com is exactly a required daily read for AMD execs. 🙂
However, when major stock analysts keep asking AMD for precisely the kind of information this inquiring mind would like to know, and AMD refuses again and again and again and again to tell them and forces them to guess, that ought to tell you something.
Namely, that’s something is not going splendidly. If everything were going wonderfully, AMD would be getting a lot more money from its CPU activities than it is. There’s no doubt in my mind about that. Just exactly what that “something” is, that’s shaky.
It’s like EMS picking up someone who is having seizures, along with high blood pressure and pulse rate. The medic may not know exactly what’s wrong with the person or how serious it is, but he’s not going to say to the patient, “I can’t figure out what’s wrong with you, so you must be fine. Get out of my ambulance!”
Yet that is just what many AMDroids say. If AMD doesn’t tell you what’s wrong with it, you can’t say there’s anything wrong.
It’s not that AMD is terribly sick at the moment, but they ought to be doing better than they actually are if both Hammer and Athlon XP sales are robustly healthy.
For instance, it’s possible that Hammer sales are rather better than I’ve thought. However, if that’s the case, that must mean Athlon XP revenues are going to hell. Given the level of AMD’s CPU revenues, either Hammer or Athlon XP revenues can be healthy. It can’t be both.
So when people take it upon themselves to defend AMD like it’s some powerless little kitten; this is not like sticking up for someone who can’t defend himself. This is a matter of defending someone who won’t.