Tailoring 300m fabs for Beijing

Constructing and operating a 300mm wafer fab in Beijing presents some unique challenges to overcome. As perennial sand storms sweep across the Mongolian deserts to Beijing at least twice a year, special design features were implemented to ensure that fab buildings would not be infiltrated by dust and sand. The three levels of protection - sand trip louvers, auto roll filters, and air filters and washers - effectively work to keep out sand particles. These measures help maintain Class 100 and Class 1000 cleanroom air standards in the photolithography bay area and elsewhere in the fab, respectively.

In addition, the design of the 300mm facility had to take into account Beijing's limited water resources and the high cost of water and wastewater treatment. This prompted additional features to the building design to maximize water conservation. Rainwater collection tanks, with a total capacity of 2600m³, collect and store rainwater for daily operational usage. Additionally, an intricate network of water collection and reclamation systems helps to re-use cleanroom and industrial water. By the end of 2006, the reclaim ratio of ultrapure water used in the facility's cleanrooms for wafer manufacturing is expected to reach 80%.

Beijing's winter temperature also provides opportunities for saving energy. During the cold season, from November to March, cooling towers take advantage of the outside air to create a "free" cooling system for chilling operations, enabling the shutdown of all chillers normally used to supply the process-cooling water system, dry coils, and make-up air units. Heat also is reclaimed from steam and reused for operational needs. Also, the Beijing site's buildings were designed and constructed to insulate against temperatures as low as -15°C.

In the initial stage of 300mm manufacturing in China, it is more advantageous for foundries to follow an IDM model by offering few products in large volumes. However, as 300mm manufacturing systems mature, foundries in China should become more flexible in processing large-diameter wafers, and that will lead to greater diversity in technology offered in 300mm fabs. Cost pressures and the need for greater capacity will influence a growing number of advanced IC players to consider manufacturing investments in China. Thus, China's IC industry will move even closer to the forefront of advanced technology and manufacturing capability by decade's end.

Semiconductor devices have long been used in electronics. The first solid-state rectifiers were developed in the late nineteenth century. The galena crystal detector invented in 1907, was widely used to construct crystal radio sets. By 1947, the physics of semiconductors was sufficiently understood to allow Bardeen and Brattain to construct the first bipolar junction transistor. In 1959, Kilby constructed the first integrated circuit, ushering in the era of modern semiconductor manufacture.

The impediments to manufacturing large quantities of reliable semiconductor devices were essentially technological, not scientific. The need for extraordinarily pure materials and precise dimensional control prevented early transistors and integrated circuits from reaching their full potential. The first devices were little more than laboratory curiosities. An entire new technology was required to mass produce them, and this technology is still rapidly evolving.

This chapter provides a brief overview of the process technologies currently used to manufacture integrated circuits.

Integrated circuits are usually fabricated from silicon, a very common and widely distributed element. The mineral quartz consists entirely of silicon dioxide, also known as silica. Ordinary sand is chiefly composed of tiny grains of quartz and is therefore also mostly silica.

Despite the abundance of its compounds, elemental silicon does not occur naturally. The element can be artificially produced by heating silica and carbon in an electric furnace. The carbon unites with the oxygen contained in the silica, leaving more-or-less pure molten silicon. As this cools, numerous minute crystals form and grow together into a fine-grained gray solid. This form of silicon is said to be polycrystalline because it contains a multitude of crystals. Impurities and a disordered crystal structure make this metallurgical-grade polysilicon unsuited for semiconductor manufacture.

Metallurgical-grade silicon can be further refined to produce an extremely pure semiconductor-grade material. Purification begins with the conversion of the crude silicon into a volatile compound, usually trichlorosilane. After repeated distillation, the extremely pure trichlorosilane is reduced to elemental silicon using hydrogen gas. The final product is exceptionally pure, but still polycrystalline. Practical integrated circuits can only be fabricated from single-crystal material, so the next step consists of growing a suitable crystal.

The principles of crystal growing are both simple and familiar. Suppose a few crystals of sugar are added to a saturated solution that subsequently evaporates. The sugar crystals serve as seeds for the deposition of additional sugar molecules. Eventually the crystals grow to be very large. Crystal growth would occur even in the absence of a seed, but the product would consist of a welter of small intergrown crystals. The use of a seed allows the growth of larger, more perfect crystals by suppressing undesired nucleation sites.

In principle, silicon crystals can be grown in much the same manner as sugar crystals. In practice, no suitable solvent exists for silicon, and the crystals must be grown from the molten element at temperatures in excess of 1400°C. The resulting crystals are at least a meter in length and ten centimeters in diameter, and they must have a nearly perfect crystal structure if they are to be useful to the semiconductor industry. These requirements make the process technically challenging.

The usual method for growing semiconductor-grade silicon crystals is called the Czochralski process. This process, illustrated in Figure 2.1, uses a silica crucible charged with pieces of semi-grade polycrystalline silicon. An electric furnace raises the temperature of the crucible until all of the silicon melts. The temperature is then reduced slightly and a small seed crystal is lowered into the crucible. Controlled cooling of the melt causes layers of silicon atoms to deposit upon the seed crystal. The rod holding the seed slowly rises so that only the lower portion of the growing crystal remains in contact with the molten silicon. In this manner, a large silicon crystal can be pulled centimeter-by-centimeter from the melt. The shaft holding the crystal rotates slowly to ensure uniform growth. The high surface tension of molten silicon distorts the crystal into a cylindrical rod rather than the expected faceted prism.

The Czochralski process requires careful control to provide crystals of the desired purity and dimensions. Automated systems regulate the temperature of the melt and the rate of crystal growth. A small amount of doped polysilicon added to the melt sets the doping concentration in the crystal. In addition to the deliberately introduced impurities, oxygen from the silica crucible and carbon from the heating elements dissolve in the molten silicon and become incorporated into the growing crystal. These impurities subtly influence the electrical properties of the resulting silicon. Once the crystal has reached its final dimensions, it is lifted from the melt and is allowed to slowly cool to room temperature. The resulting cylinder of monocrystalline silicon is called an ingot.

Since integrated circuits are formed upon the surface of a silicon crystal and penetrate this surface to no great depth, the ingot is customarily sliced into numerous thin circular sections called wafers. Each wafer yields hundreds or even thousands of integrated circuits. The larger the wafer, the more integrated circuits it holds and the greater the resulting economies of scale. Most modern processes employ either 150mm (6") or 200mm (8") wafers. A typical ingot measures between one and two meters in length and can provide hundreds of wafers.

The manufacture of wafers consists of a series of mechanical processes. The two tapered ends of the ingot are sliced off and discarded. The remainder is then ground into a cylinder, the diameter of which determines the size of the resulting wafers. No visible indication of crystal orientation remains after grinding. The crystal orientation is experimentally determined and a flat stripe is ground along one side of the ingot. Each wafer cut from it will retain a facet, or flat, which unambiguously identifies its crystal orientation.

After grinding the flat, the manufacturer cuts the ingot into individual wafers using a diamond-tipped saw. In the process, about one-third of the precious silicon crystal is reduced to worthless dust. The surfaces of the resulting wafers bear scratches and pockmarks caused by the sawing process. Since the tiny dimensions of integrated circuits require extremely smooth surfaces, one side of each wafer must be polished. This process begins with mechanical abrasives and finishes with chemical milling. The resulting mirror-bright surface displays the dark gray color and characteristic near-metallic luster of silicon.

The electronics industry has traditionally relied exclusively on electrons transmitted through metal wires to process signals on-chip, and for most off-chip applications as well. The primary exception has been in the communications industry, where lightwaves (i.e. photons) are often used to transmit data through fiber optic cables.

All that is rapidly changing. Not only are photons being used in inventive new devices - so much so that the old term "optoelectronics" no longer seems adequate - but they may soon be used to communicate signals on the chip itself, through tiny optical waveguides.

Optoelectronics, of course, has been around for decades, consisting primarily of relatively simple III-V-based devices, such as photodetectors and solid-state lasers.

Although some integrated optoelectronic devices have been produced - where digital signal processing and optics coexist on the same chip - these have not been overly successful because it's difficult to marry III-Vs and silicon, and III-V digital processors are comparatively expensive to produce.

What's new are devices that manipulate the photons while they're still in a light format. These include microelectromechanical systems (MEMS) that switch light with mirrors, and planar lightwave circuits that multiplex/demultiplex light signals through optics.

According to market researcher Cahners In-Stat, leading applications for MEMS have traditionally included pressure sensors, accelerometers, inkjet printer nozzles, and read/write heads for hard disk drives. But what's leading the charge now? MEMS-based photonic switching! In 1999, there was no such thing, but by 2006 MEMS switches are expected to become the first MEMS device to surpass the $1B mark.

Presenting even more potential are planar lightwave circuits, which are based on optical waveguides created using manufacturing processes similar to those used to produce semiconductor devices (i.e. fine-line lithography, etching, doping, thin-film deposition, etc.).

The best example of this is DWDM (dense wavelength division multiplexing) devices, but planar technology can also be used to create variable optical attenuators, optical switches, and possibly complete optical add/drop multiplexers - all on one chip. What's especially interesting is that these circuits can be active in the sense that signals can be applied through electrodes to change their optical properties - a far cry from the optoelectronics of yesterday.

Photons are also poised to make their way onto the chip. At first, it's likely that they will be used primarily for clocking on high-speed microprocessors. Here, the light will be generated off the chip and then "piped" on and distributed throughout the chip with optical waveguides made of silicon and silicon dioxide.

Beyond that, it's possible that photons will be used for all on-chip communication - what some in the industry have referred to as "a million points of light." If that sounds implausible, given silicon's historical inability to generate light, consider some new research out of the University of Surrey, in which researchers demonstrated luminescence from silicon at room temperatures with standard manufacturing techniques (see Semiconductor International, May 2001, p. 36).

Even further down the road, it's possible that optical computing - long touted but as yet unrealized - might well become a reality.

So what does all this mean? For one, you can expect to see more coverage of these kinds of topics within the pages of Semiconductor International. Expect a bunch of new equipment introductions aimed at this new and rapidly growing market. And, most important of all, expect your colleagues to be boning up on optics technology. If you want to join them, keep on reading SI, of course. We've seen the light!

During the past 10 years, the industry has been moving toward distributed computing. At first, mission-critical corporate applications stayed primarily on mainframes while fast-turnaround applications made their mark on departmental systems. Lately however, mission-critical, bet-your-business applications are being written and rewritten as distributed applications. While these business applications may continue to connect to mainframe data, they are now designed to work across an organization. But companies are now discovering that these critical applications are more difficult to manage in their distributed form than in their centralized form. Because the transition to distributed computing has not always been well-planned, organizations may begin to regress to centralized computing.

Why are distributed applications presenting so many challenges? Part of the problem is that organizations typically focus on how to design and develop applications. They do not plan for application deployment and management, which are difficult parts of the process but vital to the success of distributed computing. Successful deployment and management of distributed applications depends on a solid strategy that includes the following components: software distribution, performance management, troubleshooting, and code management.

Software Distribution. If an application is to be distributed, an organization must understand how code will physically get to each user or group that needs it. Is there staff that can manage this process? Are systems in place to manage the load? Does software exist that will allow applications to be distributed from a central site? Will the new code conflict with code that is already installed? These issues may seem insignificant when developers are hard at work writing code, but they can kill a project if you don't address them early.

Performance Management. Most complex systems begin as pilot or proof-of-concept projects. While such initial projects may demonstrate or highlight the capabilities of developers to write code, they often do not deal with scalability issues. Distributed systems applications must be architected so that they will perform well as the code base and the number of users grow.

Troubleshooting. I do not mean debugging here. Complex problems can occur after debugging has been completed and an application is deployed. More and more organizations are creating applications that are intended to work in tandem with other applications. This involves using a messaging or remote procedure call mechanism to move information among applications - not a simple task. And, as the importance of the Internet in creating interoperability among applications grows, this problem will become more complex. How can you track what is happening when one part of an application interacts with another? My guess is that most development organizations have no plan in place to deal with the inevitable problems that can and probably will arise in such an environment.

Code Management. As development organizations begin to use component libraries to extend the life of systems and expand functionality, problems will surely arise. Developers will have to be able to trace dependencies among libraries. They will also have to understand the performance implications of these changes, which is especially important if code is distributed across several different systems. Issues such as the synchronization of code across multiple systems will also be a challenge.

ЧТО И КАК ПИСАЛИ О МОДЕ В ЖУРНАЛАХ НАЧАЛА XX ВЕКА Первый номер журнала «Аполлон» за 1909 г. начинался, по сути, с программного заявления редакции журнала...

Конфликты в семейной жизни. Как это изменить? Редкий брак и взаимоотношения существуют без конфликтов и напряженности. Через это проходят все...

Что будет с Землей, если ось ее сместится на 6666 км? Что будет с Землей? - задался я вопросом...

ЧТО ПРОИСХОДИТ, КОГДА МЫ ССОРИМСЯ Не понимая различий, существующих между мужчинами и женщинами, очень легко довести дело до ссоры...

Не нашли то, что искали? Воспользуйтесь поиском гугл на сайте: