Editorial: Moore’s Law Turns 50
Should We Throw a Party or a Wake?
Turning 50 can be problematic. You can’t move as fast as you used to, you tire easily, and things just don’t work the way they should. Fifty years after Gordon Moore’s seminal article first appeared, his famous “law” is showing distinct signs of aging. As Moore’s Law fades into its twilight years, the semiconductor industry itself must change.
In an article published April 19, 1965, in Electronics magazine, Moore postulated that the number of components (transistors) per chip would double every year. A decade later, he modified his forecast to say that future doubling would occur every two years, since progress was slowing after the halcyon 1960s. Over five decades, transistor counts have risen from dozens to billions, transistors are now smaller than the wavelength of light used to draw them, wafers have grown from 1 inch to 12 inches (300mm), and instead of a few layers, modern wafers have more than 20. But through all of these changes, Moore’s forecast has stayed largely on course. For those of us in the business of making forecasts, it’s the gold standard.
Even the gold standard came to an end, however, and Moore’s Law has reached dangerous ground. Although Intel had for decades cranked out new process nodes on a strict two-year cycle, the company took 27 months to get its 22nm node into volume production and 30 months for its 14nm node. The 10nm node is likely to take at least that long. The complexity of each transition is increasing, requiring more effort to achieve the next doubling (see MPR 2/2/15, “Foundries Find FinFETs at 14nm”).
A bigger problem, unanticipated by Moore, is that the cost per transistor has stopped declining, at least at some foundries (see MPR 5/13/13, “After 28nm, No Moore?”). As a result, chips for cost-focused applications, such as most consumer devices, are staying in 28nm LP. If you interpret Moore’s Law as a statement on economics, as Moore himself often did, then it is already dead.
Some high-end products, however, continue to move to new manufacturing technologies, as their customers are willing to pay a premium for lower power and smaller size. But now that new transistors are no longer “free,” we will see judicious additions rather than wholesale doubling of transistor counts. Thus, even high-end products will fall behind the pace that Moore’s Law describes.
At some point, the increasing expense of the next node will make further progress economically infeasible, but the timing of that point remains uncertain. Mark Bohr, Intel’s top IC-process engineer, likens the semiconductor industry to a car driving down a dark road. Our headlights show the next five years or so, he says; after that, we can’t see if the road is clear or if a huge tree is blocking the way.
For more than 40 years, the road was clear, but now it is strewn with leaves and branches. It’s reasonable to assume a fallen tree lies somewhere ahead, but until our headlights reach it, we won’t know for sure. Progress to 10nm and 7nm looks good, and 5nm seems likely, but this progress will come at increasing cost. Sooner or later, Moore’s Law will reach its end; depending on how much you (or your customer) are willing to pay, it may happen sooner.
Chip designers must adapt to this new environment. Without transistor progress, changes in architecture, microarchitecture, circuit design, packaging, and software will all play a role in improving processor performance (see MPR 8/26/13, “What Comes After the End”). In upcoming weeks, we will publish a series of interviews with leading technologists such as Henry Samueli and Sehat Sutardja that discuss life after Moore’s Law.
Although Gordon Moore’s 1965 article focused on transistor density, he also opined on what those transistors could be used for. At a time when chips held only a few dozen transistors, he foresaw that “integrated circuits will lead to such wonders as home computers…automatic controls for automobiles, and personal portable communications equipment… Integrated circuits will also switch telephone circuits and perform data processing.” If he had headed to Vegas after writing that, he would have cleaned up at the roulette table.