Cao's Law
INSIDE THIS REPORT:
Cover Story: Avanex tests Corvis’s limits; MFNX lays dark plans for TDM; Xros, Calient diffuse paradigm scandal; Avanex banishes bell bottoms; Corning’s coup
Charts: Net Econ 101
Backpage: Telecosm Table
We will get to the companies of the month in due course—opportunities are breaking out all over. But first I would like to introduce the book of the month (and perhaps of the decade; time will tell). It is called Collective Electrodynamics and it was written by Carver Mead in his copious free time while launching a revolution in the camera business with the Foveon imager. With a record breaking sixteen million analog pixels on a single chip, the Foveon device represents an unprecedented extension to analog of the transistor densities of Moore's Law (predicting a doubling of transistor counts every eighteen months). In analog, where devices do not merely switch on and off but convey information through continuously varying states, chips tend to hold from tens to hundreds of devices rather than the millions that Foveon has contrived. Far superior to all other CMOS (complementary metal oxide semiconductor) imagers, the new chip will be manufactured at 0.18 micron geometries by National Semiconductor (NSM), which owns 49 percent of Foveon and which released Richard Merrill, the creator of the fabrication process, to join Foveon. Yet Mead's climactic speech at Telecosm, ending with a prolonged standing ovation, focused less on this amazing new chip and its impact on cameras than on his new book and its promise of a revolution in the physics of the electromagnetic spectrum.
Forty years ago, Mead did the basic research that underlies Moore's Law. Inspired by a 1959 lecture by his Caltech colleague Richard Feynman entitled "There's Plenty of Room at the Bottom," Mead showed that quantum effects previously believed to be a barrier to further miniaturization would actually enable transistors to be made at least a hundred times smaller in linear dimensions than then believed. He concluded that as "the circuits got more complex, they ran faster, and they took less power—WOW! That's a violation of Murphy's law that won't quit."
Also inspired by an idea of Feynman's, Mead's new book may offer a similar promise for the Telecosm and support a new law of technology progress. It recalled to my mind the account of laser inventor Charles H. Townes of a visit to his laboratory in 1956 by the two commanding giants of quantum theory, Nils Bohr and John von Neumann. The two scientists initially told Townes that the laser was impossible because the uncanny coherence of perfectly aligned photons it required violated Heisenberg's uncertainty principle. Mead's new book details how a series of ten experimental developments—from Townes's maser to currents in superconductive rings to quantum wave phenomena at temperatures below one Kelvin—embody a coherent state of matter as inconsistent with Heisenberg's uncertainty theory with regard to particles as the laser was. Swept from the stage by this mounting empirical evidence, in Mead's view, are both Heisenberg's theory as a natural law and the photon as a particle.
To Mead, the very hypothesis of a "particle" is an unnecessary legacy of classical physics, when matter was assumed to be made up of solid atoms. Mead's exploration of electromagnetic coherence—perfect continuous alignment of quantum wave functions—gives implicit support to the idea that frequency spacing in fiber optics can be reduced far below the current limits of around 25 GHz. Without obstacles of Heisenberg uncertainty, "There's Plenty of Room at the Bottom" in wave division multiplexing (WDM).
Avanex tests Corvis's limits
This new result might be termed Cao's Law, after Simon Cao of Avanex (AVNX). Cao startled the Telecosm conference with a prediction that the number of wavelength bitstreams in a single fiber thread could be increased from Corvis's (CORV) current commercial limit of 160 to the 1000 achieved in Avanex experiments, and then to an eventual level of hundreds of thousands. These predictions represent a moment in the history of optics at least as momentous as Moore's Law in microchips announced in 1965, and in Cao's vision entail the possibility of a "switchless network."
By comparison to the vision of the Avanex sage, even the moving mirrors or tiny bubbles in the best new optical switches seem as advanced and elegant as a nineteenth century railroad, embodied in a track of glass. On a railroad Cao explains, "you have one track," and for the train to go to a different place "the track must move." That is what a railroad switch is, a moving track. That's OK if you have only a few trains.
"But what if you have a lot of trains? Maybe you can build a big switchyard. Maybe you can switch a hundred trains. Maybe a thousand trains. But what if you have a million trains?"
Simon is not ready, this day, to announce a million lambdas on a fiber. But more than 1000 WDM lambdas on a fiber adds up to close to a million lambdas on an 864-fiber cable. Moreover, in the lab Avanex has modulated the sidebands of a single lambda channel with 100 different RF subchannels, in effect creating a hundred thousand optical carriers of roughly 120 Mbps on a fiber. Trouble at the train yard. Hundreds of thousands of lambdas in a fiber or tens of millions in a fiber cable cannot be switched like trains.
Even the hope of muxing large "trains" of lambdas together across the switch is dashed by new research from Corning (GLW). Eighty percent of data may be "through traffic," not being dropped or added at the node. But the "through" lambdas are scattered through the fibers. Thus, all the WDM signals must be demuxed and switched individually, with through lambdas moved onto "through" fibers. As a result of their research, Corning's own novel switching architecture—based on a sixteen-fiber network with forty lambdas per fiber—calls for 640 cross-connections, one for each lambda. Light all 144 fibers in a Williams long haul cable, however, with current state of the art 160 channels, and that makes 23,040 cross points. At 100,000 channels per fiber that would be…Oh, never mind.
Simon, however, would say that the ultimate objection to optical switches is not the accumulation of problems but the missed opportunity. Why create the flexibility of 100,000 lambdas in a fiber and then not use it at the switch point, which is just where you want it? Why invent the automobile, as it were, and build multi-lane highways for it, all to rope thousands of cars together and impel them along the highway like a train. "Cars cannot work the way trains work. We cannot have thousands or millions of cars on the highway if the highway must move when a car wants to switch lanes or exit. The car moves, not the highway. On a railroad the track steers the train, but on a highway the car steers itself. That is why people like cars, they go where you say. The highway is a network without switches. We have exit ramps or intersections, but no switches."
MFNX lays dark plans for TDM
On the first page of every elementary text on communications networks, the authors explain why networks have switches. In a network with only two nodes, there would be no switches, just a single wire from A to B, my house to your house. You could do a network with four nodes the same way. One wire each from A to B, A to C, A to D, B to C, B to D, etc. Three wires would terminate in every node. Still not so bad. But if there are absolutely no switches, then every phone gets its own wire to every other phone in an ever-expanding tangle. Connect to everyone in a reasonably sized town and the cost of the wire running to your house would be more than the house itself. That's why we have switches: because we have neither the space nor the money for all the wires that would be needed to replace them. Wires have weight, occupy space, and cost money. If wires had no mass and were free we would not need any switches at all. None.
In the Telecosm we are supposed to reverse the law of the Microcosm and spend bandwidth to save switches. Mostly we do, as when we accelerate a creaky packet network by building bigger dumb pipes rather than stuffing the routers with more elaborate Quality of Service orchestration and traffic management. Buying a dark fiber from Metromedia (MFNX), rather than T1 or T3 "services" from a telco, we substitute bandwidth for Time Division Multiplex (TDM) processing. Hollowing out the computer, transforming it into a network appliance, once again we trade bandwidth for processing overhead on a computer's backplane bus.
Nevertheless, it is undeniable that as copper turned into fiber, both packet and circuit networks switched and processed more and more, not less and less. Even at the box level, switching and routing nodes grew into the millions and at the transistor level to the trillions and beyond.
Xros, Calient diffuse paradigm scandal
The source of this paradigm scandal was the use of fiber as merely a fatter TDM pipe. The telcos assumed that the real physical layer is the fiber rather than the light. If we focus on the fiber, or regard WDM as a mere multiplier of fiber capacity, we focus on bandwidth to the neglect of connectivity. Connectivity—flexible and virtually infinite links rather than nineteenth century railroad connectivity—comes not from the fiber but the light. It comes from WDM.
As always, to find the paradigm go to infinity and work your way back. Imagine that communication power is limitless and that we can broadcast all the world's information to everyone all the time and that each user has the capacity to sort through the bit streams and find the one meant for him. In that network there is no need to switch, buffer, route or process anything, except at the terminals. But organizing that abundant bandwidth into a single shared channel would oblige us to build, at each terminal, the fastest, largest switch the world had ever seen in order to sort through all the world's information to find the bits meant for each user.
If your model blows up at infinity it may be a good idea to tweak it before you get there. As bandwidth grows faster than the number of wires, the switches will become bottlenecks, and our first impulse will be to multiply them to compensate. Only in a network in which the marginal cost and mass of the wires approaches zero will the wires become the prime source of connectivity as well as bandwidth. In that network switches tend to disappear. Lambdas are those wires.
So why aren't switches disappearing?
They are. Optical switches represent the first significant reduction in the amount of switching in the network.
They cut back on switching by reducing the number of ports for any given bandwidth and by handling all bit rates or protocols the same way (thus eliminating separate ports and paths for different bit streams). TDM cannot provide bit rates much above 40 gigs. Even today one port on a Xros or Calient switch can handle at least six times that, assuming a hundred 2.5 gig lambdas, and is ultimately limited only by the melting point of the mirror.
Switchless connectivity, however, implies virtually infinite, nearly massless wiring, at a nominal cost per wire. That is, it implies dividing the mass and cost of a fiber over such a large number of lambda channels as to make the marginal lambda virtually free.
As Simon says, even today railroads are great for some purposes. If you are moving a lot of freight between just two points, there is no need to steer. Tracks bear a lot of weight, and a few switches—which may spend days or years in one position—provide all the flexibility and connectivity you need. An undersea network link is like a freight train. But if you have a thousand lanes or a million lanes, then you want to steer. And you do not want to be steered by a central network control system. You want to steer yourself. You need what Cao calls a "smart photon."
In invoking the smart photon, Simon is not indulging the vain hope of an optical computer, a box in which photons are tortured until they behave like electrons. "No. You steer by frequency."
Imagine hundreds of thousands of frequency selective optical pathways in every fiber. The paths are defined by WDM frequency channels. "On Lambda One, you always go from A to B. On Lambda Two you always go from A to C. So imagine that I am at A. I have a tunable laser that I can toggle from Lambda One to Lambda Two. If I want to go to B, I toggle One. If I want to go to C, I toggle Two. I never need to ask the network to switch."
Such a network has already been designed and built on a small scale in Norway by Telenor, using tunable lasers from Altitun (ADCT), Marconi, and NTT. Such a limited mesh would incur traffic jams. The path from A to B may be blocked. But that's where the hundred thousand lanes come in. Avanex alone can enable thousands on thousands of lanes of long distance light.
Fiber-borne light can enter the mind only in the form of metaphors. A "channel" implies a chute or a tunnel or a pipeline. To get more of them we "space" them "closer together." Of course the channel is not really a container and frequencies aren't really spatial, so the channels can't really be close together or far apart. They flow together. Nevertheless, the light waves forming the channel can be thought of as having a "shape." Good shape is the key to good WDM.
Roughly speaking, we call a channel by its center frequency, the frequency equidistant on either "side" from the center frequencies of neighboring channels. Surrounding this center frequency is a band of frequencies that are just a bit off center. How many? Well, leading WDM systems today operate at 100 gigahertz spacing between channels. Which means between one center frequency and the next there are at least another 100 billion frequencies, just in case you were afraid we'd run out. The frequency bands close to the center frequency are important because they give our lambda channel oomph! You can't have a signal without signal power, and lots of the power is distributed just off either side of the center frequency. Altogether the center frequency and its relatively near neighbors are called the "passband."
Avanex banishes bell bottoms
Across the passband there will be an unfortunate tendency for the amplitudes to slope down and away in a Gaussian bell curve, spreading our signal power into a broad-based bell-bottomed channel. A WDM system that spreads significant signal power out to its bell bottoms, say 50 GHz off center, will just barely accommodate 100 GHz spacing, about one hundred channels in a standard fiber. So no bell bottoms. We want channels that look like shoeboxes stood on end, which allow you to approach the maximum bandwidth per channel because more signal power is focused "on target," in the identifiable frequencies of our WDM network.
How do we shear and shape photonic channels? In addition to wavelength and frequency, the language of light also offers polarity and phase. Waves that are aligned in phase add, or interfere constructively, strengthening the light stream; waves that are out of phase cancel, or interfere destructively. Waves of different polarity can take different paths. Avanex will rule the world of WDM because it uses the entire panoply of light.
Before Avanex the three contenders for dominance of the WDM market were thin-film filters (TFF), arrayed waveguide gratings (AWG), and fiber Bragg gratings (FBG). In these technologies, resolution of frequency channels depends on precise spatial alignment, difficult to achieve and sensitive to heat and other environmental factors. After enhancing commercial WDM channel count more than forty fold and bandwidth several thousand fold in five years, all three techniques now face show-stopping manufacturing challenges that Avanex's PowerMux deftly avoids.
The PowerMux is based on a sophisticated adaptation of the Fabry-Perot interferometer. A classical device that was tested in some early WDM systems, Fabry-Perots were deemed too slow for packet switching and gave way to thin-film filters which are based on similar principles. The core of a Fabry-Perot is an etalon—two highly reflective mirrors facing each other across a cavity. The input signal enters the etalon cavity through the back of one of the mirror surfaces. It then traverses the width of the cavity to the other mirror where a small portion, usually less than 5 percent, of the beam passes through and the rest reflects back to the rear mirror where part is again reflected and part passed through. As these reflections continue back and forth between the mirrors, the light fluxes leaving the filter cavity emerge at different points in their wave cycles, adding in phase for those frequencies which are related to multiples of the one-way propagation delay across the cavity, creating coherent passbands.
Thus, unlike other commercial WDM filters—which split light along different pathways that must be spatially aligned—frequencies in the Fabry-Perot-based PowerMux are self-aligning because the interference patterns are established by the different wavelengths themselves as they repeatedly traverse the same path. "Align one frequency," as Cao explains, "and the other equally-spaced frequency channels automatically line up." The light does the work.
In the Microcosm, self-aligned silicon gates, invented at Fairchild and perfected at Intel (INTC), were critical for the development of integrated circuits. Self-alignment will be similarly crucial to the triumph of WDM. It enables high channel counts because it cancels out the manufacturing showstopper—spatial alignment—that would practically limit the multiple path devices to a few hundred lambdas anywhere outside the lab.
Transforming the linear Fabry-Perot device into an asymmetric non-linear function, Cao can build exceptionally regular rectangular channels with minimal spacing; or in the different design of the PowerShaper can pre-distort the wave to adapt to a particular fiber dispersion profile. Insensitivity to both the spatial and wavelength domains make the PowerMux and the related PowerExchanger (add-drop) scalable to channel counts limited in number chiefly by the resolution of lasers and other components.
Simply by altering the gap between its etalon mirrors a Fabry-Perot could be tuned to yield lambdas across the entire spectrum of optical fiber. Even a tunable AWG would have a far more restricted range, limited by those multiple aligned paths which pre-determine the frequency intervals, and the challenge of heat tuning a device with so many coordinated elements.
AWGs and other multiple path devices are good for pitching high and hard, shooting a few lambdas at high bit rates and high power. But in the Telecosm as in the Microcosm, "low and slow," and "wide and weak," win the day. Operating across the entire available spectrum the PowerMux will multiply channels of nearly any spacing and size, in the future slicing lambdas down to megabit proportions or even less if we like.
With nonlinearities rising by roughly the square of the power, the tighter the channel the lower the power, the better the optics work. Eat your heart out, Murphy. Ultimately, networks bearing a hundred thousand lambdas per fiber will have lower bit error rates and require less complicated and costly contrivances to maintain signal quality than sixteen or thirty-two lambda systems require today. Most of the problems that plague fiber transmission result from the need to maintain coherent passbands tens or hundreds of billions of Hertz wide. The narrower the band, the less power in the drifting bell bottoms, the fewer the distortions entailed in differential wavelength response to the medium. The less the space, the more the room. It would all sound familiar to Carver Mead or to Richard Feynman.
Corning's coup
Moore's Law, as enabled by Mead's insight, functioned for two decades before its implications were accepted even by the computer industry. In fiber optics, outside Avanex, even the titans of the Telecosm still succumb to the temptations of high and hard, pushing bit rate over channel count. Visiting Corning last week—swiftly becoming along with JDS Uniphase (JDSU) our favorite broad-based components company—we listened to Wendell Weeks, Corning's EVP of Optical Communications, enthuse about Corning's conquest of 10 Gbps and 40 Gbps transmission. The maximum allowable dispersion at 40 Gbps, he gleefully tells us, is a mere four one thousandths of what can be tolerated at 2.5 Gbps. (Translation: signal quality control has to get 250 times better to sustain a 15 times increase in speed.)
Weeks loves this problem, because he thinks Corning by virtue of both its prodigious research effort (at 8 percent of sales it leads the industry) and its position as both the world's leading fiber maker and now an industry leader on the components side, is uniquely positioned to solve it for everyone. Crucial to Corning's strategy is its hybrid EDFA/Raman amplifier.
As with most new components for high data-rate systems, Raman amps react differently to different fibers. So Corning is developing a smart card that will detect the connecting fiber and adjust the gain appropriately. A new generation of variable gain EDFAs will push customization one step further adjusting for different spacing in different networks. The tweaking also works from the fiber side with new hybrid fibers bearing dispersion profiles finely tuned to the sensitivities of high bit rate networks. These heroics of integration, perhaps only possible at Corning, would create high barriers to entry for rivals.
The danger for Corning is that they may actually be smart enough to do it. Companies that excel at engineering their way through problems are often tempted to turn problems into business plans, engineering a "gotcha" strategy that holds customers hostage to a self-fulfilling prophecy of problematically optimized proprietary systems. The temptation of speed freaks everywhere, the "gotcha" strategy has killed any number of companies afflicted by engineering narcissism, including Cray and nearly IBM (IBM).
Still, don't bet against Corning. If they can escape the TDM temptation, they can seize the day in WDM. Moving optical component manufacturing out of the craft-guild era and into the industrial age represents one of the Telecosm's greatest challenges. Corning commands as many patents in process technology as in the famous materials. Corning's joint venture with Samsung to automate the manufacture of thin-film filters has already improved production yields by some 25 percent and reduced cycle times from two weeks to two days. Even in a PowerMux world thin-film filters will likely remain a crucial WDM technology for coarser separations.
Already the world's best fiber maker, Corning is successfully playing catch-up in the one area Lucent was ahead. Like Lucent with its AllWave, Corning has eliminated the water spike in its MetroCor fiber, opening up the 1400 nm region, expanding the passband by 40 percent. Holding back AllWave, EDFAs do not operate at 1400 nm, limiting transmission distance to about 40 km. Corning is at work on an amplifier (probably using thulium) that could open up the 1400 nm region for long-distance transmission. Along with JDS Uniphase, Corning can lead the way to cheap components for a ubiquitous fibersphere. But it must grasp the imperative of low and slow—low powered signals in such profusion that fast packet switching becomes entirely unnecessary in the core of the network. |
