The really neat physics argument limiting Moore's law (to my eye) is the ultimate limit on the information content in a given region of space. It turns out to be impossible to pack entropy (and thus information) more densely than in a black hole. Because the entropy of a black hole scales with surface area rather than volume, Moore's law runs up against the maximum number of bits containable within the solar system (or even the galaxy or the entire observable universe) surprisingly quickly. I haven't looked at the numbers lately, but I seem to recall that those absolute limits show up within a few centuries, tops.
Moore's law says that the number of transistors on a chip doubles approximately every two years. The Core i7 Sandy Bridge has ~2.3 billion transistors [1], and measures 434 mm^2 [1]. Silicon has a density of 2.4g/cm^3 [2] or .0024g/mm^3. Silicon's atomic weight is 28 --- meaning ~610^23 atoms of Silicon weigh 28 grams --- which means a single atom weighs 4.710^-23g.
Let us pretend that Si crystals are simple-cubic (they're actually diamond-cubic [3]). In that case, a 1x1x1mm cube of crystal weighs 2.410^-3g == 5.210^19 atoms. The cube root of this number, 3.710^6, is the number of atoms on a side, so the number of atoms on a mm^2 face is 1.410^13.
The chip is 434 mm^2, so the face has about 6.1*10^15 atoms on it, otherwise known as 6 million billion.
For roundness, say we have 2 billion transistors on the chip today. Log_2(6,000,000) == 22.
If Moore's law continues until the year 2053, there would be more transistors than atoms on the surface of the chip :-)
That is for what is essentially a 2D chip surface. Certainly layering chips has not been a viable option so far but look at the various "nano" machines that have been in science news, today a micron-sized modle truck, perhaps the next/next gen chips will have fractal surfaces that extend/configure/grow dynamically.
Do away with your heat dissipation problem by going to reversible computing, and you aren't restricted to surface atoms anymore.
The black hole argument goes way beyond that, though. It's the ultimate physical limit of the universe, rather than what's potentially practical or imaginable. Even a computer built of neutronium might not reach that limit.
You can think about decision problems (such that you answer "yes" or "no" to a question) as trying to find a path in (very specific) a graph. UL consists (in a hand-waving fashion) of problems that can be stated in such way by undirected graphs of polynomial size. Roughly it means "if I'm in state s and can go to s' I can also go back from s' to s", which is a reversible computation -- reversible processes are not required by thermodynamics to create heat.
In my experience, UL is "unambiguous log", decision problems that require no more than logarithmic space, and also have a single unique proof (if it exists).
It seems like UL-computers would be optimized for finding the path through the maze to the cheese, but log space is a pretty severe restriction...
Undirected st-connectivity is UL complete (and by showing that it's in L it was shown that UL = L). Logarithmic space is just an arbitrary constraint. It could be polynomial or exp-space.
Here's some napkin math on the black hole argument. Information is essentially equivalent to entropy: the entropy S = k * log M, where k is Boltzmann's constant and log M is the number of bits encoding the information. Now, it turns out to be fundamentally impossible to store more entropy in any region of space than would be contained in a black hole of the same size: S_max = (k A c^3)/(4 G hbar), where A is the "surface area" of the region, c is the speed of light, G is Newton's gravitational constant, and hbar is Planck's constant. That means the maximum number of bits of information in a region of space is (A c^3)/(4 G hbar), or about A * (10^69 1/m^2). Our solar system is maybe 1 light year in size: about 10^16 m, so its "surface area" is 4pi r^2 ~ 10^33 m^2. That means that the solar system can hold at most 10^102 bits of information.
If Moore's law says that computer memory doubles in capacity every 2 years, then they'll go up by a factor of 1000 about every 20 years. We currently have computers that hold about 10^12 bits of information, so just 90*20 = 1800 years would get up to that absolute limit, regardless of what technology we used, unless our single computer was bigger than a light year in radius.
Whoops! I had a feeling I was being careless at the end there. It should have been 30x20 years rather than 90x20, so just 600 years until Moore's law is absolutely finished regardless of technology. (That shouldn't feel all that much closer, but it does.)
Its not actually obvious that you can't have more transistors than atoms, but I'm not sure it isn't possible in theory. I can imagine electrons in different orbitals having different interactions, or having for different energy levels in an interaction to effectively double the number of transistors.
There are theoretical limits to the density of useful computation, but they don't correspond neatly to atoms or such.
Well, for example say you are using light based computing and transistors. Light doesn't interfere with itself for the most part so you can reuse the same transistors to add more cores by sending more channels of light through. You just need to add a gateway/filter at the end to let them communicate. Viola 1 million channels at maximum density gives a few hundred thousand times as many transistors as atoms.
Thanks for reminding me of the totally mind-blowing fact that information density of a black hole is proportional to its surface area instead of volume.
Some have argued this makes the universe analogous to a hologram: a 3d projection out of 2d.
We're nearing the limits for silicon, but we're nowhere near the limits of physics. We'll likely lose the nice smooth curve, but the long-term trend started with mechanical adding machines, and it'll probably continue well past silicon.
Within the next decade, memristors will give us a big jump for certain kinds of problems. Down the road, reversible computing would be another huge leap, essentially eliminating the heat problem.
I'm really not impressed with this guy's description. His understanding of a 3d transistor is flawed. Currently 3d transistors are just one layer of the transistor which has a rectangular cross section to reduce leakage. Future 3d processors will have transistors layered on top of other transistors. Granted heat dissipation becomes a harder problem for those processors but "a hard problem" is much easier to overcome than the theoretical laws of physics at 5nm.
Also, I don't like how his go to solution for after silicon is molecular computing. He said himself that silicon breaks down when we are around 5 atoms across. How big are the molecules which we can use as switches? Much more than 5 atoms across. Molecular computing will not get a transistor density any greater than silicon will.
Moore's law is not a law. It's an observation, it's as much a law as my observation that the amount of rain increases in podunk at a rate of 10% per day. There may be some phenomenon where it's true for a little while, but calling it a law just makes me cringe. The law is a self fulfilling prophecy because chip makers make it as a target which is neither too ambitous nor too cautious. It should be called moore's phenomenon.
Moore's law is a relationship between transistor count and cost.
The physical limits of the universe govern transistor density.
Cost may continue to drop in the face of increasing tranistor count for some years: dies will have to get bigger, but yields on big dies have a lot of improvement left. We may see chips get cheaper and cheaper even if the feature size never drops.
Personally, I'm looking forward to cheaper and cheaper computing, rather than more and more powerful.
Raspberry Pi is just the beginning. Wait until Sandy Bridges are as cheap and plentiful as 555 timers. I'd also like to see some sort of worldwide standardization on a single CPU architecture. Innovation will continue long after Moore's law.
24 comments
[ 0.22 ms ] story [ 73.0 ms ] threadMoore's law says that the number of transistors on a chip doubles approximately every two years. The Core i7 Sandy Bridge has ~2.3 billion transistors [1], and measures 434 mm^2 [1]. Silicon has a density of 2.4g/cm^3 [2] or .0024g/mm^3. Silicon's atomic weight is 28 --- meaning ~610^23 atoms of Silicon weigh 28 grams --- which means a single atom weighs 4.710^-23g.
Let us pretend that Si crystals are simple-cubic (they're actually diamond-cubic [3]). In that case, a 1x1x1mm cube of crystal weighs 2.410^-3g == 5.210^19 atoms. The cube root of this number, 3.710^6, is the number of atoms on a side, so the number of atoms on a mm^2 face is 1.410^13.
The chip is 434 mm^2, so the face has about 6.1*10^15 atoms on it, otherwise known as 6 million billion.
For roundness, say we have 2 billion transistors on the chip today. Log_2(6,000,000) == 22.
If Moore's law continues until the year 2053, there would be more transistors than atoms on the surface of the chip :-)
[1] http://www.tomshardware.com/reviews/core-i7-3960x-x79-sandy-...
[2] http://en.wikipedia.org/wiki/Silicon
[3] http://en.wikipedia.org/wiki/Diamond_cubic
The black hole argument goes way beyond that, though. It's the ultimate physical limit of the universe, rather than what's potentially practical or imaginable. Even a computer built of neutronium might not reach that limit.
It seems like UL-computers would be optimized for finding the path through the maze to the cheese, but log space is a pretty severe restriction...
If Moore's law says that computer memory doubles in capacity every 2 years, then they'll go up by a factor of 1000 about every 20 years. We currently have computers that hold about 10^12 bits of information, so just 90*20 = 1800 years would get up to that absolute limit, regardless of what technology we used, unless our single computer was bigger than a light year in radius.
There are theoretical limits to the density of useful computation, but they don't correspond neatly to atoms or such.
Some have argued this makes the universe analogous to a hologram: a 3d projection out of 2d.
Within the next decade, memristors will give us a big jump for certain kinds of problems. Down the road, reversible computing would be another huge leap, essentially eliminating the heat problem.
Also, I don't like how his go to solution for after silicon is molecular computing. He said himself that silicon breaks down when we are around 5 atoms across. How big are the molecules which we can use as switches? Much more than 5 atoms across. Molecular computing will not get a transistor density any greater than silicon will.
Are you listening?
Quantum.
(edit: not that I have a clue what I'm talking about)
The physical limits of the universe govern transistor density.
Cost may continue to drop in the face of increasing tranistor count for some years: dies will have to get bigger, but yields on big dies have a lot of improvement left. We may see chips get cheaper and cheaper even if the feature size never drops.
Raspberry Pi is just the beginning. Wait until Sandy Bridges are as cheap and plentiful as 555 timers. I'd also like to see some sort of worldwide standardization on a single CPU architecture. Innovation will continue long after Moore's law.