The Collapse of Moore's Law: Physicist Says It's Already Happening

People have been predicting the end of Moore's[1] law for years - since at least mid 90s.

Luckily today most people are at the point where they just don't need more power. Most people need enough to run a web browser and do a bit of word processing. They might need enough to open a spreadsheet or a presentation.

Most people would be happy with something like a good tablet and a monitor / keyboard dock.

It would be nice if the churn of "more power" could switch focus to "power efficient" or "better optimisations" or "better architecture" or even "future technology with advanced architectures".

Sure, there are people who need media streaming; or gaming; or compiling; or multicore crunching; or rendering farms; or etc. They'll always have machines. I guess parallelisation needs to improve a bit and clustering needs to get better.

[1] some of these people were using less formal definition of "double the 'power' / 'speed'" rather than more formal "double the number of transistors per IC".

Two things that really interest me about this:

For a couple of decades we've been trading computational resources for programmer convenience. That worked very well when cores were rocketing up in speed year on year. But since we can't count on that, I think we'll be taking a harder look at our tools.

But I think we're going to have to take an even harder look at how we work. As processors stop getting faster, software won't become obsolete as quickly. We've all left a lot of mistakes behind in code killed in platform shifts. But now there may be no escape. I don't know what will come of that, but I'm sure interested to see.

I subscribe to Kurzweil's camp in that I think the halting of Moore's law will pressure a paradigm shift in computing, like the electromechanical, relay-based, vacuum tube, transistor paradigms, the integrated circuit paradigm will cease to be relevant when something like memristors or three-dimensional molecular computing comes into play.

"Using standard silicon"...in other news, Samsung just had a breakthrough making logic circuits in graphene, and hopes to commercialize 100x faster chips by 2020: http://www.asiaone.com/News/Latest%2BNews/Science%2Band%2BTe...

And it looks like memristors will hit the market around 2015, giving us much faster storage and nonvolatile RAM. According to HP, "We put the non-volatile memory right on top of the processor chip, and, because you’re not shipping data off-chip, that means we get the equivalent of 20 years of Moore’s Law performance improvement" http://www.electronicsweekly.com/Articles/22/05/2012/53718/u...

Down the road a bit further, memristors could be used for neural-network coprocessors, since they function a lot like synapses.

The smooth progression of Moore's Law will probably get more jumpy, but the long-term trend goes back to mechanical adding machines.

Since it isn't a physical law, but an observation of average rate of advance of technology, saying that it is 'collapsing' is pretty silly.

Then the article undermines itself by talking about 3d transistors and molecular valves and other possibly ways forward.

In the end, computational density is likely to be limited by heat transfer rates. Having your CPU melt itself is rarely desirable.

Moore's law has been redefined several times already. At the beginning it was every 12 months, then 18, and now it's 24. What's being measured has also changed. It used to be frequency scaling that we measured, but that changed after the debacle of the Pentium 4.

But the essence of Moore's law has always remained: the exponential growth in the capability of ICs. The other relatively constant aspect of Moore's law is that people have always predicted it to last another 10 years. When Moore first formulated the law, that's approximately what he believed. So when somebody says that they expect Moore's law to end within 5-10 years, I take that with a big grain of salt.

What will kill Moore's law will be an unwillingness to continue to spend more money to build semiconductor plants. It used to cost only a few hundred thousand dollars to build an IC plant. This cost has increased exponentially to the point where a plant today costs $10 billion dollars. It's reasonable to suppose that a consortium of semiconductor manufacturers could invest $100 billion in a plant, but on current trajectories, a $1 trillion plant may be necessary to continue Moore's law in 10 years. Do you think that will happen?

I don't know but I've always found it odd to be called a "law", that gives too much scientific credit for something that was clearly derived by simple observation and guesswork and has no underlying science to prove it except that the statement held for some time.

2 years ago I worked at Intel and they talked about Moore's law and how people always said it would end, but it never does due to continued innovation. I think everyone knows they have a lot of people working on a lot of possible solutions to these physics problems. Intel is like a supercharged academia where hitting objectives matters. They don’t want engineers/physicists that say it cannot be done, those people are stuck on the outside writing posts like this one. Those posts come up continually, (do a search for “end of moore’s law 200X) and you will see articles written every year.

Here is a recent quote from an interview with Mark Bohr, Intel's senior fellow and director of process architecture and integration.

“"The end of Moore's law has always been 10 years away, and it will always be 10 years away," he said. He's been hearing predictions about the end of scaling since he joined the industry 30 years ago, so he isn't worried. He said 14nm is in full development and on track for manufacturing readiness in the second half of next year.”

http://forwardthinking.pcmag.com/show-reports/297801-intel-t...

Moore's Law was about the amount of transistors you can put in a chip cheaply, so it may be coming to an end. But I don't think it will stop machines from getting faster - we are seeing lots of progress in multi-core usage even on desktop and in mobile computing (multicore smartphones), and using dedicated hardware outside the CPU socket/package (GPGPU) is getting normal too - those look like new ways of making faster personal machines.

Moore's Law in Silicon will eventually run out. This is true and is physically provable. And there are already a lot of smart minds (at Intel and at a few other companies and at many universities) developing post-silicon technologies. It is possible that we will one day shrink down to true molecular electronics using organic semiconductors or similar. As Feynman said, there's lots of room at the bottom. 3D molecular electronics are exciting for me---architectures will have to be redesigned from scratch, maybe RISCs too. It's an exciting time.

(I omit talking about quantum computing because that's currently mostly theoretical. But back in the day, Turing, Church, et al were talking about computability, etc when a real computer (von neumann architecture) did not exist yet. So maybe (I would be confident enough to say probably) down the line, we'll have a quantum computer, but it may not be soon.)

I don't think we have to worry too much about it until 2030 or so. At least until 2020 it should be a smooth ride, and then they'll probably stacks chips on each other or find some work-around for the next decade or so, until something else comes out.

In terms of comparing clock speed with, say, the brain: isn't it more the parallelization that we're lacking, rather than the speed? In other words lets say we could parallelize a computer network as complicated as the brain... with current processor speeds, couldn't we already re-create effective human intelligence?

I guess I mean to suggest it's the parallelization/inter-connectiviy and algorithmic challenge rather than the speed challenge that has so far avoided the singularity.

Curious if others agree?

I'm not an electrical or thermal engineer, but even if we hit a wall as far as process sizes, couldn't they make bigger dies with more cores?

Like, if we can't get much smaller than a 12-nm process, couldn't we just make the die twice as big and put a big-ol heatsink on it to get to 128 cores or whatever?

It seems like architecture is still a bigger barrier than process size when it comes to getting >16 or >32 cores on the chip. How big will the current ring bus architecture that intel has scale?

I thought the increases in efficiency counted in place of the increases in clockspeed.

Thank god! Finally companies will start optimizing their software again.