Feature

Who's going to own their own fab?

In the past, CEOs used to brag that "Real men own fabs." With costs now pushing well into the billions of dollars, which companies can afford to make their own chips?

By Staff -- Movers & Shakers, 6/21/2007

Editor's note: This article comes from a roundtable discussion in which the top editors of Electronic News, Electronic Business and EDN sat down with John Daane, president and CEO of Altera; Roy Vallee, chairman and CEO of Avnet; Bernard Meyerson, chief technologist at IBM; Lothar Maier, CEO of Linear Technology; Walden Rhines, chairman and CEO of Mentor Graphics; James Truchard, president and CEO of National Instruments; and Brian Halla, chairman and CEO of National Semiconductor. For additional excerpts from the conversation, please see "The Brightest Ideas in electronics."

Q: Is there a future for integrated device manufacturers (IDMs)?

Meyerson: I don't think we're an IDM anymore. We built the world's only vertical ecosystem. We have parts in the vertical, but while we have a vertical stack, from shoveling sand into shipping the system, we allow people to hook into our vertical at two places and bypass everything we own. If they're better than what we've got inside, we'll use theirs. People used to think at IBM we had silo walls that were four-inch-thick steel. Those days are long dead. We've overcome that in our model. We couldn't survive. What happens in other businesses? Is there such a thing as pre-competitive collaboration or collaboration?

Daane: There will be always IDMs. If you look at Intel, Samsung, the analog companies, there are going to be many.

Rhines: Memory companies.

Daane: That's right. There are lots of examples of companies that will stay with manufacturing because it's core to their business. To say that the world will go to one manufacturer is wrong.

Rhines: Where we're talking about this is in digital CMOS, basically. In most other areas—in analog, in memory—most of the companies that are engaged as IDMs continue to find sufficient differentiation to continue to support a process development investment and the differentiations associated with having unique processes.

Q: Whatever happened to the concept of a microfab, where you didn't have to do 30,000 wafers per month?

Daane: It was replaced by the fabless industry. You can go to a foundry that has a much better economy of scale and better yields. And yields are important. The real cost for a microfab isn't the equipment. It's the ability to invest in R&D and for yields, tools and technology.

Vallee: In our business, there has been a lot of consolidation, arguably because our model is more scalable than other parts in the supply chain. We get asked all the time whether there's room for smaller distributors. Yes, but small is not a differentiation. There has to be some other form of differentiation. So if a microfab can be competitive with a giant fab running the same process as TSMC, that's a non-starter. But if it's going to run a different process technology or do something specialized, then there is potential.

Maier: That's the competitive advantage that a lot of analog companies have. They don't run on CMOS. The processes they have in their toolbox is their competitive advantage. We don't see that model changing. We're going to be a manufacturer going forward. That's core to our business model, and it's going to be a differentiator.

Vallee: Don't confuse size with scale.

Q: But if you don't do a microfab, who goes into production in digital chips other than a memory manufacturer or an Intel or Qualcomm? If you have to split a wafer, can you ever get enough yield?

Meyerson: Both Intel and AMD have announced the equivalent what is effectively a common socket? You have one standard socket, where you crank huge volumes of processors that become the main driver of the IT system. Then you have an ancillary socket that can be the accelerator. But you can start off with an FPGA that can be the accelerator, then if you get enough volume you can drop in a custom ASIC. But people are already finding schemes to address that problem.

Truchard: The real story is in software as we move forward. Maybe some companies will have five chips, and they have so much function that it's the software they add that makes the difference. I really think software will play a major role as we go forward. I admit it's not always possible to make them low power, but the software allows us to do a tremendous amount. Whatever I/O you want can be put on the chip.

Meyerson: You open a good point. In the past, the average x86 server ran about 5 to 6 percent utilization. There is an argument in the industry that the sales of x86 chips will decline for quite some time, particularly in server class, because the servers are running 5 to 8 percent utilization. If you virtualize the server class and take it up to 50 percent utilization, you've got more than 5x the output without putting in a single piece of hardware. And you save 500 percent in your power bill, not including the air conditioning. Software can have more profound impact than two or three generations of hardware.