The end of Moore's Law is like the Voyager leaves the solar system kind of post that just keeps repeating ad infinitum on HN. There are real drawbacks to popular upvotes as means to populate the homepage feed. I hope someday someone figures this out.
Unlike Voyager, this story long predates HN. I remember reading stories like this in the 90s.
At some point, of course, it'll actually happen, and that latest round of such stories will actually be correct. It could even be that this round is that round. But I certainly don't see anything that makes it worth paying any more attention to this one than the ones that have been showing up for decades.
Actually, I expect Moore's law to speed up. Yes Moore's law will end at some point, (in fact physics would tell us it must end this century) but it won't end with the death of the silicon transistor and it won't play out the way people seem to think.
Without the ability to keep shrinking silicon transistors, the industry will innovate in new directions. Eventually (perhaps after a period of relative stagnation in Moore's law.) they will hit on something that is economical and that allows further scaling (graphene? who knows?) and that's when things get really interesting. We've been using the same technology for a long time now with predictable, incremental improvements in shrinking the process. When we switch to a wildly new technology with wildly new characteristics and limits we're likely to see some order of magnitude improvements. It's this exciting transition period that will result, in my estimation, in not just the fastest growth of in the speed of new processors and memory, but a general uptick in the rate of innovation in the industry in general. We've gotten too comfortable with silicon for too long. Change is in the air.
Yes, but even graphene won't allow us to shrink transistors anymore, and this is what all these "end of Moore's Law" articles are really talking about.
>"Moore's law is the observation that, over the history of computing hardware, the number of transistors on integrated circuits doubles approximately every two years."
So that won't be possible anymore, even if we switch to graphene. But if we can actually make graphene chips that allow us to increase the clock speed every year from 5 Ghz, then 7 Ghz, then 10 Ghz, and so on up to 200 Ghz or whatever, then that can be a "solution" for continuing to double performance every couple of years or so.
But if graphene actually has that sort of potential and is not just all hype right now, then maybe that can hold us until 2030-2040, and hopefully by then we'll know how to make quantum computers that can surpass those computers in "general" performance, and can also be used to easily build "apps" on top of them.
I'm not saying that it's not far off, but the only thing I care to hear about Moores law is when it is finally obsolete. There are more interesting things to write articles about.
So what this is saying is that there are limits to how far you can push photo-lithography. Maybe. Yes, it's probably true, but as for how relevant that is...
I'd like to point out that there's an existence proof for putting the power of ~some very large supercomputer into the space of a human head, running off a hundred watts or so.
I am, of course, referring to the brain itself. It's a very different sort of architecture, that's true, but does anyone really think we'll stop before we get close to that?
Are we sure brain have more raw power than PC? Because when I try to outperform my PC at multiplying floats I got owned every time..
Maybe brain is faster at the things it's faster only because it's hardwired to do these things, and we're incorrectly comparing it to general purpose CPU? And brain has this very slow turing machine emulator on top of millions of special-purpose circuits.
And a lot of its power comes from special-purpose circuits, yep; you're only wrong in the order of magnitude. (Billions, not millions)
It also has a couple other 'advantages'. For instance, running at 100Hz means far lower power consumption. Of course, it also means many problems can only be solved through precaching; notice how much better computers are than people at, say, bouncing a ball, while using ludicrously less cpu-time.
For the raw power calculation, though...
We don't have a perfect model of neurons yet, but it appears likely that the average neuron can be modelled by a circuit of 500 or so transistors; contrariwise, a single neuron can easily model a single transistor. Neurons are more complex, no question about it; I believe their behaviour was once compared to vacuum tubes, although that's incorrect. The synapse count is more typically used for comparisons; problem is, each neuron has multiple synapses, and they're not quite independent.
To make the problem more complex, synapses double as memory devices; details of the chemistry of a single synapse can change fairly quickly, apparently allowing for long-term memory and skill formation without changing the gross structure of the brain (that is, graph connectivity). Though the connectivity also changes. The total storage capacity of the brain has been estimated at anything from a few gigabytes (very, very good compression algorithms; you might even call them AI-complete) to several terabytes, but it's nothing to write home about.
Now, there are 6 billion transistors in the most recent CPUs, and 85 billion synapses in the human brain. Naively, this would make the human brain anywhere from 14 to 7000 times more powerful than a modern i7, but of course there are confounding factors.
For starters, the aforementioned frequency differences. The i7 can run serial algorithms; the brain essentially cannot. Any given thought you have may be produced over a fairly long period, but it could be no more than a second; in the latter case, this means only 100 cycles. So it's pretty much got to be put together from previously-made components... meanwhile, the i7 can run several billion cycles of serial logic in that period. This doesn't just mean doing single things faster; per Amdahl's law there are some algorithms that cannot be usefully parallelized, so the i7 can run them while you cannot, and it may therefore be able to do the same thing using less overall resources.
Okay, so that makes the brain seem less impressive. It's especially the case for motor functions - a great deal of the brain is dedicated to that, largely because there's no time to do longwinded calculations, so it all has to be precached. Which is why you need a lot of practice to reliably throw a basketball. It's an immense waste of space - a very cheap computer can do better, where we've invented the algorithms. Our robots are getting pretty good at moving, in some constrained circumstances.
On the flip side there are problems that appear to be inherently parallel, such as face-recognition; all our best approaches so far essentially amount to checking every part of the image against every conceivable possibility. Here the brain has a major advantage, because neurons are damned well suited for the purpose; one might call them well-tuned to running neural networks. Where we're using general-purpose hardware to check several possibilities one after another the brain can have cheap neurons hanging out on the visual cortex, listening for particular signals, and never doing anything else.
Conclusion? Heh, can't say I have one, except.. the brain uses special-purpose circuitry everywhere, but half the time it's from a lack of choice. It can't really do anything else. Regardless, technologically it's still a generation ahead of us, packing more power into a smaller space and energy budget than we can hope to.. but not two, or three, and our own speedy general-purpose systems have their own advantages that will eventually let them outperform it.
EDIT: I should clarify, by 'generation' I mean a human generation - thirty years, not three. At the current rate of progress, which is likely irrelevant; we've already got good enough computers for AI (finally), we just need to use a lot of them.
You need to think bigger and deeper than "using computers for Facebook".
1 billion times faster computers could help us manipulate weather, terraform planets, solve death, have a much better understanding of physics at a much more fundamental level, create fusion or matter-anti-matter engines, etc.
Right. But can't most anything that is computable be done with existing technology operating in parallel? (I'm not sure about solving death, but in general high performance computing)
I'm interested in what's inherent about the doubling per integrated circuit that makes it the limit, rather than just using a lot of circuits. Is it power related? Does it have to do with the Von Neumann bottleneck?
For better or worse, I think 3-D videoconferencing is the short term use that will drive both bandwidth and computing needs.
Why is it so hard to understand that "Moore's Law" is coming to an end - for silicon chips at least?
The point of these stories is that we're getting to transistors being as big as atoms, and I think that's happening roughly around 2-3nm for silicon.
Whether we'll still be able to improve performance for "computers", by switching to graphene transistors, which even if we can't shrink those anymore, we could maybe "optimize" and harvest all the performance from a graphene transistor. But that is probably going to keep us another 10 years or so.
There aren't that many ways out - until we start making quantum computers, and hopefully silicon and graphene transistors will hold us until then.
It has to end at some point, certainly. But there's no particular reason to think it's now. Current technology for chips has been near the end since the 90s. So far, something has always come along to avoid the end.
In a sense it has already happened. The clock speeds in production chips have not increased markedly in the last 10 or so years.
Take a program that has lot of unpredictable branching, random memory accesses and heavily interdependent calculation (the next command always depends on the previous one) and you are not really much faster than old CPU's.
The speedups in single core performance in recent years have mostly come from increasing bandwidth, automatically taking advantage of some concurrency (Out of order execution), more and more SIMD instructions etc etc.
Interestingly enough if you want to write very high performance CPU code it's almost like writing OpenCL for GPU execution. I'd go so far to even claim that right now the best way to write high performance CPU code is to use Intel's OpenCL implementation for their CPU.
Moore's Law was never about clock speeds. It was always about feature size i.e. number of transistors per square inch.
For a long time, transistor count directly enabled higher clock speeds, and so people conflated the two. But the original statement of the Law has no mention of clock speeds.
Why does it matter why single-core performance has continued to improve? It used to be clock speed, now it's something else, but it's still getting better all the time.
