Slashdot Mirror
For AMD Success Means Problems
An anonymous reader writes "AMD's success with its dual-core Opteron and Athlon processors has created something of a happy problem for the company. It can't make its products fast enough to meet demand. Just the same, with the Intel price war heating up and new 65-nanometer manufacturing technology being implemented in its factories, AMD has a lot of balls in the air right now." From the News.com article: "AMD's current pickle is the result of its success, which makes it a little easier to swallow for company executives. Demand is high, but the company's dual-core processors still use its 90-nanometer manufacturing technology. Intel's chips, on the other hand, are built using the smaller transistors provided by its 65-nanometer manufacturing technology. Not only is AMD using larger transistors, but its dual-core Opteron and Athlon 64 processors contain two processing cores integrated onto a single piece of silicon, or a die. This design has given AMD great performance during the past few years, but resulted in processors that were almost twice the size of its single-core chips."
You don't really understand how they make chips, do you?
It costs $X to make a wafer of CPU's. The more CPU's you get from that wafer, the cheaper each CPU costs. Large CPU Dies means fewer CPU's per wafer, thus high cost per wafer. Thus, each CPU die has a higher cost to manufacture than smaller dies.
If you need web hosting, you could do worse than here
2. Larger size = higher average chance of defect per die
3. Larger size = more expensive to manufacture per processor
So you see, size DOES matter :)
Space for rent, inquire within
Thing is, Intel is already shipping 45nm chips, though 45nm CPU's won't be shipping for a while yet. It's already working on sub-45nm technology.
If you need web hosting, you could do worse than here
Yup, this was one of those "bullet point" items why Apple went with Intel over AMD. Only Intel (and IBM) have the kind of manufacturing capacity that Apple needs in order to keep the supply chain moving while still providing to other customers. It would/will take many many years of constant fab building for AMD to compete with Intel on production scale.
Space for rent, inquire within
It is also worse than that.
The larger the die the lower the yield.
Taken to the extremes, assume a huge die. A single defect and you lose the entire wafer.
On a super tiny 100 dies/wafer a single defect only takes out 1%.
Of course on large dies, you might have small pieces of redundancy built in (all memory has this) so you can recover from some number of defects before you have to scrap it.
But still, the larger the die, the fewer you can make at a time and the lower the yield on the ones you do churn out.
Increased supply requires increased production capabilities, which more or less requires a new fab (factory) being built. No, generally you can not just add additional tools to an existing factory. Nor can you just make your workers work harder/faster. Most of the steps in the process are automated and require definite process times. So it's going to cost you. A state of the art fab today runs ~$4-5 Billion. It also takes well over a year to bring it up to full production status (depending on how many tools you have. Intel's Fab 12 had 60 diffusion furnaces, Fab 22 across the street had 38). And that's part of the problem. A fab doesn't make money unless it is running at capacity.
I worked AT (not FOR) Intel for 3 years.
It means more than that...
First, you are correct. If you get more chips per wafer, you can make more chips. Since the time to process 1 wafer remains consant. However, there is also more going on.
The second thing to worry about is the cost. If it costs (making up numbers) $100 to process a wafer and you get 10 chips, it is $10 to manufacture 1 chip. If you get 20 chips from the wafer, then it only costs $5 to manufacture 1 chip.
The third item is quality control. If there are any flaws in the wafer, the chip that is created over that flaw can not be used. So that chip gets thrown out. If we can get more chips from a single wafer, our percentage yield increases as well. Imagine that there is 1 flaw per wafer. If we only get 1 chip per wafer our actual yield is 0%. This would be very bad. Now imagine we get 2 chips per wafer. While it is possible that the flaw would affect both chips, most likely it will only affect 1, giving a yield of 50%. If you get 3 chips, your yield is 66%. This yield really hits the bottom line.
If you are losing 2-3 chips per wafer from flaws, then any increase in the number of chips is going to increase your yield percentage. If you call it a 25% increase in chips on a wafer, due to the 65nm instead of 90nm process, the percentage of chips lost to flaws will also go down and you just made more money.
Reading code is like reading the dictionary - you have to read half of it before you can go back and understand it.
As if Apple sells enough computers that it would really matter anyway...
Hmm, Apple's US market share is about 5%. AMD's US market share is about 18%. So had Apple gone with AMD, they would be accounting for about 1/4 of AMD's sales in the PC market. That would almost certainly make Apple AMD's largest customer.
AMD is converting Fab 30 in Dresden from 90um and 200mm wafers to Fab 38, with 65nm and 300mm wafers. This should come on line in 2007. Longer term, AMD is building a new fab in upstate New York for 32nm features on 300mm wafers. That should come on line in 2010.
Meanwhile, AMD's main fab, Fab 36 in Dresden, is starting to produce 65nm features on 200mm wafers. AMD is also outsourcing some production to a 65nm fab in Singapore.
Down at the user level, this means that first shipments of AMD CPUs made with 65nm technology should appear in December of 2006. Coming soon to Dell Dimension desktops.
Actually, for manufacturers, size matters in a big way.
Semiconductor manufacturing is, like with most manufacturing, an imperfect process. In general, for a given die (a die is one "chip" before it's placed in/on any kind of case) and a given manufacturing method there is going to be a manufacturing error rate that is measured in terms of errors per unit area.
These defects can stem from everything from a speck of dust getting into the system, all the way to a gas depositing process making a trace too thin (so the trace, and thus the chip, burns up during use.)
Therefore the bigger the die is the higher the likelihood that any one die will contain one of these manufacturing defects. Since a defect can occur at any point on the wafer, the less of the wafer that can be associated with any particular common point of failure (i.e. the smaller the die,) the more valuable any particular wafer can be and thus increase yield and profit of that product.
Fab 36 is ALREADY online, albeit if at only 50% capacity, and using 300mm wafers.
Fab 30 is currenty running at full capacity, but next year it will drop to 40% capacity as it switches to 300mm wafers & gets echristened Fab 38 in 2008.
AMD is/will be outsourcing production to Chartered even though they just brought their new fab 36 online.
I think Nikkon is also a big semiconductor equipment manufacturer. http://www.nikonusa.com/template.php?cat=3&grp=79
Since then, anytime the get a good product, they blow it on production, leaving Intel to fill the void they created.
There's a difference between a production problem and a capacity problem, though they both result in supply failing to meet demand. A production problem is when you theoretically could manufacture enough parts to satisfy demand, but your process, technology, or chip design is flawed and is not reliable enough to meet production goals. This is especially bad when you tell your customers that they can expect so many chips, and then you fail to deliver them. A capacity problem is when you couldn't possibly produce enough chips
AMD has had plenty of production problems in the past, when their market share was low enough they've had the fab capacity to produce enough chips, but they were simply unable to. When this happens, you have to fix the design, fix the process, tweak the technology parameters, whatever to correct the problem. While this has hurt AMD quite a bit at various times, they did recover, and it didn't fundamentally limit their marketshare.
Now, AMD has reached a point where they simply have a capacity problem. Their production process and yields are very good, but their market share is such that Fab 30 is physically incapable of producing enough chips to meet demand, even if it were working perfectly and every chip on every wafer was good. Their market share is now capped. When this happens, there's only one thing you can do: Build more capacity.
AMD hasn't had this kind of problem before. They would have loved to have this problem instead of the production problems they've had in the past.
It is where they have failed again and again and again. I can't believe they haven't learned yet.
Oh, but they have. In particular, they anticipated this problem years ago and started to build Fab 36, which is about to come on line right when their capacity issues are becoming a serious limiter to their growth. Fab 36 is going to mean a lot more capacity for AMD, and let them grow their marketshare by a lot. Now they may have production problems with the new fab, but once they get those worked out they won't be having any capacity problems for a while -- and if they do, they will giddily ply investors to help them build a new fab.
P.S. speaking of investors, the point kfg was trying to make is that a capacity problem can only be solved by increasing capacity, which requires investment in a new fab. If investors fail to invest in the extra capacity, and the company hence cannot meet demand and fails, then it is the investors who are responsible. Somewhat; after all it is the company's job to seek out investors to help them build a new fab.
The enemies of Democracy are
A General Protection Fault is not a fault in the sense of a hardware flaw. It means that a process has attempted to access a segment for which it does not have the correct rights. Assuming correctly written software, the most likely cause of these is memory errors. If you are suffering from a lot of them, then I suggest you try switching to ECC RAM (or, at the very least, run memtest86 on the worst offenders for a day or two). You can typically monitor the number of errors fixed by ECC memory in software, so this should give you some indication of how useful it is being.
If some machines are getting GPFs more than others, it may just be that their memory errors are overlapping the code segments more often; other machines may be silently corrupting data more often...
I am TheRaven on Soylent News