Breaking Moore’s Law
I’ve been warning people about the end of Moore’s Law since 2007, when signs of a slowdown first began to appear. I wrote about the significance of that slowdown here, in 2011. Now, six years later, we have definitely reached that point.
If Moore’s Law had continued as futurists (Ray Kurzweil, for example) insisted that it would, we would already have processors running at gigahertz speeds. Instead, Intel’s fabricators have slowed their ‘doubling time’ down, from the Law’s 18 months, to a wait-time of many years. And, we cannot hope for more than a few additional doublings, even if we wait for decades: the physical laws constrain our use of electrons, at such small scales.
“Won’t new technologies allow us to continue our incredible improvements?”
Somewhat. Printing more layers (e.g. 3D fabrication techniques, first seen in flash memory) lets us squeeze more transistors on the same block of silicon, and keeps connections close together. Graphene, and other 2D materials, may allow us to utilize spintronics and other fancy forms of energy as our 1’s and 0’s. Yet, those advancements are on a timeline that is forever receding, like a nightmarish hallway that grows longer as you run along it.
Worse for us, there is another doubling-paradigm which has not shown signs of stopping: new fabrication technologies generally cost twice as much as the last method. This pattern has held for all of Intel’s fabricators: the price of new facilities has ballooned. That cost was supported, for decades, because computers were a growing share of the economy, computer purchases were a growing share of disposable income, global population was growing, and economic growth meant that that growing population had a growing disposable income, too! A fabricator that was twice as pricey didn’t matter, in those conditions. The increase in chips sold meant that the fabricator’s cost was spread-out over many more chips.
Now, as the rate of increase in demand for chips has slowed, those fabricators’ bills must be divided among roughly the same number of chips. The costs start to add up. My prediction was that fabricators would become too expensive to be worth it. Chip manufacturers would see better returns from older fabricators, which had already paid-down their bill, than they would from a new fabricator; new chips would be too expensive to be marketable.
Until 2D materials technology makes it to market, we will not know if those promised advances will be cheap enough. I worry that we may be at the end of the Chip’s Golden Age.
“But, I don’t really need my laptop to be that much faster…”
You may have all the computing power you need, but Jeff Dean doesn’t. Google, Amazon, Facebook, Microsoft — they are all betting on artificial intelligence, running on their clouds, as the future platform for a dizzying array of services. Those AI-based services will need ever-growing computer power to be trained and utilized. With the death of Moore’s Law, that entire premise is weakened.
Without a steady improvement in chips’ cost-effectiveness, new companies will not be able to compete with the giants of industry, and greater machine intelligence will come at a greater cost. Google hopes its Tensor Processing Units will give it an edge. That lead is only a one-off advantage. What do they do, when their computational needs grow further? My bet: they will have to twiddle their thumbs, until materials science catches up.
To advance AI, fund graphene and borophene research.
Without the materials science improvements that can make photonic and spintronic chips a reality, tech will stall. Companies will be reduced to app-specialization, tailoring AI to each special use-case without improving AI’s overall power. Like the spam of niche apps, today, but for business operations. And, each use-case will allow a one-off marginal improvement for those businesses, not an ongoing growth that fuels productivity. We need serious funding of materials research, or we face a deep, dark winter of technology.
Google, Microsoft, please: look to your future with clear eyes, and see that your growth potential vanishes without materials science. Don’t get lost in neural networks, while you speed toward this ‘soft wall’ in hardware. The global economy really does depend upon it.