Landauer's Principle
Every idea in this module has been marching towards a single sentence. The
Szilard
engine showed that a bit of information can be cashed in for
kT \ln 2 of work; the double-well lessons showed that erasing a bit means
blindly squeezing two possible states into one. Here is the sentence, published by Rolf Landauer in
1961 and now bearing his name:
- erasing one bit of information into an environment at temperature T
dissipates at least kT \ln 2 of heat into that environment;
- at room temperature (T = 300\ \mathrm{K}) the bound is
\approx 2.87 \times 10^{-21}\ \mathrm{J} = 2.87\ \mathrm{zJ} \approx
0.018\ \mathrm{eV} per bit;
- the cost is charged per erased bit — logically reversible operations
(bijections) are exempt and can in principle dissipate arbitrarily little;
- it is a lower bound, approached only in the slow (quasistatic) limit; any real,
finite-speed erasure dissipates more.
This is the only fundamental energy cost anywhere in computing. Not arithmetic, not logic,
not moving data — forgetting. Everything else is engineering overhead that better
engineering can shrink; this one line item is charged by the laws of physics themselves.
Derivation 1: entropy bookkeeping
The fastest route is pure accounting. Consider a memory bit that is equally likely to be 0 or 1, and
a RESET operation that leaves it reading 0 regardless. Before, the memory's
information-bearing degree of freedom has two accessible states — entropy
k \ln 2, as the last two lessons established. After, it has one —
entropy zero. The reset changes the memory's entropy by
\Delta S_{\text{memory}} \;=\; -\,k \ln 2 .
But the second law
forbids the total entropy of memory plus environment from decreasing. The environment must
therefore soak up the difference:
\Delta S_{\text{env}} \;\ge\; +\,k \ln 2 .
The only way a plain thermal environment at temperature T gains entropy
is by absorbing heat, \Delta S_{\text{env}} = Q/T. Hence
Q \;\ge\; kT \ln 2 .
That's the whole proof. Notice what drove it: not electronics, not friction — the bare fact that
reset is a two-states-into-one compression. Any physical implementation of any
logically irreversible operation, in any technology that exists or ever will, inherits the bound.
And notice the converse, which is just as important: a bijection never compresses
the state space, so \Delta S_{\text{memory}} = 0 and the argument bites
on nothing. Reversible operations owe no heat. That single exemption is the entire economic case for
this course.
Derivation 2: the double-well protocol, priced
The abstract accounting has a concrete mechanical shadow. Run the erase protocol from
the
double-well lesson — lower the barrier, tilt, restore — and follow the energy. While the
barrier is down, the particle wanders the merged basin; the tilting stage then acts like a piston
compressing the particle's accessible volume by half, from two wells' worth of
territory into one. This is a Szilard expansion run in reverse: an isothermal compression from
V to V/2, and the same integral now runs
against you,
W_{\text{on particle}} \;=\; \int_{V}^{V/2} \! p\,\mathrm{d}V
\;\Rightarrow\; |W| \;=\; kT \ln 2 ,
work that you must supply through the tilt and that the particle, held at temperature
T by its surroundings, immediately sheds as heat. Do the protocol slowly
and you pay exactly kT \ln 2; rush it and friction adds a surcharge. Both
derivations land on the same number from opposite directions — the second law demands at
least kT \ln 2, and the best possible protocol achieves it. That
pincer is what makes it a genuine physical constant of computing rather than a rule of thumb. (For
the connoisseur: "work in, heat out, at fixed temperature" is the signature of a free-energy
calculation — see thermodynamic
potentials for that language.)
How big is the bill?
The bound is linear in temperature — halve T and forgetting gets exactly
twice as cheap, which is one honest reason to compute cold. Here it is as a function of temperature,
and for a few bits at a time:
Now the number that should genuinely surprise you. How does real hardware compare?
const k = 1.380649e-23; // Boltzmann's constant, J/K
function bound(T: number): number { // Landauer's bound per erased bit, in joules
return k * T * Math.log(2);
}
console.log(bound(300)); // room temperature
console.log(bound(300) / 1.602e-19); // the same, in eV
console.log(bound(4)); // in a liquid-helium cryostat
const cmosSwitch = 1e-16; // rough energy per CMOS switching event today, J
console.log(cmosSwitch / bound(300)); // how far above the floor we run
A modern transistor switching event costs on the order of 10^{-16} to
10^{-15} joules — some 10^{4} to
10^{5} times kT \ln 2. Today's chips are
nowhere near the Landauer floor; their heat is still almost entirely engineering overhead
(charging and dumping wire capacitance, leakage), which is why the industry has been able to ride
the cost per operation downward for sixty years. But the floor is there, it is
fixed, and extrapolations of that sixty-year slide hit it within decades. When they do,
the only door left open is the one Landauer's exemption clause points at: stop erasing.
Rolf Landauer arrived at IBM Research in 1952,
when a computer was a room and its power bill was a line in the annual report. Asking "what is the
minimum energy a computation requires?" was then a wide-open question — respectable guesses
held that every logical step must cost at least kT-ish energy. Landauer's
1961 paper, Irreversibility and Heat Generation in the Computing Process, drew the line in
a different place than anyone expected: the tax falls not on computing but only on
discarding. He spent the following decades as the field's genial enforcer, stamping
"Information is physical" on talks, papers and at least one actual rubber stamp's worth of
correspondence, and cheerfully demolishing proposals for computers that dodged thermodynamics. The
principle got its first direct experimental confirmation in 2012 — a single colloidal bead in an
optical double-well trap, erased slowly, dissipating right down to kT \ln 2
— half a century after the theory and, sadly, thirteen years after Landauer's death.
The most common misstatement of Landauer's principle — you will find it in textbooks that should
know better — is "every elementary computation costs at least kT \ln 2."
That is exactly what the principle does not say. The bound is charged per
erased bit, and only per erased bit. A computation arranged as a bijection erases
nothing and can, in principle, be pushed below any positive energy budget — a claim we will
make precise (and constructive) with Bennett's reversible machines later in the course. A billion
reversible gate operations followed by the erasure of three bits owes physics
3\,kT \ln 2, full stop. Two further fine-print clauses while we are
reading the contract: the bound is on heat dissipated, not "energy used" (energy merely
borrowed and returned counts for nothing), and it is reached only in the infinitely slow limit —
real erasure at real speed always pays a premium on top.
What this buys the rest of the course
Landauer's principle converts a philosophical itch into an engineering programme. Since only
information destruction is taxed, the obvious question is: can useful computation be done while
destroying nothing? The answer — yes, with ingenuity and some memory overhead — is the subject
of everything that follows: reversible gates, reversible architectures, and circuits that recycle
their energy instead of venting it. But first there is an old debt to settle. The
next
lesson takes Landauer's principle back to the Szilard engine and finally, after 115
years, pays off Maxwell's demon.