Tuesday, February 18, 2014

Moore's Law: At Least a Little Longer

One can argue that the primary driver of U.S. and even world economic growth in the last quarter-century is Moore's law--that is, the claim first advanced back in 1965 by Gordon Moore, one of the founders of Intel Corporation that the number of transistors on a computer chip would double every two years. But can it go on? Harald Bauer, Jan Veira, and Florian Weig of the McKinsey Global Institute consider the issues in "Moore’s law: Repeal or renewal?" a December 2013 paper. They write:

"Moore’s law states that the number of transistors on integrated circuits doubles every two years, and for the past four decades it has set the pace for progress in the semiconductor industry. The positive by-products of the constant scaling down that Moore’s law predicts include simultaneous cost declines, made possible by fitting more transistors per area onto silicon chips, and performance increases with regard to speed, compactness, and power consumption. ... Adherence to Moore’s law has led to continuously falling semiconductor prices. Per-bit prices of dynamic random-access memory chips, for example, have fallen by as much as 30 to 35 percent a year for several decades.
As a result, Moore’s law has swept much of the modern world along with it. Some estimates ascribe up to 40 percent of the global productivity growth achieved during the last two decades to the expansion of information and communication technologies made possible by semiconductor performance and cost improvements."

The authors argue that technological advances already in the works are likely to sustain Moore's law for another 5-10 years. This As I've written before, the power of doubling is difficult to appreciate at an intuitive level, but it means that the increase is as big as everything that came before. Intel is now etching transistors at 22 nanometers, and as the company points out, you could fit 6,000 of these transistors across the width of a human hair; or if you prefer, it would take 6 million of these 22 nanometer transistors to cover the period at the end of a sentence. Also, a 22 nanometer transistor can switch on and off 100 billion times in a second. 

The McKinsey analysts point out that while it is technologically possible for Moore's law to continue, the economic costs of further advances are becoming very high. They write: "A McKinsey analysis shows that moving from 32nm to 22nm nodes on 300-millimeter (mm) wafers causes typical fabrication costs to grow by roughly 40 percent. It also boosts the costs associated with process development by about 45 percent and with chip design by up to 50 percent. These dramatic increases will lead to process-development costs that exceed $1 billion for nodes below 20nm. In addition, the state-of-the art fabs needed to produce them will likely cost $10 billion or more. As a result, the number of companies capable of financing next-generation nodes and fabs will likely dwindle."

Of course, it's also possible to have performance improvements and cost decreases on chips already in production: for example, the cutting edge of computer chips today will probably look like a steady old cheap workhorse of a chip in about five years. I suspect that we are still near the beginning, and certainly not yet at the middle, of finding ways for information and communications technology to alter our work and personal lives. But the physical problems and  higher costs of making silicon-based transistors at an ever-smaller scale won't be denied forever, either.