Slashdot Mirror
IBM Scientists Measure the Heat Emitted From Erasing a Single Bit
ananyo writes "In 1961, IBM physicist Rolf Landauer argued that to reset one bit of information — say, to set a binary digit to zero in a computer memory regardless of whether it is initially 1 or 0 — must release a certain minimum amount of heat, proportional to the ambient temperature. New work has now finally confirmed that Landauer was right. To test the principle, the researchers created a simple two-state bit: a single microscopic silica bead held in a 'light trap' by a laser beam. (Abstract) The trap contains two 'valleys' where the particle can rest, one representing a 1 and the other a 0. It could jump between the two if the energy 'hill' separating them is not too high. The researchers could control this height by changing the power of the laser, and could 'tilt' the two valleys to tip the bead into one of them by moving the physical cell containing the bead slightly out of the laser's focus. By monitoring the position and speed of the particle during a cycle of switching and resetting the bit, they could calculate how much energy was dissipated."
Wonder how much heat is dissipated when you mod a post down?
Sheesh, evil *and* a jerk. -- Jade
Yeah, it's kind of like a piece of armor being considered arrow-proof, and then you fire an arrow out of a railgun.
I wonder how you would even measure it, though, and distinguish the heat from a bit changing from the ambient heat from drive operation.
Random Thoughts From A Diseased Mind (Not For Dummies)
It probably reflects the spirit of Landauer's claims. Claims such as this depend upon an understanding of physics, which was much more common in computing back in the days when innovation depended upon an understanding of physics in order to develop new hardware. You also have to consider that a variety of different techniques were used to make computer memories back then, so his claims had to be based upon the underlying physics rather than a particular memory technology. So it is fair game to apply different physical models to prove his claims.
I mean, this is demanded by Maxwell's demon, right? You need to expend energy to store information in order to not violate the 2nd law of thermodynamics. Awesome that they measured it, for sure.
except for endothermic reactions
Say you have two valleys named 0 and 1, and a mountain between. Setting our bit by rolling a ball from 0 to 1 would require energy expenditure, but once the ball is in the valley it is stable and won't roll out again without further input. 0 and 1 may be at different heights relative to each other, but need not be. They might even be at the same altitude. But if 1 were higher than 0, then yes, you would be storing energy in some sort of potential energy form, and may be able to recover that energy when coming back to zero. But you cannot expect to recover all the energy it took to push the ball up the mountain. Any energy required to raise the ball above its destination will have been wasted.
Specifically in the calculation of the Landauer limit, E = kT(ln2), the minimal energy needed to transform a single bit. The interesting thing is that 10^20 bit operations is just a watt. This means that the efficiency of today's computers is just 0.00001%. More details at http://tikalon.com/blog/blog.php?article=2011/Landauer.
Oh it has a law on Wikipedia, must be a waste of time to test or verify it then! Seriously, have a read about how science works before attempting to comment again. A "law" in science is not like a legal law - i.e. it is not a fact merely by self-assertion (a legal law is a law because law makers say so). Scientific "laws" require test and proof; they often require refinement in details. Scientific "laws" do not exist as abstract facts about the universe - they are human attempts to model the universe from the knowledge we currently have. Our limited knowledge means that the detail may be imperfect. A quick survey of the history of science demonstrates that we often get them wrong.
I'm not attempting to challenge the "laws" of thermodynamics - my guess would be that we have the broad picture right (we have a lot of evidence in favour), but again, given the history of science I would be surprised if every detail of taught theory in that area survives the next few hundred years without some modification.
Yes the scientists doing this probably expected some heat to be measured. They were more interested in precisely how much. This is science - an ongoing process.
Yes, the memory will absorb heat, but it costs heat from the hot room. You have to consider the total energy of a closed system and it your naïve approach, the best you can get is a net neutral energy balance. The argument is primarily about the fundamental increase of entropy associated with erasing a bit, and thermal equilibration (between a hot and a cold object) definitely represents an increase in entropy.
Avantslash: low-bandwidth mobile slashdot.
its not necessarily stupid test.. in terms of science, we can estimate the amount of energy from various sources, suchas nuclear plant, or total earth energy, or our solar system, or galaxy... using that estimate, we can put an upper bound on the maximum amount of computational power we have at our disposal.. such as, a certain problem is shown to require X calculational complexity, and X exceeds or the amount of disposable energy in our solar system, thus, X is uncalculable given current technology.
Now, let X be some sort of encryption complexity. now do u see how it could be useful?
Your example only erases in one direction.
To be correct, your experiment must complete erasure in both directions (0 to 1, 1 to 0).
As such, I think you'll find going the other direction is radically more difficult to get energy neutral since you'll be trying to keep thermal bleed from happening whilst flipping your bit.
Does this suggest that by saving up erasures to be done more slowly, perhaps by flipping bits to 0 near the time when they are flipped to 1, could energy be saved and the Landauer limit approached? Also, are there architectures in which a flipping a bit in one direction uses less power, or when blocks of bits can be deselected by some pointer instead of actually erased, trading memory hardware space for power usage?
It's theoretically possible to change the state of a bit without spending energy. Here's a dumb example: think of a closed system (so no energy is being gained or lost) consisting of a box filled with oxygen and only one molecule of water. Divide the box in two halves and say a bit is "0" if the molecule of water is in the left half and "1" if it's in the right half. If you wait a while, eventually the bit will flip with absolutely no change in energy. That's a dumb example, but it shows that there's nothing that requires a "intrinsic state" and energy loss when you move away from it, like you described.
The only time energy dissipation is unavoidable (in theory) is when you erase information. That's a strange concept because, usually, we don't think about "conservation of information" in the same sense of conservation of energy, but there's a relation. A little more discussion with more relevance to computing can be found here: http://en.wikipedia.org/wiki/Reversible_computing.
Not really that surprising, a silicon atom is about 0.11nm and the lattice grid in a silicon crystal 0.54nm, which is still way smaller than the 32nm processors he's talking about. I don't know how many electrons flow down each 32nm path but they're between 0.1nm and 0.000006nm in diameter depending on what model you use - quantum mechanics makes a mess of this anyway - so it's way more than one. If you want single electron calculations you'll have single electron signals, one quantum event and your signal is lost. So the limit is likely to remain a very theoretical limit.
The other thing is that this only includes the operation itself, no clock cycle, no instruction pointer, no caching, prefetching, branching, this is the ideal you could get out of a fixed-function ASIC that only does one thing, not even as programmable as a GPU shader. We already know that there's a significant gain to that, but even supercomputers aren't built that specifically to the task. Formulas must be tweaked, models adjusted, parts must be able to be used in many computers. We've already seen that a GPGPU can beat a CPU by far on some tasks, but even they aren't close to such an ideal.
If you think about this in encryption terms it's not that much... it says you can at most improve 23-24 bits, in encryption most have used the Landauer limit to "prove" there's not enough energy to break a 256 bit chipher by brute force. In some places I don't think it's that relevant either, in for example mobile I think the energy involved in bandwidth use will be more significant. Want to stream a HD movie? It's not the decoding that kills the battery, it's the 3/4G data connection. Just like cameras get better but good optics still isn't small, light or cheap.
Live today, because you never know what tomorrow brings
Landauer's claim was about the relationship between entropy as used in information theory and entropy as used in thermodynamics: specifically, that entropy in information theory is identical to the entropy in thermodynamics. The scientists used this set-up so they could measure a change of exactly one bit (the information-theoretic conception of entropy) while controlling outside heat influences (the thermodynamics conception of entropy), and see if the change in information corresponded to the change in heat as predicted by thermodynamics and information theory.
Without precisely controlling the change in information and precisely measuring the change in heat, the result is much less clear. That's why they used this methodology and equipment. Moreover, as this is empirical evidence for a very general identity between heat and information, the result will hold for computer memory as well.
By monitoring the position [AND] speed of the particle...
Unpossible! Measure one or the other, but not both...
"If anything can go wrong, it will." - Murphy
By your own example, it took energy to erase the bit, just that the energy came from the pre-erasure difference in temperature between the bits and the environment. And the end result of the erasure in your example is an increase in entropy for the (assumedly closed) system of the room plus the bits. So, no, your example does not come close to disproving Landauer.
Wonder how much heat is dissipated when you mod a post down?
Less than the heat that is saved by not displaying the down-modded post in millions of basements all over the world.
Ezekiel 23:20
It says that information is disorder. And thermodynamic entropy is (for some definitions of order) order as well. If you have all of the air molecules in a room compressed into the corner, maybe that's ordered? But that's one small lump of air, and a whole lot of vacuum. Evenly distributed air is more ordered because it is uniform. If you let a system starting in any arbitrary corner-gas configuration (and there are a lot, since each molecule can have any number of different values describing it) progress for X amount of time, you find that almost certainly you have ended up in an even-gas configuration. On the other hand, if you start in an even-gas configuration, and progress for X amount of time, you will almost certainly still be in an even-gas configuration. This may seem at odds with the fact that laws of motion are time reversible (at least if you assume that molecules are like frictionless billiard balls, as physicists are wont to do). But it's not. If you take some specific corner-gas start A , and run it for X time, you will (probably) have an even-gas configuration B. If you take B, reverse the velocity of all molecules, and run it for X time again, you will be at A (again, assuming molecules are frictionless billiard balls). But, with discrete space and velocity, you can count the possible velocity and position vectors. There are a LOT more even-gas configurations than there are corner-gas configurations. So, with a tiny room and only a few molecules, you can establish the chance that after X time starting at even-gas, you end up at corner-gas. And even for very small systems it basically 0. Entropy is the concept of changes to a system that are not reversible, not because of laws of PHYSICS but laws of STATISTICS. The second law is the observation that, by statistics, you will tend to a uniform (ordered) system because there are a lot of ways to go that direction, and very few ways to go the other direction.
Landauer's observation is that any computational device, at the end of the day, stores information mechanically (again, I refer you to the fact that for our purposes, subatomic particles are frictionless billiard balls, so even things like the atom-trap from TFA are mechanical devices). So if you have a 32 bit register, it has 2^32 configurations. If you consider how many possibilities there are for ordered bit flips involving X bit flips total, it's 32^X. And if you start at 0, almost all of those ordered flips will take you to a pretty chaotic state. But if you start from a random state, almost none of those same bit flip orders will get you to 0. So treating the system as a completely mechanical one, thermodynamics applies and puts limits statistical limits on such changes. What Landauer did is establish a maximum circuit temperature T for your memory/CPU, and observe that you won't want Brownian motion breaking your system, so 0/1 need a minimum separation for the system to be useful at temperature T. This puts a lower bound on the state counts, and lets traditional thermodynamics establish a minimum energy dissipation to go from a high entropy state to a low one (like a 0'd out register). What information entropy does is take the same thing and say that therefore the disordered information has intrinsic entropy, since regardless of system design it requires a certain minimum entropy to store that information. It's avoidable if your system is reversible, which is possible if you have more ways to represent a bit pattern the more ordered that bit pattern is. So if you have fewer ways to store 10010101 compared to how many ways you have to store 00000000. It's also beatable if you find a way to store information non-physically. But good luck on that front.
Neat, huh? I took a course on Kolmogorov Complexity, which is somewhat related, and pretty cool.
ASCII stupid question, get a stupid ANSI
I've seen serious claims that "reversable computation" can be done with no energy input at all. What this doesn't cover, of course, is setting up the initial conditions, or extracting the results of the computation. One requirement is that at the end of the computation, the state of the system should be identical to the initial state.
I must admit that I don't understand either the utility, or the feasibility, of such a system. But there have been serious claims that computation does not, itself, require any energy at all.
OTOH, I don't see this experiment as any proof that all writing of a bit requires a minimal amount of energy. It shows that using THIS technology it requires THAT minimal amount of energy. This is a *far* different statement.
I think we've pushed this "anyone can grow up to be president" thing too far.