Products You May Like
For the longest time, there’s been a golden rule in technology, often shorthanded as Moore’s Law: Every year, transistors get smaller, and devices get faster and more capable as a result. Do you remember the days of the TigerDirect catalog, when your sweet, bleeding-edge mini-ITX vanity build was somehow obsolete at six months old? That mad scramble was brought to you by the same golden rule. But these days, Moore’s law is less often spoken of as a welcome challenge than as the end of an era, an uneasy concession to a limit. Is Moore’s Law really dead? How small can a transistor be? And what in the world is “dark silicon?” Read on to find out.
What Is Moore’s Law?
Named for Intel co-founder Gordon Moore, Moore’s Law is the observation that the number of transistors on an integrated circuit will double every two years with minimal rise in cost.
In a 1965 article entitled “Cramming more components onto integrated circuits,” Moore noted that the number of components per integrated circuit had been doubling every year, and predicted that the trend would continue for at least another decade. Ten years later, with an eye to the decade to come, Moore revised his theory from doubling once a year to doubling every two years. By general consensus, his prediction has held since 1975, and has since become known as a law.
Technically speaking, Moore’s law is not a law in the same sense as the laws of thermodynamics. Moore’s law doesn’t describe a natural process. Instead, it is an experience-curve law—a way of putting specifics behind the observation that progress builds on progress.
Credit: Hannah Ritchie/Max Roser, OurWorldinData.org
Moore’s law has more or less held since 1975, and it’s impossible to give a complete list of all the advancements we now enjoy thanks to the shrinking transistor. Devices of all kinds are smaller, lighter, and faster.
But the race to obsolescence is a lot more sedate these days. The amount of RAM your console or computer needs has basically stopped rising. Device scaling has slowed down a lot, thanks to a kind of corollary, a principle that goes hand in hand with Moore’s law: Dennard scaling.
Moore’s Law vs. Dennard Scaling
In semiconductor electronics, Dennard scaling, also known as MOSFET scaling, predicts that as transistors get smaller in accordance with Moore’s law, their power density stays constant, so that an IC’s power use rises in proportion with its area; both voltage and current scale downward with decreasing feature length. The idea was proposed in a 1974 IEEE paper co-authored by Robert H. Dennard, after whom the law is named.
Dennard first made a name for himself not for a philosophical observation but a technological innovation. At the same time that Moore was making his first predictions about transistor size, Dennard—then an electrical engineer with IBM—was working on a design for a single-transistor solid state memory cell. In 1966 he received a patent for a game-changing way for computers to store information: dynamic random-access memory, or DRAM.
The viability of Moore’s law is still hotly debated, but Dennard scaling gave its swan song around 2005, thanks to a phenomenon called the “power wall.” Switching CMOS circuits require an amount of power to function that scales exponentially with their clock frequency. This means that as chip features get smaller, they cram in tighter and tighter, like electrified sardines. Long before Dennard scaling reached its “limits at infinity,” the reality of manufacturing tolerances proved its downfall.
There was some warning: In 2001, former Intel CEO Pat Gelsinger remarked that if current trends held, the power density of a given high-speed processor would approach that of a nuclear reactor by 2005, a rocket bell by 2010, and by 2015 it would compare apples to apples with the surface of the Sun. “Business as usual,” said Gelsinger, “will not work in the future.”
Five years later, Mark Bohr, former head of Intel’s manufacturing division, wrote in a 30-year retrospective on Dennard scaling that it is “no longer a sufficient strategy to meet future transistor density, performance, and power requirements.” However, Bohr said, he remains optimistic because of forthcoming technologies such as strained silicon, high-kappa dielectrics, and new metal gates and multi-gate MOSFETs.
“So although the letter of ‘Dennard’s Law’ can no longer be followed, it has gotten us very far over the past 30 years and the spirit is alive and well in transistor R&D facilities around the world.”
The End of Moore’s Law
So, is Moore’s law dead or not? The answer depends on who you ask.
There are good reasons to say that Moore’s law is both dead and not-dead, like a philosophical cousin to Schrödinger’s cat. No matter what, the laws of physics impose a floor on raw feature size: You can’t have a wire thinner than a single atom.
But making wires ever thinner also makes it easier for electrons to get where they shouldn’t, hopping between traces via quantum tunneling. This “current leakage” introduces noise that mucks up the square waves of a binary signal like how JPEG artifacts ruin old photos. Worse, it also makes chips vulnerable to thermal runaway, which can destroy a chip altogether. Like the brain, only a fraction of an IC can be active at any given moment without violating power constraints. (The parts that aren’t active are referred to as dark silicon.)
Under a narrow interpretation, in which Moore’s Law refers to transistor density scaling and cost but not the larger question of whether silicon is still improving, yes, Moore’s Law is winding down. While new lithographic nodes still offer some generational density improvements, performance, power, and area scaling aren’t what they used to be. The node-to-node gains are smaller, and they don’t affect all types of silicon equally. TSMC, for example, recently made headlines for being able to offer a modest improvement in SRAM density at its 2nm node, after having to hold SRAM density constant at 3nm.
On the other hand, as they say, (transistor) size isn’t everything.
If we evaluate the life left in Moore’s Law under a broader, more general “Are computers still getting faster / more efficient?” framework, things look more cheerful. Packaging innovations have let the semiconductor industry carry the baton of Moore’s law farther than Dennard scaling would otherwise allow. Companies like AMD and Intel have responded to the general slowdown with what Moore called “circuit and device cleverness.” Intel has introduced CPU tiles connected together via Foveros, its 3D chip stacking technology, while AMD has its chiplet strategy and its new, bottom-mounted 3D V-Cache. Both companies introduced their new on-chip neural processing units with explicit references to the integration of integer and floating point units in x86 CPUs some 35 years ago.
When it comes to semiconductor manufacturers, opinions on the end of Moore’s law vary with a company’s bottom line.
Established semiconductor firms with an interest in the status quo (e.g. AMD and Intel) are more likely to paint Moore’s Law as alive and well. Intel holds that Moore’s Law “only stops when innovation stops, and innovation continues unabated.” Meanwhile, companies that want to paint themselves as outsiders looking to shake up an established market—we’re looking at you, “Huang’s law” apologists—are more likely to declare that Moore’s Law has died. What better way to paint their own business as the very option that will save the silicon industry?
I’d be remiss in leaving out what Gordon Moore himself thought of his eponymous law. Moore viewed the persistence of his own law as a kind of “violation of Murphy’s law. Everything gets better and better.” As feature size approaches the size of single atoms, Moore said in a conversation with the Economist, there are bound to be some limits—but his model of density scaling has run into obstacles before that looked insurmountable, until they were in the rearview mirror. In a 2015 retrospective, on the 50th anniversary of Intel’s founding, Moore called his own prediction a “wild extrapolation” that ended up being “far more accurate than he could have anticipated.”
Ultimately, I’d be surprised if people didn’t talk about Moore’s law 50 years from now, even if such articles bring journalists like myself out of retirement to grouse about formal definitions, arthritis, and the historical record. It’s an observed trend whose time is drawing to a close—and a bit of verbal shorthand too useful to ever let die.