Moore’s Law: Alive? Dead? Or resurrected?

The indispensable foundation underpinning the Fifth Techno-Economic Revolution is Moore’s Law, the relentless cost-performance improvement in electronics enabled by the ever-increasing density of transistors in integrated circuits.

This amazing economic phenomenon has been going on for at least 60 years. It was first widely-recognized back in 1965, when Intel’s Gordon Moore observed that “the number of transistors per square millimeter was doubling approximately every two years.” For over 5 decades that has translated into lower costs, higher performance and more reliable computing and networking.    

In 2018, some experts argue that Moore’s law is dead. Other say it’s alive and well. And others claim it has been resurrected in a new form. Who’s right?

Consider the facts.

We last examined Moore’s Law, in 2013, as we were writing Ride the Wave. Now, five years later, we’re looking back at what actually transpired and attempting to understand its likely forward trajectory.
In an era when digital computing drives nearly everything we do, from plowing and watering fields of corn to diagnosing early-stage cancer deep inside the brain, no issue is more important for our economic future. Today, the ever-falling cost of ever-increasing computing power lets us embed connected “machine intelligence” into nearly everything, everywhere. So, the sudden end of this phenomenon would irrevocably change the trajectory of human progress.

Fortunately, if the leaders at Intel are right, that won’t happen anytime soon. They say, “We have good insight into how we will solve the problems [with Moore’s Law] during the next five years.” They also do a lot of path-finding for the five years beyond that point. The bottom line: as of today, “Moore’s Law is alive and well, for Intel.”
 
So, why do a lots of other experts challenge Intel’s claim, insisting that Moore’s Law is sick, if not totally dead. The answer lies in definitions.

Consider the literal wording of Moore’s 1965 observation, “the number of transistors per square millimeter doubles approximately every two years.” Much of the ability to pack more transistors onto a millimeter of silicon involves making individual feature on the chip smaller; for instance, going from 65 nm features to 45 nm features permitted engineers to pack roughly twice as many transistors onto a single millimeter of silicon. Gordon Moore assumed that the cost of a millimeter of silicon would remain constant in real dollars meaning that the cost per transistor would be half as much as you move from one generation to the next about every two years.

But two things have happened to those assumptions in recent years:

1. The time intervals between generations of technology have become longer; and
2. The cost of producing a finished millimeter of silicon has increased because the newest technology is so expensive to develop and deploy.

Some people say Moore’s law is dead because it’s no longer possible to cut the size of features in half every two years. All other things being equal, that implies that the amount of improvement in cost and functionality would no longer be doubling every two years.

However, there’s more to the story. For current and emerging 10-nanometer and 7-nanometer technology, the density is actually increasing faster that implied by the shrinking size of chip features. So, Intel is getting the same year-on-year improvement, even as the time intervals between 22-nanometer and 14-nanometer as well as between 14-nanometer and 10-nanometer, have become longer. In fact, Intel is a larger-than-normal density benefit as they went from 14-nanometer and as they go to 10-nanometer. In essence, they’re taking bigger steps from generation-to-generation, which is enabling them to stay on the historical cost-performance trend line.
 
Intel has been able to do that because of a strategy called hyperscaling. There are several underlying technologies that enable this, but the really important ones are called Self-Aligned Double Patterning, and Self-Aligned Quad Patterning.
 
This new technology raises the cost per millimeter but also permits more transistors per millimeter. As a result, in every generation cost per square millimeter to manufacture wafers goes up, but Intel shrinks the transistors. And at the end of the day, they get a declining cost per transistor. As a result, their cost per transistor is coming down at a slightly better rate than the historical trend. That is why they say, “Moore’s Law is alive and well.”

For consumers and application developers the benefit of continuing Moore’s Law is that the semiconductor industry can improve performance, add features, and reduce costs, all at the same time.
Given this trend, we offer the following forecasts for your consideration.

First, it will be business-as-usual for CMOS integrated circuits over the coming decade.

In addition to making individual features smaller, it’s also possible to produce multi-layer chips. But the problem in both cases is heat dissipation. The next five years is largely a matter of good execution. Beyond that point, we’ll have we’ll have to rely on the ingenuity of the industry to overcome significant barriers.

Second, in the medium term, another paradigm called Huang’s law may become more important for many commercial applications that can harness the power of GPUs.

CEO Jen-Hsun Huang recently pointed out that today, Nvidia’s GPUs are 25 times faster than they were just five years ago. That’s big; if they were advancing according to Moore’s law, they would only have increased their speed by a factor of ten. Even more impressive, the time required to train AlexNet, a neural network trained on 15 million images took six days, five years ago; now it takes just 18 minutes. That’s 500 times faster! Why? Because, these machines in these applications benefit from simultaneous advances on multiple fronts: architecture, interconnects, memory technology, algorithms, and more. The innovation is across the entire stack.

Third, so-called neuromorphic chips will paly an important role in commercializing artificial intelligence.

The biggest commercial opportunities of the 2020s are likely to be driven by AI, and Neuromorphic chips will play a big role there. A research paper by Intel scientist Charles Augustine predicts that neuromorphic chips will be able to handle artificial intelligence tasks such as cognitive computing, adaptive artificial intelligence, sensing data, and associate memory.  They will also use 15-to-300 times less energy than the best CMOS chips use. That’s significant because today’s AI services, such as Siri and Alexa, depend on cloud-based computing in order to perform such feats as responding to a spoken question or command.  Smartphones run on chips that simply don’t have the computing power to use the algorithms needed for AI, and even if they did, they would instantly drain the phone’s battery.

Fourth optical terahertz chips will enable CMOS to integrate with super-fast photonic computing technology.

As documented recently in Laser and Photonics Review, Hebrew University’s Nano-Opto Group has created an optic technology that integrates the speed of optic communications with the reliability and manufacturing scalability of conventional electronics. Optic communications are super-fast but in microchips they become unreliable and difficult to replicate in large quantities. Now, by using a Metal-Oxide-Nitride-Oxide-Silicon (or MONOS) structure, the Hebrew University team has come up with a new integrated circuit that uses flash memory technology in microchips. Once fully developed, this technology could enable standard 8-to-16 gigahertz computers to run 100 times faster and bring us closer to a commercial terahertz chip. And,

Fifth, by the time Moore’s Law hits a serious bottleneck between 2025 and 2035, Graphene-based logic will begin to play a huge role in commercial electronics.

Since 2013, researchers have made a great deal of progress in terms of transforming graphene nanosheets into an electronic technology. By 2030, it will be possible to manufacture cost-effective graphene microprocessors operating at hundreds of gigahertz. At that point, CMOS will begin to yield share to graphene. Moore’s law will finally be finished. And, we’ll have jumped onto a new and equally exciting performance curve.

References

1. Meir Grajower, Noa Mazurski, Joseph Shappir, Uriel Levy. Laser & Photonics Reviews, 2018; 1700190. Non-Volatile Silicon Photonics Using Nanoscale Flash Memory Technology.

https://onlinelibrary.wiley.com/doi/abs/10.1002/lpor.201700190

2. Steven Vannelli, CFA. Knowledge Leaders Capital Blog, April 7, 2017. Moore’s Law: Knowl- edge Economy Firmly On Track.

http://go.gavekalcapital.com/l/109202/2017-04-05/4d6d-nd/109202/37625/Moore_s_Law_and_the_Knowledge_Effect.pdf

3. Dave James. Pcgamesn.com, January 1, 2018. Intel, Nvidia, Please Shut Up About Moore’s Law. It’s Not Dead Yet, or Interesting... or Even a Law.

https://www.pcgamesn.com/intel-moores-law-no-more

4. Tom Simonite. MIT Technology Review, May 30, 2017. How AI Can Keep Accelerating After Moore’s Law.

https://www.technologyreview.com/s/607917/how-ai-can-keep-accelerating-after-moores-law/