The Near Impossibility of Making a Microchip

The microchip seen here

Somebody had to invent machinery is actually smaller than your thumbnail—as shown that could reliably and affordably lay on the last page. Amazingly, fabricators can down patterns with details as small (and must) mass-produce as 80-millionths of an inch. such chips while replicating every detail flawlessly.

BY DANIEL P. BURBANK but economical. Achieving this demanded innovations at every step of the manufacturing process. In the late 1950s Jack Kilby at Texas Instruments To visualize the Kilby/Noyce/Moore in Dallas, and Robert Noyce and Gordon Moore, at process, imagine a manic real estate Fairchild Semiconductor near San Jose, California, developer building an entire town at once. independently came up with ways to cram many First hundreds of bulldozers dig all the tiny transistors and resistors onto a small sliver of basements; then a fleet of cement trucks silicon. By adding microscopic wires to pours all the foundations; then an army of interconnect groups of adjacent components, they carpenters completes all the floors, then made the first integrated circuits, which are now the walls, then the upper stories and roofs; known informally as chips. The potential range of and finally all the streets are paved. In this applications was vast, limited only by the power of analogy the individual buildings 1950s imaginations to conceive of refrigerator-size correspond to individual transistors and computers shrunk to the size of a postage stamp. resistors, the streets to connections As with most technological breakthroughs, between them, and the blocks of buildings however, a brilliant idea was only the first step. and neighborhoods to functional circuits. The earliest chips of any complexity were logic In real life, of course, this method would circuits that went into the guidance systems of make no sense for building a town. But for intercontinental ballistic missiles. For this fabricating integrated circuits, which are application, mass production was not necessary, built from successive layers of silicon and cost was of secondary importance. For chips compounds and aluminum laid down with to become as ubiquitous as they are today, though, microscopic precision, it's the key to mass the technology had to be made not only feasible

COURTESY OF INTEL MUSUEM

When its first projection device needed 16 lenses, Perkin-Elmer decided to try a new approach production. Many things had to be an acid bath, leaving the mask pattern Different combinations of ultraviolet light through a mask; worked out along the way: isolating in aluminum topped with exposed chemicals are used on different dissolve the unhardened photoresist; and purifying the different photoresist. That hardened layers, but the basic procedure is the etch away some or all of the revealed substances, finding ways of photoresist is removed with another same: Deposit a solid film; cover it film; and then dissolve the remaining depositing them, making them adhere solvent, leaving just the network of with photoresist; imprint a pattern on photoresist. The earliest integrated to the silicon wafer and to one aluminum lines. the photoresist by shining intense circuits usually required about eight another, and dozens of additional masking steps. details. One of the most important Experience made the masking was ensuring that the minute and procedure work reasonably well, and complicated patterns for each layer, by the mid-1960s many chip makers as drawn up by the circuit designer, were building their own machines to were accurately reproduced on the do the job for about $10,000 apiece. surface of the chip. The trouble was that the patterns being reproduced on the chips had Chip fabricators achieve this details as small as 10 wavelengths of precision with the help of light- ultraviolet light. So the mask had to sensitive compounds and masks. For be clamped tightly to the wafer to example, let's say it's time to deposit prevent blurring at the edges of the the thin lines of aluminum that image. Despite elaborate procedures connect the components with one aimed at maintaining pristine "clean another. Fabricators start by laying rooms," tiny particles of dirt or dust down a solid layer of aluminum and would get stuck between the mask covering it with a solid layer of and the wafer. By the time eight steps photoresist, a plastic resin that had been completed, a high hardens when exposed to ultraviolet percentage of chips were defective. light. A glass mask with a negative Worse, particles often damaged the image of the interconnect pattern is mask, and even after they had been placed on top of the photoresist, removed, the defect remained to be image side down. Then a bright replicated on all later images. The ultraviolet light is trained upon the smaller and simpler a circuit was, the mask, "exposing" the photoresist like greater chance it had of getting film in a camera. Afterward a solvent through the process intact. The dissolves the unhardened photoresist, dismal economics of defects revealing the underlying portions of Pioneers of the Micralign include Peter Moller, fourth aluminum. These are removed with from left; John Bossung, fifth from left; Harold Hemstreet, third from right; Abe Offner, second from right; and Jere COURTESY SVG Buckley, far right, with their creation in 1977.

© FALL 1999·INVENTION & TECHNOLOGY 45 from contact printing placed a low equivalent of more than 300 million an engineer who worked on the to resolve those billion pixels all at ceiling on practical circuit pixels (picture elements) on a two- project for Perkin-Elmer, "but we once. He needed only to achieve complexity, even for well-funded inch wafer, which may contain hadn't really accomplished what high precision in a small area and U.S. Air Force programs. The Air hundreds or thousands of chips. the government wanted, a viable use it to scan the mask a little at a Force was looking hard for a way out (By comparison, a high-quality commercial product." Abe Offner, time, the way a photocopier scans a from under this ceiling. The 35mm camera lens resolves about who was one of Perkin-Elmer's sheet of paper. He accomplished this obvious solution was to separate 6 million pixels in a 28-by-35- most formidable optical designers, by cleverly combining two spherical the mask and wafer using a lens to millimeter format, roughly half the agrees: "It was at the very limits of mirrors in such a way that there was project an image of the mask onto area.) To achieve such high what could be manufactured at the a small ring in their image field in the wafer, as in a photographic resolution, the Microprojector time. No one thought that we could which the mirrors' aberrations darkroom enlarger. used 16 lens elements. All those make it in production." canceled each other out. Perkin-Elmer of Norwalk, lenses meant that it worked only Offner's design paired a Connecticut, one of the foremost in a narrow 200-angstrom sliver erkin-Elmer's experience "primary" concave spherical mirror optical-instrument companies in the of the ultraviolet spectrum (which with the Microprojector about 10 inches across with a country, had a long history of extends from 40 to 4,000 Pconvinced Harold smaller, convex "secondary" one building special-purpose optical angstroms). Outside that range, Hemstreet, general manager about 2 inches across. (See Fig. 1.) systems for scientific work and the dispersion would cause a faulty of the electro-optics division, that By arranging them along a common defense industry. So it was no image. This meant that most of lenses were the wrong approach. axis of curvature, he made a distor- surprise when Perkin-Elmer was the light from the system's 1,000- With the Micro-projector still in tion-free ring about a millimeter awarded an Air Force contract to watt mercury-vapor lamp had to be development, he called once again wide, with a diameter of about 5 build a lens-based projector in June thrown away. on Abe Offner, who decided to inches around the axis of 1967. The use of multiple lenses has investigate reflective systems—those symmetry. (See Fig. 2.) Light from Within a few years the company another consequence that made the using mirrors instead of lenses. above the mask would travel in a W had come up with its Microprojector hard to operate. Reflective systems image light of path, with the mask at the top of Microprojector, which met the Air Before exposure, the mask and the all wavelengths in exactly the same one arm, the wafer at the top of the Force specifications but was less wafer must be aligned with great way, solving the problems of other, the concave primary mirror than satisfactory for industrial use. precision by a technician. The alignment and wavelength range. forming a base for the W and the The main problem was that lenses Microprojector system was made But they create distortion of their convex secondary at the middle refract light of different wavelengths to transmit ultraviolet light, and own, which is known as aberration. peak. Moreover, the magnification at different angles—a basic property designers found it impossible to For that reason, designers had was inherently IX, which meant that of glass called dispersion. Designers get enough visible light through the previously avoided them. And to mask features would be imaged at of optical systems have learned to labyrinth of lenses for humans to make matters worse, by this time the the same size on the wafer, and the combine elements of different see. Fortunately, at about that time, resolution requirements had gotten system was telecentric, meaning that shapes, made from different types night-vision apparatus developed for even tougher. Now the system it would still work if the mask or of glass, to compensate for this Vietnam was declassified. Perkin- would have to resolve details of two wafer was moved slightly out of phenomenon. These methods involve Elmer installed an image-intensi- microns, or 80 millionths of an inch focus. This allowed much greater tradeoffs affecting resolution, fication system to let operators use (sometimes expressed as 250 line tolerance for mask or wafer surfaces image size, and the acceptable ultraviolet light for alignment pairs per millimeter). On the newly that were not perfectly flat. And range of wavelengths. Perkin- The night-vision fix worked, but it popular three-inch wafers, that Elmer's contract required its projec- was a stopgap that only increased worked out to more than a billion tor to resolve details as small as 2.5 the unit's cost. "We had pixels. microns, or 100 millionths of an accomplished the contract The task sounded daunting until inch, at an exact 1:1 scale. That's the requirements," says John Bossung, Offner realized that he didn't have

The parts of the roof mirror are held together without glue, by intermolecular forces alone

In fig. 4 a roof mirror replaces one of the flat fold mirrors, giving the images on mask and wafers the same orientation. Fig. 5 shows how mask and wafer are held in place in a scanning carriage that rotates by means of a flexure bearing (not shown).

Fig. 1 shows the basic idea behind Offner's breakthrough: Light passes through the mask, is reflected by the concave primary mirror twice and the convex secondary mirror once, and then exposes the photoresist on the wafer. Fig. down 2, looking along the axis of symmetry, shows the thin ring of near-perfect resolution and how it is used to scan the image on the mask. In fig. 3 a pair of flat fold mirrors has been added, allowing mask and wafer to be scanned in the same direction.

© FALL 1999·INVENTION & TECHNOLOGY 47 The company learned that to sell a tool, you have to develop expertise in its application. finally, since the system was entirely had drawn it up and simply pull the the ends of the W of light, to required by such a scheme would reflective, it would work at all mask and the wafer through at the submicron tolerances. A have been so complicated and wavelengths. same speed. That would avoid the photocopier or fax machine can expensive that the scanner could Now Offner could project a sliver need for any more mirrors. stretch or compress an image never have been commercially viable. of light—carefully aligned to Unfortunately, it would also require slightly without major correspond with a segment of this that the mask and the wafer be consequences, but even a tiny Markle and Buckley next distortion-free ring—and move the driven in precise, coordinated linear amount of scanning distortion considered what would happen if mask through the sliver until it had fashion at exactly equal velocities would have rendered a chip useless. they placed a pair of flat fold been completely covered. The but in opposite directions, across The mechanical and servo design mirrors into the light path, resulting scan, reflected through the reflecting the ends of the W system onto the silicon wafer, would outward. (See Fig. 3.) Doing so have excellent resolution could simplify the scanning process everywhere. His plan had the by allowing the mask and the wafer disadvantage of introducing to be held a fixed distance apart, moving parts into the system, but facing each other, and moved in the that could be coped with. Much same direction by a common more important was the advantage scanning platform, or carriage. This of replacing a complicated, unwieldy would have the advantage of system of 16 lenses with an requiring only one precision arrangement of two mirrors, a servomechanism instead of two. system so simple and elegant that it However, it would still invert the is found today in optical textbooks. image—that is, it would turn a "left-handed" image into a "right- Bossung built a bench-top proof- handed" one. This meant that a of-concept model using photographic company adopting the scanner film instead of photoresist to record would have to discard its existing the scanned image. The demon- inventory of masks. stration was enough to get another Even worse, it meant that if the $100,000 from the Air Force. With carriage was bumped perpendicular the design's feasibility to the scan direction, the image on demonstrated, Perkin-Elmer the mask would move one way but addressed the task of building a the image reproduced on the wafer sturdy, reliable machine that could would move the other. This would be produced at reasonable cost. In produce an unwanted jog in the May 1971 Hemstreet put Jere imaged line. To prevent such a Buckley, a mechanical designer, and In a "clean room" used for chip situation, the carriage would have to Dave Markle, an optical-systems fabrication, elaborate precautions be driven in exactly linear fashion engineer, in charge of the project. are taken to prevent contamination and in exactly the right direction, The most obvious approach was to by even the minutest particles. with any deviation causing configure the system just as Offner distortion. The mechanical COURTESY OF SEMI © FALL 1999·INVENTION & TECHNOLOGY 48 requirements would be simpler than axis of rotation turned out to be the scan could be performed in 10 to 12 mirror array mount was contrive a in Offner's original system but still simplest: the flexure bearing. A seconds. system in which there were three too complicated for commercial familiar application of flexures is the Other problems cropped up as the adjustment knobs. You turn one, production. Markle made a plastic cap with flip-up lid found on team began building prototypes. and it is strictly a focus adjustment. breakthrough by changing one of ketchup and shampoo bottles. The "Our first system had all the There is no cross coupling to the two flat fold mirrors into a roof Micralign's more sophisticated electronics in the desktop—a bad anything else. You turn another, and mirror. (See Fig. 4.) A roof mirror high-performance flexure bearing decision because the electronics it exclusively adjusts distortion in the consists of two flat fold mirrors uses thin, flexible stainless steel kept growing and the desk didn't," direction of the scan. You turn the joined at a right angle. Large ver- leaf springs arranged to allow says Markle. Another challenge third knob, and it adjusts distortion sions of roof mirrors are often found limited rotation around one axis was giving the operator a way to perpendicular or across the in clothing stores, barbershops, and only. "We did look at other bearing align the mask image with previously direction of the scan and affects hair salons, where they allow types, a rotary air bearing in patterned wafer features. "This was nothing else. With that system you customers to see themselves from any particular, but chose the flexure solved by putting a dielectric coat- could lay out a cookbook angle. Inserting a roof mirror bearing for its simplicity, accuracy, ing on one of the fold mirrors for procedure for the service engineer, diagonally added an extra and reliability," Buckley says. "You fine alignment and on the secondary where in a matter of minutes he reflection to the light path, yielding could quite literally throw a handful [spherical] mirror for coarse could make all these adjustments an image that was not inverted. of sand into the bearing with no alignment." This coating reflected with a minimum of confusion." Now customers could use up their loss of performance." the ultraviolet wavelengths onto the Through all the technical ups stock of masks that had been made Although the team completed the wafer for exposure but transmitted the and downs, Harold Hemstreet for contact printing. More impor- basic design of the Micralign (as it yellow and green wavelengths provided the vision and management tant, the three-reflection design was dubbed) by November 1971, a through the mirror to the operator's skill to keep the project on track. eliminated the requirement of a number of hurdles lay between that viewing scope for alignment. Buckley says, "I remember Harold precisely linear scanning movement. design and a practical Fabrication of the roof mirror sitting in his office and saying, As long as the mask and the wafer manufacturing tool. As Markle also required a major development 'Someday we're going to sell a moved in tandem with each other, recalls, "One challenge was finding effort. Because of the extreme thousand of these machines,' and the direction of the scan did not a way to efficiently and uniformly precision required of the right everybody thought he was totally have to be held constant. illuminate a curved slit 1 millimeter angles, it could not be built the bananas." The development staff In fact, Markle realized, the scan wide and about 80 millimeters long." conventional way, by gluing two mir- was much less sanguine than Hem- would not have to be linear at all. Simply projecting an ordinary lamp rors together with optical cement. street, especially as pressure This was an unexpected benefit. through a slit-shaped hole would Instead the bonding surfaces had to mounted with the product launch Incorporating a linear translation not do, because it would waste too be made so smooth and clean that approaching in the summer of into the system would have been much light and thus require too inter-molecular forces alone would 1973. Late one night Peter Moller, a difficult, requiring an expensive and much time per scan. Instead, says hold the components together, essen- marketing executive with Perkin- high-maintenance carriage supported Markle, "the answer turned out to tially making the two pieces Elmer, was in the laboratory with a on an air bearing and floated over a be a curved capillary mercury arc become one. small group of engineers. As granite slab. But the new design lamp which Ray Paquette from ARC Jere Buckley designed the Buckley recalls, "The machine allowed the linear scan to be made and tested for us in about two alignment system, of which he is wasn't behaving properly, and it replaced by rotation around an axis, hours after I phoned him one day." still proud: "If you've ever had wasn't really all that clear that much more reliable and less Offner designed a light source much exposure to optical systems, we were on solid ground with this complicated. (See Fig. 5.) around this special lamp to you may have noticed that they thing. Peter Moller said, I’lll give collimate the ultraviolet sliver— tend to be full of adjustments, you a trip to Bermuda when we s often happens, the best that is, make all the light waves many of which are cross-coupled. sell the hundredth machine or a A choice of bearing for the parallel. Using this source, a typical What I was able to do on that fold cup of coffee now.' We all took

the cup of coffee. And I wasn't potential customer similarly production-floor wear and tear. time the step is repeated half a even a coffee drinker!" unimpressed. "Operators would put their feet up dozen times, the probability of The launch had its bumpy As it turned out, the trouble did on the aligner during exposure," getting a good chip will be much passages, as Moller remembers. not lie with the Micralign itself. Bossung says. Process engineers greater. "Texas Instruments came in, and "By the time I got home," Moller then had to figure out why the In 1975 a report by a consulting we ran a lot of wafers for them. continues, "Raytheon had decided images looked so much worse firm summarized the savings. For They looked at them through that we probably didn't know than they had in the small integrated cir cuits, such as microscopes and so on, and they anything about photoresist. They demonstration. But once the the SN7400 TTL logic series, seemed to be fairly impressed with were basically right about that." problems were ironed out, cost which could fit on a silicon chip 35 what we had done. . . . We then Perkin-Elmer was not a chip savings were dramatic. A Texas by 48 thousandths of an inch, launched and went to the West fabricator, and its lack of Instruments manager said the units yields improved from 75 percent Coast with what we called our experience in processing the demon- paid for themselves in 10 months. with contact printing to 90 percent 'golden wafers' to show the stration wafers had nullified much Perkin-Elmer turned out Micraligns with the Micralign. A wafer three industry." At Raytheon "we gave of the precision engineering that as fast as it could, but new inches in diameter could hold them the wafer and they looked at went into the Micralign. Raytheon customers had to wait as long as a 4,000 copies of this chip, so the it under the microscope. And the sent an experienced fabricator to year. Micralign printed 600 more good head of production at the time let show Perkin-Elmer how to Early marketing materials chips per wafer. Results were even out a loud bellow, and he said, process photoresist. Through this emphasized the enormous more dramatic for larger chips with 'This is s—!' We were absolutely experience Moller and the company improvement in mask life. "Contact higher profit margins. In the mid- stunned. We left there. It was time learned that in order to sell a tool, printing was godawful," Bossung 1970s logic circuitry for a four- for lunch, and we sat around lunch you have to develop expertise in says. "The emulsion [on the masks] function calculator could fit on a and decided that he probably didn't its application. would lift off. . . . After ten uses, the chip 140 thousandths of an inch know what he was talking about. In spite of these initial setbacks, masks were useless." Buckley agrees: square. Contact printing yielded a We went into National Semi- manufacturers came to appreciate "Places like TI were buying dismal 30 percent usable chips, conductor, and we dealt with a the Micralign's superior masks, literally by the truckload, while the Micralign yielded 65 fellow there who was more mask- performance. The first one was using them six to ten times, then percent. oriented. He was mostly interested sold to Texas Instruments in 1974 putting them in the landfill." By in whether [the image magnifi- for $98,000. Intel and Raytheon contrast, the Micralign offered mask cation] was 1:1. He looked at the were also among the early pur- lifetimes of at least 100,000 ltimately, the Micralign wafer and the mask, and, boy, they chasers. The Micralign would exposures. scanner made the personal were exactly 1:1. So we thought prove to be a high-tech cash Gradually, though, the computer possible by that was great. That was the first machine for its users, but fitting it semiconductor companies began to Uallowing the necessary day, a 50-50 deal." into existing production processes realize that the greatest savings microprocessors to be manufactured The next day was less was much more complicated than from projection alignment came not cheaply enough for a reasonably successful. A wafer they brought to plugging a new component into a from reduced mask costs but from priced unit. In June 1978 Intel Fairchild "came back with electron stereo system. Fabrication tech- improvements in yield. No longer introduced the 8086 chip. A year microscope pictures of godawful- nicians were accustomed to did particles transfer from wafer to later came its sister chip, the 8088, looking edges. . . . They said this rugged, low-tech contact-printing mask, to be replicated with each which was used in the first IBM PC. just wasn't hacking it." Another equipment and had to be taught to subsequent use. And even if the These chips were about one-fifth of visit that afternoon left a fourth respect a unit that was much less yield for a single step is increased by an inch square and contained some forgiving of vibration and only a few percentage points, by the 29,000 transistors each. With © FALL 1999·INVENTION & TECHNOLOGY 50 contact printing, only 20 percent of however, it began to encounter Today's chip-fabrication tools can Present-day projection aligners these chips were salable. With the serious competition from firms in combine laser light sources and don't look very much like the Micralign, the yield shot up to 60 the United States, the Netherlands, sophisticated lens systems with original Micralign, but their percent. and Japan. After some rough times scanning technology and "step-and- performance continues to improve Perkin-Elmer reached total market Perkin-Elmer sold its Micro- repeat" methods to print chips one relentlessly on the upward dominance in the early 1980s, when lithography Division to Silicon at a time across a wafer that may trajectory set by the launch of that more than 2,000Micraligns were in Valley Group, a manufacturer of be 4 by 6 inches or larger. machine a quarter-century ago. use worldwide. Later in the decade, wafer-processing equipment.

Daniel P. Burbank is a photolithography engineer with Seagate Technology, a maker of computer disk drives. He lives in Minneapolis.