Intel’s former chief architect Bob Colwell delivered the keynote address at the Hot Chips conference on Sunday, in a speech I personally wish I’d been able to attend. Colwell, who served as a senior designer and project leader at Intel from 1990 to 2000, was critical to the development of the Pentium Pro, Pentium II, P3, and P4 processors before departing the company. I’ve had the opportunity to speak with him before, and his speeches on processor technology and the evolution of Intel’s most successful designs are always fascinating. The Pentium Pro’s architecture (also known the P6) is arguably the most successful design in the history of microprocessing — echoes of its design principles persist to this day in the latest Haswell CPUs.

Today, Colwell heads up DARPA’s Microsystems Technology Office, where he works on developing new cutting-edge technologies across a variety of fields. In his talk at Hot Chips, he faced up to a blunt truth that engineers acknowledge but marketing people will dodge at every opportunity: Moore’s law is headed for a cliff. According to Colwell, the maximum extension of the law, in which transistor densities continue doubling every 18-24 months, will be hit in 2020 or 2022, around 7nm or 5nm.

“For planning horizons, I pick 2020 as the earliest date we could call [Moore's law] dead,” Colwell said. “You could talk me into 2022, but whether it will come at 7 or 5nm, it’s a big deal.”

Dennard and Moore

It’s important to realize, I think, just how oddsemiconductor scaling has been compared to everything else in human history. People often talk about Moore’s law as if it’s the semiconductor equivalent of gravity, but in reality, nothing else we’ve ever discovered has scaled like semiconductor design. From mud huts to skyrscrapers, we’ve never built a structure that’s thousands of times smaller, thousands of times faster, and thousands of times more power efficient, at the same time, within a handful of decades.

Once you recognize just how unusual this has been, it’s easier to accept that it’s also coming to an end. With Dennard scaling having stopped in 2005 (Dennard scaling deals with switching speeds and other physical characteristics of transistors, and thus heat dissipation and maximum clock speeds), the ability to cram ever-more silicon into tiny areas is of diminishing value. The explosion of accelerators and integrated components into SoCs is partly about fighting the growth of “dark silicon” (that’s silicon you can’t afford to turn on without blowing your power budget) by building specialized functions that can be offloaded into cores that only fire up on demand.

That Moore’s law will continue until 7nm or 5nm is actually extremely reasonable. I’ve heard other engineers speak of being dubious about 10nm and below. But the problem is simple enough: With Dennard scaling gone and the benefits of new nodes shrinking every generation, the impetus to actually pay the huge costs required to build at the next node are just too small to justify the cost. It might be possible to build sub-5nm chips, but the expense and degree of duplication at key areas to ensure proper circuit functionality are going to nuke any potential benefits.

What’s striking about Colwell’s talk is that it echoes what a lot of really smart people have been talking about for years, but has yet to filter into general discourse. This isn’t something that we’re going to just find a way around. DARPA continues working on cutting edge technology, but Colwell believes the gains will be strictly incremental, with performance edging up perhaps 30x in the next 50 years. Of the 30+ alternatives to CMOS that DARPA has investigated, only 2-3 of them show long-term promise, and even there, he describes the promise as “not very promising.”

Innovation is still going to happen. There are technologies that are going to continue to improve our underlying level of ability; a 30x advance in 50 years is still significant. But the old way — the old promise — of a perpetually improving technology stretching into infinity? That’s gone. And no one seriously thinksgraphene, III-V semiconductors, or carbon nanotubes are going to bring it back, even if those technologies eventually become common.

Now read: The death of CPU scaling: From one core to many — and why we’re still stuck