Dedicated Fabs Reinvent Industry
Alexander E. Braun, Senior Editor -- Semiconductor International, 1/1/2000
 |
At a Glance | | |
|
| ||
| |||
This concept appears well on its way to the dustbin of history.
"The dedicated fab is a very major step toward the disintegration of the traditional vertical semiconductor company," said Roger Fisher, senior director of marketing for TSMC-USA (San Jose, Calif.), "It's major, because the dedicated foundry's emergence has enabled the rapid growth of the fabless semiconductor industry."
"The pure-play semiconductor foundries are setting the new value-added business model of IC design. Now IC design houses can turn to us for their silicon; they no longer require the considerable cost of having to own a fab," said Jim Ballingall, vice president of worldwide marketing for UMC Group (Sunnyvale, Calif.). "Fabless companies are focusing on their core strengths and turning to the pure-play semiconductor foundries for wafer development and assistance with their migration paths to smaller process geometries, such as deep 0.18 µm and below, and new processes like copper."
"Everybody talks about the convergence of the foundry roadmap with the SIA Roadmap. That's a little bit deceiving," said Kevin Meyer, Chartered Semiconductor's (Milpitas, Calif.) vice president of worldwide marketing. "People tend to focus on the logic process — 0.25, 0.18, 0.16 µm — but there are variants of it that are increasingly important, particularly as markets begin to segment."
| 1. Dedicated fabs are quickly moving to become the norm in semiconductor manufacturing. The Fabless Semiconductor Association estimates that by 2012 fully 50% of IC manufacturers will be fabless companies. (Source: Chartered Semiconductor Manufacturing) |
The push to fabless
Fabless semiconductor firms existed before the coming of dedicated fabs, but they led a somewhat marginal existence, depending on established semiconductor vendors to get their products built. Often a significant price had to be paid for this option, although not necessarily in dollars: They had to allow the semiconductor company building their product to second-source it. "The second blow to that fabless model," said Fisher, "was that they also had to allow the semiconductor company doing their manufacturing to use and incorporate their IP into their own products. In short, they were creating their own competition."
A major hurdle also was that the fabless company was never assured of capacity. Companies were, in effect, taking in "piece-work" from the fabless companies, to keep their fabs busy during slow periods. Whenever demand for their own products increased, their entire capability would be shifted to meeting it, leaving the fabless semiconductor company's needs a far second at the bottom of the priorities list.
|
| 2. Foundry market growth is currently proceeding at a compound annual rate of 19%. This rate is not expected to decline and may accelerate, depending on market demands. (Source: Dataquest/UMC) |
It soon became apparent there was a vacuum that could only be filled by dedicated fabs prepared to put out their customers' products and not their own (Fig. 2). "You can see the results of this innovation all around you," said Fisher, "in the technology and vitality that this capability to compete without the overhead of a fab has released. Right now, there are over 500 fabless companies, with a market value in excess of $20B, and this number is growing rapidly." According to TSMC, over the last decade fabless businesses have had a compound annual growth rate of over 36% — more than twice that for the overall semiconductor industry.
A different business model
The impact of the dedicated fab concept on the semiconductor industry is not unlike that which Henry Ford's assembly line had on the automobile industry. It has developed into a business model that is in reality a combination of two separate companies — a semiconductor company and a fab company. Together, they have been able to compete very effectively with established, vertically organized semiconductor companies, to the point where these traditional firms are beginning to migrate to the fabless model.
What do fabless companies have that their traditional counterparts do not? For one thing, they have access to a state-of-the-art wafer fabrication capacity generally operated more efficiently than those of most traditional semiconductor vendors. Since it can serve a broader number of customers and market spaces, this tends to average out some of the upturns and downturns in particular market segments.
Over the last decade, fabs have become better able to deliver the latest technologies to customers. "When TSMC started in business, in 1988," recalls Fisher, "our technology lagged two or three generations behind that of the leading semiconductor vendors. We started with 2.0 µm technology at a time when the Intels and TIs were at 0.8 µm. We steadily closed the gap, and early last year we introduced our 0.18 µm technology." Fisher added this puts them slightly ahead of major leaders such as IBM. "In fact, if you look deeply into actual physical structures, you'll find that, in metal interconnects, we have tighter geometries than they do. In metal 1, their pitch is 0.6 µm, and ours is 0.05 µm less than that. It's not major, but it's there and is an indication of advances made in developing our own technology." In the beginning, TSMC, like most of its competitors, bought or licensed its technology. Then it began building its own R&D effort. "Last year we spent $300M in process R&D, over and above what we spent in manufacturing capacity," said Fisher (Fig. 3).
|
| 3) Major dedicated fabs' capacity is increasing as the fabless model gains momentum. Market shares among the leading players also are shifting as competition becomes more pronounced. (Source: TSMC) |
This final development put to rest one major objection integrated, vertical manufacturers had for staying with their own manufacturing facilities and technology: that dedicated fabs may be efficient, but lack the leading technology needed for their particular product.
Leadership's price
Although being technology leaders is not an inexpensive proposition for the dedicated fabs, it is not a painful one either. "It isn't cheap to upgrade facilities," conceded Fisher. "However, since our growth is on the order of 30% a year, it's not much of a strain, either; we build new facilities every year. Last year, our total capital expenditures approached $2B."
When they build a new facility, fabs outfit it with first-grade equipment capable of handling the latest and, ideally, next generation of process technology. This allows them to remain about two generations ahead of the rest.
Customers & third-party suppliers
It is in the fabs' interest to progress as quickly as possible from an older process technology generation to a newer one. As Fisher puts it, "We're always working closely with third-party design library vendors to help them implement their products in our new processes as soon as possible. That way, when a customer is ready to go to a new process generation, the foundation design libraries and design environments are already there, and he doesn't have to put these in place himself."
When the fabless model began, the fab's primary deliverable to its customers, before they started buying wafers, was process parameters and design rules and models. These would then be used to create their transistor-level designs.
Today, another aspect of the industry's disintegration is that about 50% of the fabs' clients expect basic building blocks — NAND gates, NOR gates, memories, I/O — to be done and proven before starting a design. "They no longer have to do it or invest in it," said Fisher. "When a customer gets a product from a library vendor like Artisan, Avanti, or others, he knows it meets our design rules, has been prepared using our models, and that all the timings associated with it are consistent with our process." An entire new associated market has arisen from the production of these libraries.
More change
Even as this is being written, the fab business model is changing. Fabs' relationships with their customers are deepening, transmuting from a close partnering into what could almost be described as a symbiotic relationship. Fabs also are looking to work more closely with the third-party core vendors who provide microprocessor, signal-processing and other high-level function cores to make these even more readily available to customers.
Another major transformation already on the horizon, which will ripple throughout the semiconductor industry, is the transition from 200 to 300 mm wafers. "We estimate that a 300 mm wafer fab will cost in the neighborhood of $2.5B," said Fisher. "If you do the math on what that fab's output will be — assuming a 30,000-wafer-a-month fab — it is enough to produce end-product revenues somewhere in the $6-8B range. A single fab, supporting some $8B worth of product revenue! There aren't too many semiconductor companies with sufficient revenue to justify a single 200 mm fab."
|
| 4 . By 2012 — sooner by some projections — it is expected that outsourcing will be at the same level as captive fabs. (Source: Fabless Semiconductor Association) |
UMC recently announced it is building a 300 mm, 0.13 µm, copper-capable fab, at a cost of $3B. "We're putting all the bells and whistles in it," said Ballingall, adding that aside from the affordability aspect, it really no longer makes any sense for semiconductor manufacturers to build their own fabs. "Within a few years, even Motorola will farm out 50% of its needs," he said. "The model is validated, and projections indicating that foundries will produce about 50% of the world's ICs sometime in the next decade appear to be on target."
Some in the industry have argued that if fabs continue their rate of growth they, rather than their customers, will become the technology drivers, and that because of the skyrocketing costs of setting up new fabs, semiconductor technology progress may slow down. As long as there is more than one major dedicated fab, this is doubtful. At present, there are several highly competitive, extremely viable ones. Even if all vertical semiconductor manufacturers were to become fabless operations, (a not-too-unlikely possibility), the often vicious competition between some of the fabs appears to supply a built-in safety factor that precludes this.
The possibility of the dedicated fab turning into a semiconductor manufacturer does not appear to be a major concern, either. Ballingall indicated that although UMC is working more closely than ever with customers, EDA tool developers, design service companies, IP providers and package and test companies, it intends to remain a silicon manufacturer. "We're not going to become an ASIC company," he said. "There's a temptation for some foundries to bring in that capability, and some have done it; they have an internal ASIC division. We're convinced that's a step backward. You are basically attempting to re-integrate what is disintegrating. We can give the same level of service simply by working closely with design service companies, IP providers, EDA tool developers and the rest."
The fabless business model appears to benefit everyone in the industry, including equipment manufacturers. The semiconductor industry's disintegration and the move to fabless models has moved sales up rather than down. Like the IP model, all of these are catalysts that build and diversify, rather than reduce, the market.
|
| 5. The demands of the various sectors of the communications market are expected to fuel market growth, while driving the fabless manufacturing model. (Source: Dataquest) |
Fabs also are active in developing standards, by working with the various groups, for the interoperability of IP and tools to ensure these will work seamlessly.
According to Meyer, Chartered's strategies are all toward system-level capabilities. "We're not only developing 0.25, 0.18, 0.15 and even looking to 0.12 µm capabilities. We're also looking at what must go along with that baseline technology, such as analog for RF-CMOS," he said. "We're taking a systems perspective, not just racing to the next smaller geometry."
The shift from memory
Meyer is of the opinion that semiconductor manufacturing and process development is becoming less of a competitive advantage for those semiconductor companies that are still doing the actual IP development and implementing it in silicon. "Increasingly, the ability to implement systems-level technology and the IP necessary to implement that system solution is becoming the competitive advantage — whether it is a Broadcom with a cable modem design, or a Motorola with an embedded microprocessor for some advance communication capability such as a router on a chip. The competitive advantage is shifting to systems-level knowledge — not actual production capability," said Meyer.
|
| 6. Of the various markets driving the demand for semiconductor products, communications is expected to surpass the current leader, computers, with the rest maintaining their shares. (Source: WSTS/Dataquest) |
According to Meyer, the fab business model is changing. "We changed our R&D model to focus on systems-level technology. Almost two years ago, we refocused our sales and marketing energies on what we think will be the main semiconductor driver in the future: communications." He added that the company has moved away from the commodity business, which comprised a significant portion of its 1997 sales. "We had 25% of our sales in the memory sector. We were heavily committed to commodity memories. Memory is now less than 7% of our overall sales. Our communications industry segment has gone from about 26% in 1997 to over 43% today."
The fabless era is on its way, as the industry quickly — and successfully — turns to this business model. The Fabless Semiconductor Association (FSA) targets 2012 as the year when 50% of all ICs will be produced by foundries. The momentum in that direction is here, and makes economic sense. •
Chartered Semiconductor Manufacturing Ltd. www.charteredsemi.com
TSMC www.tsmc.com
UMC Group www.umcgroup.com
