Archive | 10.01 - Jan 2002 | Feature

Moore's Quantum Leap

Why has the microchip's explosive growth rate never happened before? George Gilder explains the micro microeconomics and why silicon is just the beginning.

In 1965, when the Internet was the inkling of an "intergalactic computer network" in the mind of a mildly demented psychologist by the name of J.C.R. Licklider, Silicon Valley produced more apricots than electronic devices; Steve Jobs was growing hair and learning subtraction; and no one had imagined a silicon DRAM or a microprocessor or a computer smaller than a refrigerator. The prevailing wisdom of theorists at IBM posited the inevitable triumph of a Few Good Mainframes. In the midst of this antediluvian world, the young director of R&D for a subsidiary of Fairchild Camera and Instrument, Gordon E. Moore, contributed an article to an industry journal, exploding a mind-bending prophecy.

In futurism, the favored rule is "you can say what, or you can say when, but not both at once." What made Gordon Moore's essay so delphically dazzling was his prediction of how the marvels of integrated electronics would be engineered - over time. He included a graph with his journal article. With the year on the horizontal axis and the log of the number of components in an integrated circuit on the vertical axis, the graph mapped just four data points - the number of transistors on ICs in 1962, 1963, 1964, and 1965. These points produced a nearly straight diagonal line at 45 degrees across the graph, indicating that the number of components had doubled every year, beginning with 2³ or 8 transistors, continuing with 2⁴, and up to 2⁶, or 64 transistors. The Moore coup was to boldly extend the line through 1975 when 2¹⁶ or 65,000 transistors would be inscribed on a single chip. This feat was achieved in the designated year in a lab at IBM.

The annual doubling pace slowed to an ultimate rate of a year and a half, but with each generation the devices were eminently manufacturable at yields approaching 100 percent. This year, after 27 doublings since 1962, the billion-transistor DRAM chip should once more fulfill the 18-month pace of advance that is now known far and wide as Moore's law.

Every technology touched by integrated electronics has advanced at a radically new speed. In the next two years a single fiber installation will carry more than a month's worth of Internet traffic in one second.

Ask a historian what other technologies have approximated the pace of Moore's law, and he'll tell you none. No other innovation by any metric has come close to doubling at such quick intervals for such a sustained period. Why? The answer lies at the intersection of quantum physics and a phenomenon related to the learning curve called the experience curve.

First documented in the late 1960s under the guidance of Bruce Henderson of the Boston Consulting Group, the experience curve ordains that the cost-effectiveness of any manufacturing process increases 20 to 30 percent with every cumulative doubling of volume. Whereas the learning curve attempts to measure the increase in productivity, the experience curve quantifies the decrease in cost. BCG and its spinoff Bain & Company documented experience curves for cars, golf balls, paper bags, limestone, nylon, and phone calls. In farm products, they limned a curve for chicken broilers.

As an empirical phenomenon, the experience curve describes efficiency increasing with experience and scale in the manufacturing of any product - from pins to cookies, steel ingots to airplanes. At the outset of any production process, uncertainty is high: No one knows how hard the machinery can be pushed; managers must supervise closely, keep large reserves of supplies on hand for emergencies, and maintain high manufacturing tolerances, or margins for error. Without a substantial body of production statistics over time, managers can't even tell whether a defect signals a serious problem recurring in one of ten cases or a trivial one occurring once in a million.

Considered more deeply, BCG's theorem captures the explosive increase in efficiency resulting from the mixture of mind and matter, information and energy. Governing each is entropy. Informational entropy measures the content of a message through the "news" or surprises it contains - the number of unexpected bits. While in communications you want unexpected news (high entropy), in a manufacturing process you want predictability (low entropy). Thermodynamic entropy measures wasted heat and movement: unrecoverable energy. High informational entropy produces high physical entropy, but in any industrial experience curve, the two forms of entropy are being reduced: energy waste and informational uncertainty. The combination of these two negentropic trends accounts for the 20 to 30 percent improvement in productivity.

One striking early demonstration of experience curve magic is found in the history of television, when the chair of the FCC decreed that all future TV sets must contain UHF tuners. Gordon Moore's colleague at Fairchild, salesman Jerry Sanders (now the chair of AMD), knew that among all companies in the world, only his possessed a chip that could do the job: the 1211 transistor. At the time, he was selling the device to the military in small numbers for $150 apiece; since each one cost $100 to build, this brought a $50 gross margin. But Sanders salivated at the prospect of lowering the price a bit and selling large quantities, making Fairchild the world's largest vendor of components for TVs. Then came the bad news. RCA announced a newfangled vacuum tube called the Nuvistor that could also do the job (though not as well) and priced it at $1.05, more than 100 times less than the 1211 transistor.

With production volumes set to rise from the hundreds for military applications into the millions for TVs, Fairchild's Bob Noyce and Gordon Moore foresaw economies of scale that would allow a drastically lower price: They told Sanders to sell the 1211 to TV makers for $5. Sanders ended up diving further, meeting the Nuvistor's price of $1.05 and then going far below it as volume continued to increase. Between 1963 and 1965, Fairchild won 90 percent of the UHF tuner market in the US. The more chips the company made, the cheaper they got, the larger the market they commanded, and the more money Fairchild made on the product. By the early 1970s, Fairchild was selling the 1211 for 15 cents apiece.

But if every production process obeys the experience curve, what made the 1211's saga so striking? Time. In Henderson's theory, volume is crucial to efficiency and learning, but there is no measure of how fast the larger volumes can be produced. Moore's law, on the other hand, is not only explicit on the subject of time, but it is also unprecedented in its pace. By contrast, beginning in 1915, it took automobile production volume not 18 months - but 60 - to double, and another 60 to double again.

What governs production time is the availability of key resources, the elasticity of demand (how much more of the product is purchased when the price drops), and the physical possibilities of the materials and systems applied. With respect to resources, as Moore was also the first to point out, integrated circuits have a vast advantage over other products: Silicon, oxygen, and aluminum are the three most common elements in Earth's crust. Unlike farmers or freeway contractors, who inevitably face diminishing returns as they use up soil and real estate, microchip manufacturers chiefly use up chip designs, which are products of the human mind.

When it comes to demand, the magic of miniaturization allows Moore's law to respond rapidly to almost any increase in the market. Take the case of the 1211. In those days, each TV contained essentially only one transistor, and the number of potential TV sales was limited more or less to the number of households on the globe. That would mean mere billions of transistors. With a total volume of billions, discrete transistors like the 1211 could decrease in cost to the price of their packages, about a dime apiece, but no further. But with the integrated circuit, you could put ever-expanding numbers of transistors together on a single silicon sliver; today just one typical television set alone contains billions of transistors.

More than an abundance of materials or elasticity of demand, however, what makes Moore's law so powerful are the properties of the microcosm. The ultimate science of semiconductors is quantum physics, not thermodynamics. Rather than managing matter from the outside - lifting it against gravity, moving it against friction, melting or burning it to change its form - Moore and his team learned how to manipulate matter from inside its atomic and molecular structure. In the microcosm, as Richard Feynman proclaimed in a famous speech at Caltech in 1959, "there is plenty of room at the bottom." As Moore's law moves transistors closer together, wires between them become shorter. The shorter the wires, the purer the signal and the lower the resistance, capacitance, and heat per transistor. As electron movements approach their mean free path - the distance they can travel without bouncing off the internal atomic structure of the silicon - they get faster, cheaper, and cooler. Quantum tunneling electrons, the fastest of all, emit virtually no heat. Thus, the very act of crossing from the macrocosm to the microcosm meant the creation of an industrial process that burst free of the bonds of thermodynamic entropy afflicting all other industries. In the quantum domain, as individual components became faster and more useful, they also ran cooler and used less power.

If Moore's law were a mere oddity in the ongoing advance of technology, it would be an extraordinary one. More remarkable, though, is that this unprecedented change is not a blip but a beginning. From processors to storage capacity, every technology touched by integrated electronics has advanced at a radically new speed. Today, in fact, the 18-month pace of Moore's law appears slow compared with the three-times-faster rate of the advance of optics.

Emerging as the spearhead of global industrial progress is the fiber-optic technology called wavelength division multiplexing. WDM combines many different "colors" of light, each bearing billions of bits per second on a single fiber thread the width of a human hair. The best measure of the technology's advance is lambda-bit kilometers, multiplying the number of wavelengths (lambdas) by the data capacity of each and the distance each can travel without slow and costly electronic regeneration of the signal. In 1995, the state of the art was a system with 4 lambdas, each carrying 622 Mbits per second some 300 kilometers. This year, a company named Corvis introduced a 280-lambda system, with each lambda bearing 10 Gbits per second over a distance of 3,000 kilometers. This is an 11,000-fold advance in six years. With several hundred fibers now sheathed in a single cable, a fiber installation in the next two years or so will be able to carry more than a month's worth of Internet traffic in a single second.

This process moves a step forward from the seminal effect of Moore's law and the collapse of the price of computation. While the power of microelectronics spreads intelligence through machines, sector by sector, the power of communications diffuses intelligence through networks - and through not just computer networks but companies, societies, and the global economy.

And unlike silicon transistors, with their mass and expanse, photons are essentially without mass, making the dematerialization that began with semiconductors complete. Photonic carriers can multiply without weight in the same physical space. Virtually any number of colors can occupy the same fiber core. The new magic of optics feeds on the ultimate low-entropy carrier - the perfect sine waves of electromagnetism - and can plunge down curves of experience without mass or resistance through worldwide webs of glass and light.

George Gilder (gg@gilder.com) is the author of Telecosm: How Infinite Bandwidth Will Revolutionize Our World and Microcosm: The Quantum Revolution in Economics and Technology, and the author and editor of the Gilder Technology Report (www.gilder.com) .