0.35 micron across. The delivery of devices composed of 0.25- and 0.18-micron features is virtually assured; such chips are in the development phase and will ship in the next several years.
But there are signs that this technology is reaching its limits. While the features on the chip die have shrunk, the cost of the equipment necessary to fabricate these devices has ballooned. Intel alone has spent over a billion dollars apiece for the construction of several new "fabs" (the manufacturing plants that fabricate the chips) located in Oregon, New Mexico, and Arizona. Both IBM and Motorola have also broken ground on new high-price fabs.
The soaring costs of these facilities may eventually slow or halt the development of chips sporting ever-smaller features before the technological limits do. Once that happens, what does the microcom
puter industry do next?
As small as these chip features are, they are still made up of huge aggregates of atoms. New computing technologies might operate on smaller scales, possibly at the molecular or even the atomic level. Or fundamentally new ways to handle information might be the answer, such as storing binary data as a holographic pattern whose data can be written or read in parallel.
This month, let's look to the future--specifically at two new storage media and one new CPU technology that may one day supplant silicon. But to do that, we must first examine the technology already in place.
It's Not Just a Good Idea, It's Moore's Law
Since the IC was developed, the number of transistors that designers can pack on a chip has increased at a phenomenal rate. This rate, where the transistor count doubles approximately every 18 months, has become an axiom known as Moore's law. It's named after Gordon Moore, who first noticed this trend in the early 1960s. Within the span
of 10 years, for example, the logic density in the x86 processor has increased 20 times, as shown in the figure
"x86 Transistor Counts"
.
The basis of these ever-higher logic densities is
photolithography
-- the same technology that etches the plates that print this magazine, only more complex. Here's how it works: Companies make ICs by layering patterns of metal or chemically treated (i.e., doped) silicon, one atop another, onto a die of silicon. The layout of these patterns, composed of either conductive or insulating material, builds the transistors that make up the IC's logic gates.
Adding a new layer first involves covering the die with a photosensitive coating. A mask in the shape of the desired pattern blocks light from reaching the coating, as shown in the figure
"The Limits of Silicon Fabrication"
. Chemical processing etches off those sections of the coating that are exposed to the light. Logic gates thus get built, step by step,
in a cycle where another doped layer gets applied, followed by another coating, another mask exposure, and more etching.
To accurately reproduce features onto the die, the wavelength of the light must be at least as small as the features themselves. Current lithographic processes employ a mercury light source whose 0.365-micron wavelength creates the 0.35-micron features. Successfully achieving the smaller 0.25-micron feature size requires the utilization of a krypton-fluoride ultraviolet laser that has a 0.248-micron wavelength.
Still-smaller features will be handled in the future by the use of argon-fluoride lasers with a 0.193-micron wavelength. But achieving 0.1-micron feature size requires optical trickery involving masks that phase-shift the light to improve the resolution. Building even-smaller chip features requires using light sources with even shorter wavelengths. In doing so, chip designers have traversed the electromagnetic spectrum from visible light, to ultraviolet light, and final
ly into X-ray territory.
But using X rays for the photolithographic process introduces a whole new set of production problems. With visible and ultraviolet light, masks are typically four to five times larger than the feature size. When the fab machinery projects the masks onto the die, lenses perform a reduction operation. With X rays, the masks must be the size of the features themselves, since X rays can't be focused with optical lenses. In short, making defect-free masks is as difficult as making the chip itself. Also, materials that are opaque to light aren't necessarily opaque to X rays.
Finally, there's the issue of having a reliable X-ray source. Mark Bohr, an Intel Fellow, hints at the scope of that problem by joking, "Part of the price tag of a future fab, if X-ray lithography is used, might very well be for the construction and operation of an on-site synchrotron."
John E. Kelly III, vice president of systems, technology, and science at the T. J. Watson Research Center, says th
at his group has fabricated logic gates as small as 0.07 micron using X-ray lithography. "They work--they switch--but there are still manufacturing challenges to be addressed," he admits.
Despite these hurdles, Intel and IBM say that current CMOS technology still has a lot of life in it. Says Bohr: "There's no sign of the technology slowing down. If we're going to run into a wall, it's more than 10 years out." Kelly agrees. "With CMOS technology and a lot of hard work, in a decade we'll use X-ray lithography and other techniques to deliver a processor that has 50 million to 100 million transistors and operates at 1 GHz," he predicts.
Light Storage
Future compute-intensive jobs will present technical challenges in other areas besides the development of new processors. Whether they're made of CMOS or a fundamentally new technology, the quantity of data that these processors demand will tax the capabilities of other subsystems in a computer.
The capacities of today's mass-s
torage devices are indicative of this trend. Today, CD-ROMs are a common staple for distributing software, multimedia, and games. That's because they store up to 650 MB of error-corrected data on a single side of a platter. Magnetic-storage techniques are advancing rapidly as well. Within the last year or so, the typical storage capacity of the hard drive in a desktop computer jumped to more than a gigabyte.
Still, future computers will routinely handle hundreds of gigabytes or terabytes of information--orders of magnitude larger than the capacity of any existing CD-ROM or disk drive. Managing such vast quantities of data and delivering it in a torrent to an ultrahigh-speed processor requires a radically different type of storage system.
An optical recording technology known as
holography
shows great promise because it achieves the necessary high storage densities as well as fast access times. This capability occurs because a holographic image, or hologram, encodes a large block of data
as a single entity in a single write operation. Conversely, the process of reading a hologram retrieves the entire data block simultaneously. (For more on the fundamentals behind holographic recording, see the sidebar
"Creating Holographic Storage"
.)
Holographic data storage uses lasers for both reading and writing blocks of data, or "pages," into the photosensitive material. Theoretically, thousands of such digital pages, each containing a million bits, can be stored within the volume of a sugar cube. This is a storage density of 1 TB per cm^3. Practically, researchers expect to achieve storage densities of 10 GB per cm^3--still impressive compared to today's magnetic-storage densities, which are around 100 Kb per cm^2 (not including the drive mechanism) .
At this density, a block of optical media roughly the size of a deck of playing cards would house a terabyte of data. Because such a system can have no moving parts and its data pages are accessed in parallel, it's estimated that da
ta throughput on such a storage device can hit 1 Gbps or higher.
The extraordinary capabilities of holographic storage have attracted the attention of universities, industry research labs, and the government. This interest has sparked two research projects. One is the Photorefractive Information Storage Materials (PRISM) program, a 2-1/2-year project jointly funded by the U.S. Department of Defense's Advanced Research Projects Agency (ARPA) and other project members,
such as IBM's Alamaden Research Center
(the principal investigator), GTE, and Rockwell International. The purpose of PRISM is to research optimal photosensitive materials for storing holograms and to understand their potential for storage.
The second research project is called the Holographic Data Storage System (HDSS). It has the same principal investigators as the PRISM project and includes such participants as IBM's Watson Research Center, Rockwell International, and GTE.
While PRISM investigates media
, HDSS is developing the hardware technologies necessary to implement a practical holographic data-storage system. HDSS concentrates on building several key system components: a high-speed data-input mechanism, a sensor array to recover the data, and a high-powered red-light semiconductor laser (required for holographic I/O). These components will be integrated with the PRISM medium into prototype storage platforms to demonstrate the potential of this technology.
Molecules as Bits
Even smaller objects might serve as storage devices or replace conventional semiconductor memory. Professor Robert R. Birge, director of the W. M. Keck Center for Molecular Electronics, has implemented a prototype memory subsystem that uses molecules to store digital bits.
The molecule in question is a protein called
bacteriorhodopsin
. This purple, light-harvesting protein is present in the membrane of a microorganism called halobacterium halobium, which thrives in salt marshes, where tempera
tures can hit 150. It uses the protein for photosynthesis when the oxygen levels in the environment are too low for using respiration to obtain energy.
Birge selected bacteriorhodopsin because its
photocycle
, a sequence of structural changes that the molecule undergoes in reaction to light, makes it an ideal AND data-storage gate, or flip-flop (see the figure
"Storing Bits in a Molecule"
). According to Birge, the
bR
(where the state is 0) and the
Q
(where the state is 1) intermediates are both stable for many years. This situation is due, in part, to the remarkable stability of the protein, which appears to have evolved to survive the harsh conditions of a salt marsh.
He estimates that data recorded on a bacteriorhodopsin storage device would be stable for approximately five years. "We have lab samples that have held information reliably for two years," he says. Another important feature of bacteriorhodopsin is that these two states have widely d
ifferent absorption spectra. This makes it easy to determine a molecule's current state using a laser tuned to the proper frequency.
Birge has built a prototype memory system where bacteriorhodopsin stores data in a 3-D matrix. He builds this matrix by placing the protein into a cuvette (a transparent vessel) filled with a polyacrylamide gel. The cuvette is oblong and 1 by 1 by 2 inches in size. The protein, which is in the
bR
state, gets fixed in place by the polymerization of the gel. A battery of kypton lasers and a charge-injection device (CID) array surround the cuvette and are used to write and read data.
To write data, first a yellow "paging" laser fires to pump up the molecules to the
O
state. A spatial light modulator (SLM), which is an LCD array, slices this beam so that it excites a 2-D plane of material inside the cuvette. This energized plane of material is a data page that has the abil-ity to hold an array of 4096 by 4096 bits. (See the figure
"H
ow Molecular Memory Works"
.)
Before the protein can return to its resting state, a red data-write laser, located at right angles to the paging laser, fires. Another SLM displays the binary data, and it sections up this beam so that certain spots on the page are irradiated. Molecules at these locations convert to the
Q
state and represent binary 1s on the page. The remainder of the page returns to the rest state and represents binary 0s.
To read data, the paging laser fires again, which excites the targeted page into the
O
state. This is done to further widen the absorption spectra differences between the digital 0s and 1s (the
Q
state). Two milliseconds later, a low-intensity red laser bathes the page. The low intensity is required to prevent the molecules from flipping into a
Q
state. Molecules representing 0s absorb the red light, while those in the binary 1 state let the beam pass through. This creates a checkerboard pattern of light and dark spots on t
he CID array, which captures the image as a page of digital information.
To erase data, a brief pulse from a blue laser returns molecules in the
Q
state back to the rest state. The blue light doesn't necessarily have to be a laser; you can bulk-erase the cuvette by exposing it to an incandescent light with ultraviolet output.
To ensure data integrity during selective page-erase operations, Birge caches several adjacent data pages. The read/write operations also use 2 additional parity bits to guard against errors. A page of data can be read nondestructively about 5000 times. Each page is monitored by a counter, and after 1024 reads, the page is refreshed via a new write operation.
How fast can data be accessed with this design? While a molecule changes states within microseconds, the combined steps to perform a read or write operation take about 10 milliseconds. However, like the holographic storage system, this device obtains data pages in parallel, so a 10-MBps rate is possible
. This speed is similar to that of slow semiconductor memory.
By ganging up eight storage cells so that entire bytes can be accessed in parallel, Birge believes an 80-MBps data rate is possible. Maintaining this throughput depends on how you implement the memory subsystem. In some versions, the SLM does page addressing. Less-expensive designs use galvanometric mirrors that slew the beam to the correct page. While the SLM offers a millisecond response time, it also costs four times as much.
Says Birge: "Such a system would operate nearly as fast as semiconductor RAM until a page fault occurs. Then we have to reposition the laser beam to access pages on the other side of the container. Depending on the design, we can keep the page-fault access time in the milliseconds so that the memory system behaves like a hard drive during paging. Page caching appears to solve the access-time problem, but it's expensive because of the large page size [about 1.7 MB per page]. If you're willing to spend the money
to cache about 10 pages, then you can eliminate the paging effects."
Theoretically, the cuvette described could hold 1 TB. Practically, Birge has stored about 800 MB on the cuvette, and he hopes to achieve a storage capacity of approximately 1.3 GB. Problems with the lens system and protein quality limit the system to this amount for now.
However, the merits of molecular storage have garnered sufficient interest that three of NASA's Space Shuttle missions explored methods to improve the manufacture of the data cubes by using microgravity. The resulting material was more homogeneous and provided an enhanced storage density. It remains to be seen, however, whether microgravity manufacturing will be sufficiently cost-effective to justify the observed factor-of-four improvement.
Birge's system, which he categorizes as a level-I prototype (i.e., a proof of concept), sits on a lab bench. He has received additional funding from the U.S. Air Force, Syracuse University (Syracuse, NY), and the W.
M. Keck Foundation to develop a level-II prototype. Such a prototype would fit and operate within a desktop personal computer. "We're a year or two away from doing internal testing on a level-II prototype," says Birge. "Within three to five years, we could have a level-III beta-test prototype ready, which would be a commercial product."
Can molecular storage compete with traditional semiconductor memory? The design certainly has its merits. First, it's based on a protein that's inexpensive to produce in quantity. In fact, genetic engineering is being used to boost the output of the protein by the bacterium. Second, the system has the ability to operate over a wider range of temperatures than semiconductor memory.
Third, the data is stable. If you turn off the memory system's power, the bateriorhodopsin molecules retain their information. This makes for an energy-efficient computer that can be powered down yet still be ready to work with immediately because the contents of its memory are preserve
d.
Finally, you can remove the small data cubes and ship gigabytes of data around for storage or backups. Because the cubes contain no moving parts, it's safer than using a small hard drive or cartridge for this task.
Quantum Computing
The scale of the function of new mass-storage and memory subsystems has grown progressively smaller. Holographic storage imprints data on crystalline lattices, and the rhodopsin memory system operates on batches of molecules.
But what about the processor itself? Is there a way to replace its machinery? Perhaps with something even smaller: individual atoms. For years, physicists have manipulated individual atoms in the lab. Now they're trying to coax computations out of them. But this work is like nothing you can imagine. At this scale, you get a whole new set of rules: The normal physical behavior that you expect even for minuscule CMOS logic gates no longer applies. Instead, quantum mechanics dictates the manner in which subatomic particl
es behave.
Quantum mechanics has every atom act as either a particle or a wave (the so-called wave-particle duality). This means that when subatomic particles behave as particles, they occupy only discrete energy states, called
quanta
.
When particles behave as waves, they exhibit strange counterintuitive behaviors. As the quantum wave that represents, say, an electron spreads out over time, its location becomes vague, and the laws of probability reign supreme. (The situation is analogous to throwing a rock into a pond: The wave centers around the point of impact the moment the rock hits the water. Over time, the wave spreads out over the surface of the pond and is everywhere.) The electron, in a sense, can be everywhere at once.
This fuzzy state of affairs continues until the electron interacts with another particle or photon that reveals its position, at which point the spread-out wave "collapses" into several localized waves (the electron and the other particle). As an example
of this bizarre action, suppose a minute junction holds an electron. Its presence can be represented as a wave. This wave function has a certain probability that the particle can also be
outside
of the junction. Under the right conditions, the electron escapes from one junction to another by "tunneling" through the junction's walls, simply because the electron's wave function makes it probable to do so.
In the 1960s and 1970s, Rolf Landauer and Charles H. Bennett at the IBM Thomas J. Watson Research Center did research that investigated the basic physics of computing, which laid the groundwork for quantum computing. Notably, Bennett demonstrated abstractly that you could build a molecular computer that implemented a Turing machine.
Around 1980, Paul Benioff of Argonne National Laboratory showed that computing could be done on a system that exactly obeys the laws of quantum mechanics. David Deutsch at the University of Oxford pointed out in 1985 that such a system could do quantum parall
elism. While this research was still in the abstract stage, it indicated that a quantum computer could have greater capabilities than a classic digital computer.
In 1993, Seth Lloyd, who was then at Los Alamos National Lab, showed that many quantum systems, including an ordinary grain of salt, could function as quantum computers. That same year, Peter W. Shor of AT&T Bell Labs demonstrated that a quantum-mechanical computer could execute a practical task faster than any digital computer--factoring large numbers. All these findings have triggered a renaissance in quantum-computing research, where various groups are working on the construction of prototype components that represent quantum-computer "circuits."
The theoretical proposals to implement a basic quantum "gate" vary as widely as the number of research teams that are currently working on the problem. However, two groups have taken some important steps in demonstrations of actual laboratory implementations. This work has been carried out b
y David J. Wineland's group at the National Institute of Standards and Technology (NIST), which has built an XOR gate using an atomic ion held in a trap, and by H. Jeff Kimble's team at CalTech, which uses an optical cavity with a trapped atom to build a quantum phase gate (QPG). This latter gate's output, which modifies the phase shift of input laser beams, might be used to implement a variety of functions.
Constructing these building blocks isn't easy. NIST's logic gate involves a vacuum chamber with four electrodes, as shown in the figure
"How a Quantum-Logic Gate Works"
.
Although the NIST group built a logic gate that implements the truth table of a classic electronic gate, it's important to note that quantum logic doesn't have to function that way. As mentioned earlier, quantum computing can exploit a kind of parallel processing because of that fuzzy-wave behavior of particles, and even the NIST gate exhibits this feature. "The state space of a quantum-computing syste
m is far larger than the state space of a classic computer system, because the quantum system can exist in exponentially many states all at once," says Kimble.
Because of this, quantum bits are termed
qubits
to distinguish them from conventional bits. "A 3-bit register holds only one number, but a 3-qubit register can hold all eight possible numbers until you read it out," according to Chris Monroe, a member of the NIST team.
In theory, this quantum parallelism allows you to perform complex tasks quickly. For example, factoring a large number normally requires a computer to perform numerous divide operations, which can quickly reach an exponential amount of computations for large numbers. "A quantum computer would attack the problem by raising a smaller number to all different powers at once," explains Bennett. "A repeat period for a particular power function tells you how to factor the original number."
Furthermore, a quantum computer does not have to perform digital computation
s. The late Richard Feynman proposed that quantum computers could simulate other quantum-mechanical systems--in other words, operate as analog computers.
This idea is championed by Seth Lloyd, who's now with the department of mechanical engineering at MIT. As an example, Lloyd wants to simulate the time evolution of 40 particles that make up the matter at the core of an exploding star. Performing these calculations digitally would require setting up and working on 2^40 by 2^40 matrices that would accurately describe all the quantum characteristics of these particles, such as their spin.
"It would take 10^24 digital operations to compute the result," says Lloyd. "A TFLOPS system would require a trillion seconds--31,709 years--to compute the outcome. However, by using lasers to program the behavior of 40 ions in an ion trap, a quantum computer would have to operate for only a hundred quantum interactions." Such a quantum analog computer would use the very quantum properties of these particles, suc
h as the spin, to compute the quantum effects of the simulation. Most of the purposes of quantum analog computing are similarly specialized.
Although quantum computing has lots of potential, there are still many problems yet to be solved. According to Landauer, there's the formidable issue of maintaining a coherent quantum system. "A quantum computer has to operate under two conditions that are hard to reconcile," he explains. "The qubits must interact strongly with one another to perform the computations. Yet they must do so without interacting with the environment itself. That's very difficult to do, especially if you're trying to perform computations over any length of time. For example, the thermal vibrations of the frame that holds the bits in their proper positions will cause the quantum logic to lose its coherence. Another problem is that flaws in the equipment cause errors to build up--unlike with digital computation, where at every stage the system is pushed back to a level of 0 or 1."
Monroe admits that "nobody's really studied these issues. Even the XOR gate loses coherence after 10 or 20 operations, perhaps due to minute instabilities in the laser." Bennett and others have investigated the use of error-correcting quantum codes to tackle the problem. According to Bennett: "Peter Shor discovered promising leads in quantum data storage for correcting errors. He proved that we could use 9 qubits to maintain an error-correcting code. It's not efficient, but it works. However, these codes require reliable quantum processing to function. Unfortunately, it looks like we're going to need a breakthrough just to achieve reliable quantum processing."
Although the picture appears bleak, remember that quantum computing as a technology is still in its infancy. The situation is similar to when Bell Labs built the first transistor in 1947. Researchers are just starting to cast some of quantum computing's decades-old theories into real-world components that can do something. Says Kimble of the situ
ation: "Implementing the quantum analog of classical circuits probably isn't the optimum strategy. Quantum physics is a rich and unexplored land where we're still discovering how to do things."
Even if quantum computing's problems are intractable, future processors will be built--somehow. "Between the limits of conventional lithography and moving atoms around, there's a lot of space to build logic gates," says Kelly.
History is littered with technologies that showed great promise but failed to live up to expectations or usability. (See the sidebar
"Whatever Happened to Josephson Junctions?"
as a case in point.) This applies to all the technologies described here, not just quantum computing. Any one of them might founder due to unforeseen technical problems or because of cost issues. However, it's equally possible that offshoots from other disciplines might usher in a breakthrough, just as an eighteenth-century technology -- photolithography -- did for digital electronics.
illustration_link (36 Kbytes)
Chip vendors use photolithography to etch patterns onto doped silicon layers. The smallness of the features is limited by the frequency of the light beam and the resolution of the lens.
illustration_link (10 Kbytes)
The x86
processor's transistor count has increased by 20 times in 10 years.
illustration_link (72 Kbytes)
1. An LCD array steers a yellow paging laser so that the beam excites a layer (i.e., a data page) of bacteriorhodopsin to the O state.
2. Another LCD array sections up a red data-write laser.
3. Spots on the page struck by this beam switch to the Q state, encoding binary 1s.
4. To read out data, the yellow laser fires, pushing the page into the O state.
5. Now a low-level red laser bathes the page. Molecules in the O state (0s) absorb light, while those in the Q state (1s) let the bea
m pass through, striking a CID array.
illustration_link (42 Kbytes)
A photocycle is the sequence of structural changes that a molecule undergoes in reaction to light. The molecule remains at a resting state, known as bR. Yellow light starts the photocycle, where the molecule goes through several intermediate states, known as K, M, and O. If left alone, the molecule returns to the bR state. If the molecule is illuminated with red light during the O state, the photocyle detours into a P state, and then Q. The molecule remains at the Q state until irradiated by blue light, at which point it returns to the bR state. Both bR
and Q are stable configurations and represent a binary 0 or 1, respectively.
illustration_link (50 Kbytes)
An XOR gate built by the NIST research team. The chamber produces an electromagnetic field that suspends one beryllium atom. Two tuned ultraviolet lasers shine through quartz windows and manipulate the atom's state, namely its oscillation and spin. These two characteristics are used to implement a 2-bit register that behaves as an XOR gate. Another laser measures the atom's current state. If the atom fluoresces in response to the read-out beam, it's called a 0. If it doesn't, it's a 1.
photo_link (11 Kbytes)
Future mass-storage devices might use holograms to record digital information on a doped crystal, in a way similar to that of the test apparatus shown here at IBM's Almaden Research Center.
Commercial-scale equipment would be much smaller, have no moving parts, and use a high-powered semiconductor red laser. A crystal the size of a pack of playing cards would hold a terabyte of data.
Tom Thompson is a BYTE senior technical editor at large with a B.S.E.E. degree from the University of Memphis. He writes extensively on Mac-related and general computing issues. You can rea
ch him by sending E-mail to
tom_thompson@bix.com
.
When Silicon Hits Its Limits, What's Next?
