Proof #4 COLOR 8/5/08 @ 5:15 pm BP

beyond silicon’s Elemental

logicIn the quest for speed, key parts of micro- processors may soon be made of arsenide or other “III-V” By Peide D. Ye

he first general-purpose lithography, billions of them are routinely microprocessor, the Intel 8080, constructed en masse on the surface of a sili- released in 1974, could execute con wafer. about half a million instructions As these got smaller over the t per second. At the time, that years, more could fit on a chip without rais- seemed pretty zippy. ing its overall cost. They also gained the abil- Today the 8080’s most advanced descen- ity to turn on and off at increasingly rapid dant operates 100 000 times as fast. This phe- rates, allowing microprocessors to hum along nomenal progress is a direct result of the semi- at ever-higher speeds. all artwork: bryan christie design bryan artwork: christie all conductor industry’s ability to reduce the size But shrinking much beyond their of a microprocessor’s fundamental building current size—a few tens of nanometers—will be blocks—its many metal-oxide- a herculean challenge. Indeed, at some point in field-effect transistors (MOSFETs), which act the next several years, it may become impossi- as tiny switches. Through the magic of photo­ ble to make them more minuscule, for reasons

www.spectrum.ieee.org september 2008 • IEEE Spectrum • INT 39 TOMIC-LAYER DEPOSITION provides one means for coating a semiconductor wafer with a high-k aluminum oxide insulator. The Abenefit of this technique is that it offers atomic-scale control of the coating thickness without requiring elaborate equipment.

1. Apply the gas trimethyl 2. Apply water vapor, which aluminum, which reacts with the reacts with the adhered hydroxyl groups attached to the trimethyl aluminum, forming a surface of the wafer, creating thin coating of aluminum oxide. a one-molecule-thick veneer.

The repeated Water vapor application of trimethyl Trimethyl aluminum Methane by-product aluminum and water vapor in alternating steps serves to build up the insulator into a many-atom-thick layer (shown schematically here as just a thin vertical slice).

of fundamental physics rather along with other researchers it as a III-V semiconductor. There than nuts-and-bolts engineering. in industry and academia, have are more than a dozen such com- So people like me have been look- recently made some advances that pounds, including ing at other ways to boost their might soon allow transistors built and , but gal- speed. In particular, we’ve been with or a related lium arsenide is the most com- laboring to build them using com- compound to be used for large- mon example and therefore the pound semiconductors like gal- scale digital ICs. That capability best studied. It currently accounts lium arsenide, which would allow would go a long way toward bring- for about 2 percent of the semi­ such transistors to switch on and ing us microprocessors that can conductor market. off much faster than their silicon blaze along at triple or even qua- Gallium arsenide devices cousins can. druple the speed of today’s best. cost a lot more than ones built of This strategy is by no means Achieving that goal will no doubt ­silicon—the raw materials are new. Practically ever since the sili- require other improvements in about 10 times as expensive—but con MOSFET was invented in 1960, semiconductor technology to take they serve well for certain special- engineers have been attempting to place in parallel, but gallium arse- ized applications, including high- come up with a gallium arsenide nide or something close to it could efficiency solar cells, diodes, version suitable for large-scale be key. No wonder some of us have and one very special kind of field- integrated circuits. No one has yet been unwilling to give up on this effect : the high-electron- succeeded. Those repeated fail- remarkable material. mobility transistor, or HEMT, ures have led to the oldest joke in which is used in cellphones, com- Silicon Valley: gallium arsenide— allium arsenide’s munication systems, and radars, it’s the technology of the future, two main compo- among other things. and it always will be. nents come from HEMTs are remarkable devices But that perennial skepticism the third (gallium) because they overcome a funda- may be about to vanish. My col- G and fifth (arsenic) mental problem of solid-state phys- leagues and I at Purdue Uni­ columns in the right-hand portion ics. Semiconductors, as their name versity’s Birck Nanotechnology of the periodic table of elements, implies, normally don’t conduct Center, in West Lafayette, Ind., which is why cognoscenti refer to electricity all that well. Usually,

40 INT • iEEE Spectrum • september 2008 www.spectrum.ieee.org www.spectrum.ieee.org september 2008 • IEEE Spectrum • INT 41 they must be doped with other at the interface fails to bond with with higher dielectric constants, kinds of atoms to become electri- the adjacent silicon dioxide, leav- which could be made physically cally conductive. But those impu- ing what’s called a dangling bond. thicker without compromising rities tend to interfere with the These defects disrupt the flow of the electrical functioning of the movement of electrons through electrons in the channel, but they transistor. Eventually, suitable the semiconductor’s crystal lat- are rare enough that they don’t compounds were found. Intel, for tice, limiting the conductivity that materially degrade the perfor- instance, is now using a hafnium- can be obtained. mance of a transistor. based gate insulator on some of its In HEMTs, electrons are intro- Gallium arsenide is a different most advanced microprocessors duced into a III-V semiconductor story. When it oxidizes, it forms a [see “The High-k Solution,” IEEE

not by doping but by placing the complex mixture of Ga2O3, As2O3, Spectrum, October 2007].

material in contact with another and As2O5. Beginning in the 1960s, To control the thickness of III-V compound that is doped. some researchers tried using these “high-k” dielectrics (a des- In essence, electrons fall a short these native oxides for a gate ignation that refers to the symbol distance into the undoped mate- insulator, but that tactic proved used for a material’s dielectric rial, allowing a thin layer of it— worse than useless because the constant, the Greek letter kappa), the channel—to conduct electric- native oxides create all kinds of manufacturers apply them to the ity extremely well whenever the defects at the interface with the silicon substrate using a tech- transistor is switched on. gallium arsenide, which destroy nique called atomic-layer depo- HEMTs can be used singly or the electrical conductivity of the sition. It’s quite ingenious, really. in integrated circuits with, say, adjacent channel. Clearly, a bet- The trick is that you employ a 100 or even 1000 of them clustered ter material needed to be found if chemical carrier molecule that together, but they can’t yet work there was to be any hope of mak- sticks to the target surface but for microprocessors. The problem ing gallium arsenide MOSFETs not to itself. Such a chemical thus is that too many of the electrons for digital ICs. deposits a one-molecule-thick that are supposed to flow through veneer. A second treatment with the channel from the transis- esearchers con- another carrier molecule removes tor’s source electrode to its drain tinued the decades- the first carrier, leaving a two- instead seep out the controlling long quest, testing atom-thick layer of the desired input electrode—the gate—creat- silicon dioxide, sil- material behind. Repeated appli- ing heat. With millions of leaky r icon nitride, silicon cation of the two carrier gases, one transistors crowded together oxynitride, and aluminum oxide, alternating with the other, allows on the same chip, things would among other candidates. They also chip makers to deposit various quickly get hot enough to melt. tried adding a third material, such high-k gate insulators on silicon In a silicon MOSFET, a layer of as sulfur, silicon, or , with atomic-level precision. intervening insulation (traditionally between the substrate and the insu- In 2001, Glen Wilk, then my col- silicon dioxide) prevents electrons lator to negate the pernicious effects league at Bell Labs, and I decided from slipping out of the channel of dangling bonds. Yet the results to try to put a high-k gate insula- into the gate. In a HEMT, the chan- always proved disappointing, and tor—in this case, aluminum oxide

nel is separated from the gate by a by the early 1990s most investi- (Al2O3)—on top of gallium arse- semiconductor, which, as you might gators had simply given up. Two nide using atomic-layer deposi- expect, is somewhat conductive. notable exceptions were Minghwei tion, which was all the rage at the What’s needed here, of course, is an Hong and Matthias Passlack at Bell

insulator, but for decades there have Labs, who had developed a way of Al2O3 gate insulator p-type In0.65Ga0.35As been no good gate insulators avail- depositing a combination gallium able for gallium arsenide. From time oxide–gadolinium oxide insulator Gate to time over the years, researchers on a III‑V substrate, a strategy that Source Drain seem to uncover a promising mate- Passlack later refined at Motorola

rial, but nothing ever really panned and at a company Motorola spun off Channel out—until recently. in 2004, Freescale Semiconductor. It’s easy enough to understand, At that time, the engineers n-type regions at least in general terms, why find- making silicon MOSFETs were

ing a gate insulator for silicon was beginning to experience trou- p-type In0.53Ga0.47As easier than it has been for gal- ble with their gate insulator too. lium arsenide. Silicon dioxide is a As the dimensions of these tran- p-type InP substrate native oxide of silicon—a naturally sistors shrank, the silicon diox- forming coat that grows when sili- ide insulating the gate no lon- con is exposed to oxygen. By good ger functioned well. Indeed, it An experimental transistor of indium gallium fortune, silicon dioxide makes for became so thin that electrons arsenide (blue) is built on a bed of indium phosphide (gray). Positive voltage applied to the gate draws electrons into the an excellent chemical marriage could pass through it as if it were channel between the silicon-doped n-type regions beneath with the silicon it covers: only a sieve. Great effort went into the source and drain electrodes, allowing current to flow. one out of 100 000 silicon atoms finding alternative materials

40 INT • iEEE Spectrum • september 2008 www.spectrum.ieee.org www.spectrum.ieee.org september 2008 • IEEE Spectrum • INT 41 once. If you want additional coats— content allowed us to engineer

that is, a thicker film of Al2O3—just the substrate’s electronic prop- Years in the Making repeat the application of trimethyl erties as required, instead of try- aluminum and the second carrier, ing to work around the givens of 1960 Dawon Kahng 1995 Minghwei and Martin M. Atalla Hong and Matthias water vapor, in alternating steps. a particular material. After much at Bell Labs invent Passlack at Bell Labs Once you grow a suitably thick experimentation, we settled on a the MOSFET. deposit a gallium layer of aluminum oxide on gallium composition that had a 65:35 ratio oxide–gadolinium arsenide in this way, you use tradi- of indium to gallium. With it we 1965 Hans Becke, oxide insulator on Robert Hall, and a III-V substrate. tional lithography to construct the were able to build a MOSFET that Joseph White at drain, source, gate, and other com- carried more than 1 ampere per RCA devise the first 2001 The author ponents of a MOSFET. No special millimeter of channel width—the gallium arsenide of this article, Peide processes are required. The rub highest current density ever pro- MOSFET using Ye, and Glen Wilk is that the transistor you’ll end up duced in four decades of work silicon dioxide for deposit aluminum the gate insulator. oxide insulator on with will be a dud: it won’t pass any on gallium arsenide MOSFETs. a gallium arsenide more current through its channel Indeed, it was so large that it ini- 1979 Takashi substrate using than did some of the failed designs tially sent our semiconductor Mimura at Fujitsu atomic-layer of decades past. parameter analyzer off scale! Laboratories deposition. invents a type of One well-known difficulty gallium arsenide 2005 Intel hen I came to with this approach is that indium FET: the high- announces interest in Purdue three and gallium arsenide has very poor electron-mobility III-V semiconductors a half years ago, Yi mechanical properties, so poor transistor (HEMT). for future Xuan, a postdoctoral that it would be problematic, if microprocessors. w investigator in my not impossible, to use it to make 2007 The author research group, and I took on this wafers. Pure gallium arsenide and his colleagues problem of dismally poor current is much more robust. Our wafer measure record- capacity. At about that time, Intel supplier, a UK-based company breaking current for a III-V MOSFET. announced that its engineers were called IQE, was able to overcome seriously considering the use of this hurdle by growing a thin III-V semiconductors in its future layer of indium gallium arsenide time. After 2003, our team con- chips. IBM, too, made its interest in on a thick base of indium phos- tinued to study this approach at this technology known. The quest phide. These two compounds Agere Systems, in Allentown, Pa., for speed, it seemed, was driving a have crystal lattices of similar a spin-off of Bell Labs and Lucent renaissance in research on how to sizes, so they bond reasonably Technologies, where our group make III-V semiconductors for dig- well together. And the mechan- had been moved. The maneuver ital applications. But despite all this ical properties of indium phos- succeeded better than we could attention from some of the biggest phide, while not ideal, proved have dreamed. guns in the industry, nobody had good enough to allow us to con- That’s not to say we were able a clear idea about how to achieve struct various test transistors. to make a perfectly functioning sufficient current-carrying capacity Passlack and his co-workers at MOSFET out of gallium arsenide for III-V MOSFETs. The challenge Freescale Semiconductor and the straight off. Rather, what stunned was greatest for those operated in University of Glasgow have also us early on was that atomic-layer enhancement mode, meaning that been experimenting with indium deposition allowed us to apply the electrons flow from source to drain gallium arsenide over the past

Al2O3 despite having done nothing only when a voltage is applied to the few years, using a gallium oxide– to remove the troublesome native gate, as is the case for the silicon ­gadolinium oxide insulator. Hong, oxide from the gallium arsenide. MOSFETs found in digital ICs. who is now at National TsingHwa The reason, as researchers at the Based on published work carried University in Taiwan, continues University of Texas at Dallas have out almost a decade earlier at Bell work on this combination as well. recently detailed, was that the first Labs and on my own research on Although such MOSFETs have carrier, a molecule called trimethyl depletion-mode MOSFETs, which shown a reasonably good ability to aluminum, eats away at gallium switch off when voltage is applied carry current, they would be diffi- arsenide’s native oxides, which to the gate, I realized that a related cult to manufacture. The problem despite all reasonable precautions III-V semiconductor—indium gal- is that they require two applica- tend to cover the substrate. It’s the lium arsenide—would serve bet- tions of a high-vacuum deposition atomic-scale equivalent of the mold ter for the channel. In this com- technique called molecular-beam on an old porch floor. And as any pound, indium atoms substitute epitaxy: one to lay down the homeowner knows, if you want to for galliums to a degree that can be indium gallium arsenide and repaint those boards, you’d better adjusted arbitrarily. That is, you can then a second to coat it with the scrape off the gunk first. have mostly indium atoms, mostly gate oxide. Doing molecular-beam Using trimethyl aluminum was gallium atoms, or a 50:50 mix of the epitaxy twice, all the while keep- like having an all-in-one product two bonding to the arsenic atoms. ing things under high vacuum, is that strips, primes, and paints all at Tinkering with the indium possible in the lab, but it would be

42 INT • iEEE Spectrum • september 2008 www.spectrum.ieee.org www.spectrum.ieee.org september 2008 • IEEE Spectrum • INT 43 a challenge for industrial-scale want these transistors to be able to any advantage over silicon for pos- production. operate at low voltages, for instance, itive charge carriers—the “holes,” Research groups at the National so as to reduce another trouble- which are sites in the semicon- University of Singapore and at IBM some source of heating: the power ductor’s crystal lattice that are are pursuing yet another design that expended at the moment the tran- deficient in outer-shell electrons. has lately shown promise, one that sistors switch states. (Indium-rich So it would be very difficult to uses chemical means to add a layer indium gallium arsenide holds great make a high-performance p-chan- of amorphous silicon between the promise in this regard.) Designers nel MOSFET using gallium arse- semiconductor and the gate insula- will also want to ensure that very nide or another III-V compound. tor. This approach resembles a strat- little current flows when a tran- The current consensus is that egy that was attempted two decades sistor is nominally “off,” so that the semiconductor industry will ago and is similar to work going on power isn’t expended—and heat probably employ germanium for now at the University of Texas at isn’t generated—uselessly. Doing those transistors. The Duallogic Austin and at the State University so, all while making these transis- ­academia-industry consortium in of New York in Albany. tors as tiny as today’s silicon won- Europe, for example, is working What’s more, some researchers ders, will be no small feat. to combine germanium and III-V are looking to build a very different In addition, it is likely that semiconductors in this way. kind of III-V field-effect transistor manufacturers will have to find The III-V devices that my suitable for digital applications, ways to place III-V semiconduc- Purdue colleagues and I have one that can function without a tors on top of a silicon wafer. That recently constructed represent a gate oxide at all. These devices is, chip makers will surely aim to whopping leap forward, as these operate similarly to HEMTs in use the compound semiconductors MOSFETs are both easy to fabri- that a semiconductor provides only where they’re needed rather cate and able to carry record cur- the barrier between the gate and a than trying to replace silicon rents. The competing designs offer highly conductive, undoped chan- entirely, although getting all some attractive features too. Still, nel. Intel and UK-based QinetiQ in these materials to work properly many barriers stand in the way of particular have over the past few together turns out to be a tricky their widespread use. In particu- years achieved impressive perfor- undertaking and the subject of lar, chip makers will have to learn mance with a transistor fashioned much research. to mix and match some very dif- this way using an indium anti- One reason for keeping as much ferent kinds of semiconductors on mony channel. Jesús A. del Alamo silicon as possible around is that it a single wafer. Perhaps chip mak- and his colleagues at MIT are also has considerably better physical ers will have to weave together a investigating how to make such properties for making the large patchwork quilt of indium gal- HEMT-like transistors smaller wafers used in semiconductor man- lium arsenide and germanium on and less prone to gate leakage so ufacturing. Also, silicon is cheap a bed of silicon, or maybe it will that they may one day serve for and environmentally friendly, be something even more compli- digital applications. whereas gallium arsenide is expen- cated. But if there’s any lesson to be Indeed, much work goes on sive and, because it contains arse- drawn from the past four decades around the world on bringing nic, potentially quite toxic. of dizzying advances in comput- III-V semiconductors into what Another reason not to expect an ing power, it’s that this industry has long been the sole domain of all–gallium arsenide microproces- thrives on a challenge. o silicon. In addition to the efforts sor anytime soon is that III-V semi- being mounted in industry, aca- conductors can speed up only half To Probe Further Details of the demic teams have formed centers the transistors in a CMOS chip: author’s work in this area are avail- of research at the University of the n-channel ones, which carry able in “High-Performance Inversion- California at Santa Barbara, the current in the form of negative Type Enhancement-Mode InGaAs University of Glasgow, and the charges—electrons. CMOS inte- MOSFET With Maximum Drain University of Tokyo, specifically grated circuits require a combina- Current Exceeding 1 A/mm,” by Y. to carry out these investigations. tion of both n-channel and p-chan- Xuan, Y.Q. Wu, and P.D. Ye, IEEE nel MOSFETs, which together Electron Device Letters 29:294, onsiderable prog- draw power only when they switch April 2008. ress will yet have to states, such as when an n-channel To learn more about the use of a be made before any transistor turns on and the p-chan- gallium oxide–gadolinium oxide insu- of these new kinds nel transistor that’s wired in series lator on a III-V substrate, see “High c of field-effect tran- with it turns off. When not switch- Mobility III-V MOSFET Technology” sistors replace their slower silicon ing between states, such a com- by M. Passlack, R. Droopad, K. counterparts in microprocessors, plementary pair draws no power, Rajagopalan, J. Abrokwah, P. memory chips, and other digital ICs. which is what makes CMOS chips Zurcher, R. Hill, D. Moran, X. Li, H. In particular, device engineers will so energy efficient. Zhou, D. Macintyre, S. Thoms, and I. have to optimize many parameters Although gallium arsenide Thayne at http:// www.gaasmantech. besides current-carrying capac- allows electrons to move through org/Digests/2007/2007%20Papers/ ity and gate leakage. They’ll also it especially easily, it doesn’t offer 12c.pdf.

42 INT • iEEE Spectrum • september 2008 www.spectrum.ieee.org www.spectrum.ieee.org september 2008 • IEEE Spectrum • INT 43