Slashdot Mirror
Intel Researchers See Moore's Law Becoming Obsolete
prostoalex writes "A paper, published by Intel researchers, claims we might be the witnesses of Moore's Law becoming obsolete, as the rate of shrinkage for transistors goes lower with each year. In 2018 we might be able to get the chips manufactured with 16-nanometer technology, then one or two more manufacturing processes will shrink it even further, but after that we're facing the physical limits."
Engineers will be able to continue the shrink for another 15 years based on what we know now. However, the cost for designing setting up manufacturing for a chip will continue to increase exponentially. It will only be worth the money to do this for a part that can be sold in the billions, and there will be few such parts. The end will come not because the technologists can't reduce feature sizes any further, but because no one will be willing to sink an investment equal to the GDP of a mid-sized country into a fab.
At least, that's the case for CMOS silicon chips. To get Moore's Law to continue to operate in a meaningful way, something completely new is likely to be needed: maybe molecular gates that self-assemble or something equally exotic.
In 20 years how fast of processor will grandma need to check her email..
You obviously didn't RTFA.
These fundemental limitations are not material specific. When you get geometric feature sized on your transistor where the source and sink are within 4-5nm of each other, the electrons can tunnel from source to sink more than 50% of the time, regardless of the field imposed on the electron. Therefore it cannot be used as a basis for a logic circuit. Essentially you are killed by Heisenberg uncertainty.
I agree, Diamond based transistors look very very promising, mostly for their thermal properties. When you can maintain a very high thermal gradiant, while maintaining your semi-conductor properies, you can clock the chips much faster without having to worry about overheating and thermal effects, but this research article that Paulos wrote if about a much more fundemental problem.
The content of this paper is pretty much old news, but it is actually promising to see this published by Intel researchers. Intel is well aware of the fundemental limits of its current design, this does not mean the end of Moore's law, in it's most general meaning, this just means that Intel will find new better ways to keep increasing it's core competancy, making amazing CPUs with very low manufacturing costs.
-PT
to predict technologies and processes 20+ years down the road is beyondd amusing. You cannot predict breakthroughs and discoveries.
Actually, somebody wrote a paper making just the same observation. It was mentioned a while back in EE-times, can't find it now (obviously). If somebody knows what I am talking about, please post a link. In summary, the authors reviewed publications over the last 20 years or so and found out that the average time to the predicted end of Moore's law is 10 years, IIRC. E.g., papers published in 1980 predicted the end of Moore's law in 1990. Looks like Intel's researches couldn't find the paper either when they made up the 15 year number.
Every year or so, an article is published along this lines. Moores law is obsolete, no more bigger hard disks etcetera. The thing is that Moores law isn't a law as such, but the prediction that a series of revolution will increase computer power by a seemingly nice and constant line. Every time we get to the physical limits, we find other limits to go to.
- - - - - - -
Sample my Google Hacks
There are some other ideas being discussed. One is a return to BJT technology from FET technology. Simplified explanation: A FET is kinda like a capacitor, it pulls a little current each time it changes state. This makes heat frequency dependent.
A BJT generally uses more power, because it is controlled by current rather than voltage, but the current it draws is mostly fixed, it doesn't vary with switching rate like a FET.
At some speed point, it will be more heat efficient to use BJTs than FETs. At least that's the theory.
I've had enough abrasive sigs. Kittens are cute and fuzzy.
The article wasn't based on silicon or anyother substance, but fundimental physics.
From my understanding of the article, the limit toted by Intel is based on leakage due to quantum tunnelling over distances of 5 nanometers or less. Now, IAAP (I am a physicist), and I know that tunneling probabilities have an exponential dependence on both distance and the height of the potential barrier which is being penetrated through. This barrier height depends on the particular materials used to manufacture pn semiconductor junctions; therefore, the OP was correct in pointing out that using different materials can get around the problems which silicon will soon meet.
Tubal-Cain smokes the white owl.
1GB/3GHz = 0.3 Byte/Hz
32MB/100MHz = 0.3 Byte/Hz
Basically, as processors have gotten faster, the resources attached to the processor have gotten correspondingly larger. Thus it takes more clock cycles to initialise these resources and get them ready for use (ie. 'boot' them). The end result is boot times will be approximately constant (as observed).
By way of comparison, my first computer had 6kB of RAM and a 3.6MHz processor. 6kB/3.6MHz = 0.002. As one would expect, this computer booted in milliseconds!
Another odd effect is the current erosion. Once the circuitry of a semiconductor got so small, there mere flow of electrons will actually erode the wirings away, causing breaks in the circuit after a period of time. In another word, CPU lifetime will be shorter and shorter.
In US, you can easily buy enough major firearms to wipe out your neighbourhood but a few little fireworks are banned.
I think we're all missing the point here. Say we have a 2 gHz computer. (By the way, gHz really means nothing.) The speed doubles. Is that same machine twice as fast? Of course not. For sure, the limiting factor is the hard drive. While SCSI may be the most viable option for speed, we don't see drive speeds following any sort of marked increase.
I'll throw some numbers out. These are fictitious. Say we have an application that is processor intensive and read/writes a massive amount. It takes 10 seconds to process a request. (50% hard drive, 50% processor, other factors, like RAM, we don't factor in, we're looking at the point here.) The speed of the processor doubles, cutting the time for the processor in half, leaving us with 2.5 seconds instead of 5 seconds. Continuing the pattern, we soon learn it's exponential decay, with a rock bottom at 5 seconds.
Intel even acknowledged that speed wasn't everything, their very own centrino technology contradicts that Hz are everything to a computer. (running at 1-1.5 gHz clock range).
Sure, this might be useful for people with massive beowulfs, but for 90% of the home and business computing applciations, excluding servers, 1.5 ghz will do just fine, coupled with a decent drive, RAM, and a lack of any sort of windows variant, well, XP's okay. cough, cough
There is no barrier.
I am sure we all remember when we were told that phone lines could not physically hold more than 2,400 bps.
Well, we are at 56k now, and the only reason we stopped there is because cable modems have been invented and there is not as much money in it anymore.
If there is enough money to be had, humans will always find a way to push the limits further and further.
Its not a technology issue though, electron tunneling is a fundamental limit that says you just cannot pile any more transistors into chips made of any solid.
:-)
I think this paper is 'more' (sigh) significant than many are taking it to be. What they are saying is that the electron will no longer be able to provide us with greater computing power in twenty or so years time. Super computer builders prepared to pay will get a little extra milage out of stacking and clever parallelism but your desktop computer will never get any faster after this time using electronics as we understand it.
What we need is a breakthrough as fundamental as the discovery of a new law of nature to get any further.
Quantum computers show some possibility along with self organising molecules to instantate them - but we are still at the practical ability to do this, that we were at with electricity when kite flying in the clouds was a good way to study electrons.
It is prahaps somewhat significant that the number of gates on a chip will be comparable to the number of neurons in the human brain by the end of this decade. Maybe we dont need faster computers at all, maybe the clever thing will be expecting a computer to do something that it cannot do at the momment - think for itself. Sadly creating artificial intelligence has proven a brick wall that has almost no mainstream spin off so far unless you count Microsofts ghastly paper clip...
However my bet is that when the megahertz race is over, the new race will be how to make the compute element more intelligent - through a mixture of software and hardware. Sadly it seems to be a lot more than twenty years away as we cant even program all human brains to read and write despite the several hundred thousand years of development that have been applied to the grey matter
Facts are history now plebs have politics for religion on social media.
Its not a technology issue though, electron tunneling is a fundamental limit that says you just cannot pile any more transistors into chips made of any solid.
When light-through-air microscopes reached the physical limit, we came up with light-through-oil to get a greater magnification than was "physically possible". Then when that reached its limit we replaced the light with electrons....Even if this is a fundamental limit of electrons-through-solid, who says we're limited to that technology?
Jason
ProfQuotes
It's a bit like when my daughter was born, one of the photos I put on her website was captioned "she's doubled in age in the last 24 hours - surely this can't continue". You can get seemingly odd curve shapes when things are young, but you don't take them and extrapolate to the longer term. Everyone knows that, and that's what made Moore's curve amusing.
The staggering thing about Moore's Law is that reality then proceeded to follow it. Unprecented !
You mean that by 2020 we won't be able to keep up with Moore's law?
Golly-gee! That means that we'll only have another 11 doublings of transistor count, meaning we'll be limitted at about 2000 times the number of transistors we have today. Geez, what how would I ever survive with only the equivalent of 2,000 P4/Opteron processers in my desktop?
steve
Oh, you're not stuck, you're just unable to let go of the onion rings.
not a dumb question at all, though I dont realy know the answer as to how much it helps, its a while since I did this stuff. Could someone help me out here..
My take on why this would not help is..
Cooling a semiconductor reduces the number of electrons available to participate in the transport of current through a pn junction. A cmos transistor contains two pn junctions joined by a region of semiconductor which is switched between a p type and an n type semiconductor by means of the gate.
The gate is a conducting electrode held away from the chip by a very thin layer of insulator - thats where the "transistor" name comes from - a voltage applied to the gate sees a very high resistance and enables a current to flow between the pn np or np pn junctions at a low resistance - hence transistor is trans / resistor (pn np and np pn because CMOS is Complementary Metal Oxide Switch, the idea being that whilst one is on the other is off and current only flows through the two when driving the join between the two between one power rail and the other by switching one on and the other off, all in the interests of reducing power dissipation. These days the gate is silicon and not metal, SMOS never caught on in market speak)
( in theory no current flows through the gate insulating layer of silicon oxide except when the voltage is flipped between positive and negative - however in practise making this gate very thin without a current flowing all the time by tunneling and leakage is a complete nightmare for the chip manufacturer - because leakage = current = heat = meltdown).
The big deal however is that, if you cool a semiconductor down enough, be it p or n type - is that the holes and electrons that make the semiconductor able to carry current - they get fewer and fewer because they only exist because the thermal vibrations in the semiconductor crystal have set them free. You end up with a thing that isnt a transistor any more because there is nothing left to carry a current in it.
Tunneling and therefore the size of the circle an electron could be located in is dependant on the temperature of the material it lives in so cooling sounds like a good idea because tunneling would then depend upon the speed it is moving at due to the applied voltage - so you could get somewhere by lowering the voltage after you have cooled the material down enough so that the electron speed depends on the voltage and not just its speed given to it by thermal excitation. But as noted above it couldnt be done in a normal transistor structure anyway because transistors rely on thermal energy to create carriers in the conduction band - and make the n and p type regions that allow it to work in the first place.
so yes cooling makes the circle smaller but it also breaks the transistor.
Facts are history now plebs have politics for religion on social media.
Quoted from Ray Kurzweil in "The Age of Spiritual Machines":
;)
How will the power of computing continue to accelerate after Moore's Law dies? We are just beginning to explore the third dimension in chip design. The vast majority of today's chips are flat, whereas our brain is organized in three dimensions. We live in a three-dimensional world, so why not use the third dimension? Improvements in semiconductor materials, including superconducting circuits that don't generate heat, will enable us to develop chips -- that is, cubes -- with thousands of layers of circuitry that, combined with far smaller component geometries, will improve computing power by a factor of many millions. And there are more than enough other new computing technologies waiting in the wings -- nanotube, optical, crystalline, DNA, and quantum (which we'll visit in chapter 6, "Building New Brains") -- to keep the Law of Accelerating Returns going in the world of computation for a very long time
either way, moore's law is dead. Kurzweil suggests that after moore's law, an exponential boost will occour, with the advent of a new technology, not simply in materials, new research all together, in accordance with the aforementioned "thillbert's law"
A very good point, but I do not quite share your optimisim - digital electronics will hit the endstop in your lifetime and I havent read any post with an alternative yet - though I'm still reading :-)
Also I think your otherwise excellent technology development analogy is broken in this case because there are no known semiconductors that are a heck of a lot better than silicon and a lot of other semiconductors have been studied since way back (1930's ?) So any silicon replacement is only going to be a bit better - not more than ten times better, then thats it, the endstop for digital electronics.
A possible replacement that might have some milage could be a transistor like structure based on superconductors, SQUIDs or Superconducting Quantum Interference Devices have a Trans / Resistor (Transistor) like property and they actualy rely on tunneling to function. These have been capable of running at many GHz for at least the last decade and are used in some specialised frequency and magnetic field measuring meters. However nobody has been able to build anything more sophisticated than a few devices on one substrate and they need to live in liquid helium to function. Maybe room temperature superconducters have a future here. The big deal at the momment is that even if you could make a chip out of them the Intel finding is probably still relevant.
Any three terminal device where one terminal gates a current flow between the other two terminals will run into the tunneling failure and hence the minimum size problem because you just cant get two connections to any physical device any closer together and not expect electrons to start randomly jumping across the device - its a physical limitation of the electron.
So ok you might get realy clever and figure out how to use some new material that has better properties than silicon and better pattern writing tools to draw your circuit even smaller but at some point the physical elements of the circuits get so close together that electrons have a high probability of being anywhere in your circuit - at random! Hence the digital circuit no longer exists in a 1 or 0 state. End of digital electronics.
Interestingly though seem to recall that there is no such thing as a quanta of magnetic flux - so if you could build a computing device that ran on magnetism rather than electrons then you might get somewhere.... mind you electrons and magnets seem to live in the same house, so I dont know, what do you think?
Facts are history now plebs have politics for religion on social media.
Progress is not going to end, but the automatic metronome of Moore's law will no longer be the driver. The rate of progress will slow for a while then start to pick up. The Intel paper says as much.
Incidentally the point of the paper seems to be to push out the end date and fend off rivals proposing the same ideas. The tunneling effect is quite definitely the end point of traditional logic gates. The astonishing part of the paper is that the end they cite is a 16nm process (with a 5nm gate), the smallest scale currently in use is 37nm. In other words there are only four more generations to go, two generations resulting in the feature size halving which means four times the number of transistors. So if the old two year schedule were kept Moore's law comes to an end in 2011.
They also point to the fact that Intel themselves have pushed out their dates for adopting new processes and are planning for three year gaps between generations. I have suspected that Intel has been the main factor in keeping the industry to the roadmap of Moore's law for some time.
Looking for an Information Security student project suggestion?
Try http://dotcrimeManifesto.com/
Back around the turn of the century, folks could have been talking about how we'd reach the physical limits of hot-air ballooning for mass transportation. At a certain point, a blimp can only move so fast against the air, etc.
Then we took a lateral step into airplanes. Whoops, those folks who were thinking linearally were surprised.
Sure, contemporary microprocessors might reach their limits, but some big brains will come up with The Next Way Of Doing It before then.
The big problem isn't the total voltage. It's the electric field--potential change (voltage) per unit distance. As the transistors in a circuit shrink, the field across them goes up. Electrons get pulled across--the system is 'leaky'. This problem imposes a minimum limit on the size of each transistor, and also increases the current draw of the chip.
Heck, this is a problem elsewhere on the die, too. (Electrons always go where you don't want them.) Forget the switches--you get into trouble with a low breakdown voltage between little tiny wires on the chip. Let's say that it takes 20000 volts (generous estimate) to break down a millimetre of insulation. If there are traces on a chip one micron apart, that's a breakdown voltage of ~20 volts. If traces are 20 nanometres apart (0.02 micron), it only takes 0.4 volts to short the chip.
~Idarubicin
Since commercial developers seem to have unlimited ram, disc space and processing speed and creat bloated software. Maybe soon they will focus on good and efficient software design instead of time to market. Ofcause most successful open source software have good design already. =)
I mean, we've seen this kind of crap so often, it is no longer funny, but anyway, I will bite. :)
:) Of course, any constant (moreso exponential) growth will have to stop. How is that news?
1) End draws nearer for Moore?s Law - we do not know that and this might even be false. Remember, Moore himself thought that his observation will only be valid for a decade or so. But instead the end of Moore's Law has been constantly postponed for almost half a century now. It might be that, with increased R&D, in 10 years we will expect the end of Moore's Law in 2025. Then the opposite to the article title is true - the end of Moore's Law is always pushed further into the future.
2) Ignoring the stupid and factually incorrect headline, let's turn to the idea itself that this Law will stop working some day. Well, duh. Obviously, if we are talking about transistors on silicon, we can't increase the density infinitely, because every transistor must have at least one atom and we can only pack the atoms so tightly before they start to fuse.
3) Why do we ignore all computing technologies and concentrate on transistors and silicon alone? Like Kurzweil writes, they are just a small part of the big picture. It might very well be possible to make a computer based on the electron tunneling effect, which complicates traditional transistors.
The truth is - it is possible to fit a shitload of computational capacity in a very small volume. As a minimum, we can fit a computer able to run a human-level AI in a cube 10x10x10 cm. And most likely, we will be able to do 5-20 orders of magnitude better. Most likely, not without Intel's help. Computers will not stop becoming much faster, simply because it is fashionable (or rather it was 10 years ago) to bash Moore's Law.
In short, journalists are complete idiots, we are tired from sensationalist bullshit.
Future Wiki -- If you don't think about the future, you cannot have one.
You're missing the big picture. Whatever happens to the development of semiconductors is utterly and totally moot. Moore's Law is a single trend line among a forest of trend lines all of which describe a process of evolution and expanding intelligence, beginning at a singular start of life, growing and accelerating towards some kind of ultimate sentient informational singularity.
Follow the trends for biological systems leading to a sentient life-form... us. Then the new trend lines concerning language and symbolic thought, the trend lines that describe the advance of technology. Currently we see Moore's Law. Soon we'll be facing genetic processors, molecular assemblers / processors / and unpredictable nonotechnology, quantum computing, and new applications that are beyond the event horizon of our current conceptions.
The rate at which human knowledge doubles is now down to less than 3 years. This curve may in fact be superexponential. There is no reason to believe that advancing technology won't spawn newer technologies that will continue to cause our knowledge to explode ever faster. Moore's Law will yield to manipulations of matter and new ways to process information, that will almost certainly involve added dimensions. With modalities of information processing that are only now beginning to show faint glimmers of the near future, I have absolute faith that this future is imminent, certain, and unavoidable.
Genda Bendte You begin to see
Thanks for the insight; I wasn't aware of the SNL limit. But is it specific to semiconductors, or does it also apply to optical/quantum computers?
Tubal-Cain smokes the white owl.