Silicon process technology and tools

[VirtualLarry]

Just wondering, after reading that GlobalFoundries will have 20nm on-line at around the same time that Intel is going to have their 14nm process on-line.

If it's just a matter of buying the tools, then why doesn't GF just buy 14nm tools from the same places supplying Intel, and skip right to the newest generation?

My naive viewpoint is that of someone buying a computer. If you're still on a C2Q, why buy Nehalem, when you could be buying Sandy Bridge? IOW, why, if a fab is behind on process tech, why invest in just next-generation tools, when they could be investing in next-next-generation tools?

 

[Charles Kozierok]

The short answer is that it isn't just a matter of buying the tools. There's a lot more to the process, all of which gets steadily more and more complex as the node size shrinks.

 

[esun]

It isn't just about making small transistors. It's about making them so that they perform well. I'm sure you could buy the most expensive equipment and pattern transistors with 14 nm gate lengths (in reality that doesn't refer to the gate length anymore but let's just go with it for simplicity). However, they will likely perform terribly.

The key is to have a process in place that produces transistors that perform well at the given node. For example, Intel's 45 nm node used high-K gate dielectrics with a metal gate. Their 22 nm node has been using tri-gate structures. In the research world the tri-gate (more precisely the FinFET) structure has been shown to scale down to 5 nm gate lengths. However, Intel has to come up with a process that ensures they can pattern a few billion of those onto each chip on a wafer and have an acceptable (i.e., profitable) yield. That's a very hard problem to solve.

I don't know much about the fabrication industry in terms of equipment manufacturers, but it wouldn't surprise me if Intel uses some equipment that you can't buy off the shelf.

Another factor could also be that the cost of a 14 nm fab could be prohibitive at this time. If it costs you 5 billion to build your 20 nm fab, but you think you could produce a 14 nm fab given an extra year and an extra 3 billion dollars, it may make more business sense to stay a little behind Intel and have less risk than skipping a node.

 

[bryanl]

Another factor could also be that the cost of a 14 nm fab could be prohibitive at this time. If it costs you 5 billion to build your 20 nm fab, but you think you could produce a 14 nm fab given an extra year and an extra 3 billion dollars, it may make more business sense to stay a little behind Intel and have less risk than skipping a node.

In the history of IC production, have the costs of a newer technology fab ever come down, either in absolute or chip output terms? I'm thinking of something equivalent to mini mills for steel production (poor analogy, I know) or IBM's shirtsleeve environment fab of the 1980s or 1990s.

 

[Denbo1991]

Intel does a lot of their process technology in-house. Where Intel excels is their domination in the process space. They aren't just going to sell it to their competitors to use. Attempts by other companies to accelerate their process so quickly can result in unacceptably low yields.

The costs of going to a smaller node have never gone down. It's quite the opposite, in fact. The cost of the equipment required grows exponentially as you try to move down to smaller node sizes due to the increasing complexity and precision required. For example, the improvement in optics required to move to a smaller process might increase the cost of the optics by an order of magnitude. Another example is the cost of a precision motor/actuator. Increasing the precision of your motor by a factor of 2 might again increase the cost by an order of magnitude. You CAN make the money back, but the initial investment grows larger and larger each time you go smaller. This is why a lot of ICs are still made on larger process sizes. A typical NAND gate you put on a breadboard doesnt need to have gate sizes of 22nm, so why would you risk buying a machine that costs hundreds of millions of dollars to do so over a machine that costs maybe a hundred thousand dollars?

 

[silverpig]

The answer is basically: If I gave YOU the tools, you wouldn't be able to produce a 14nm chip. There's more to it than just the hardware. Photoresist chemistry, recipe, process, exposure times, developing times etc are all incredibly important.

 

[pm]

There's got to be the price involved - maybe the latest generation tools are super expensive, and if you are one generation behind you can save some money. But even so, why not skip a generation... while I agree with everyone that the equipment is only part of it and you need a recipe and settings and everything else... and yields are important... all agreed. But you need all of these things for 20nm too. Just ask, why work on 20nm, when you could work on 14nm? Because 20nm is a stepping stone to 14nm? But I don't believe that answer because 14nm is certainly going to be finFET for GF and I can't imagine that working on 20nm planar FETs is going to help you step closer to making a 14nm finFET. But I can imagine that it's cheaper and less risky.

My guess is that it must be very expensive, and this investment can only be regained if you are shipping a very large volume of products. The first generation equipment is expensive, experimenting to get all of the important steps in the recipe correct in order to get the good things like high yields and excellent electrical characteristics are expensive, and the longer you are from the bleeding edge, the more these costs go down. So then it only makes sense to be on the leading edge if you are shipping a lot of product - and can thus share all of the expensive and difficult R&D costs over a large volume of prodcuts and fabs and if you aren't doing very high volume then you step back a bit from the edge, spend less on R&D and fab equipment and thus have lower fixed costs which then don't need as high volume/profit margins to pay off that fixed cost.

 

[Revolution 11]

It is not as simple as buying the tools for a given node, as silverpig points out. But I am sure that Intel develops much of its own hardware at this point (the engineers certainly have the expertise and resources to do so and it seems like an Intel philosophy to do everything themselves).

 

[pm]

I thought this was interesting:
http://www.electroiq.com/articles/s...-transistor.html?cmpid=ENLWaferNEWSJune122012

 

[extide]

I'm sure Intel either works closely with the people who make the tools, and/or build the toold from the ground up themselves. Their leading edge process, at least, is most definitely not off-the-shelf components.

 

[EltonL]

It isn't just tools. As another user pointed out, you have to have viable yields that won't leak an insane amount of electricity. It's one thing to be capable of making things in that node. It's another to make something that will consistently be produced from that same node.