Symptom

Your datacentre’s power budget is dominated by two numbers: the energy to switch transistors, and the energy to remove the resulting heat. Every generation of process shrink has reduced the first, and the projections say it keeps falling.

So you ask the natural question: how far can it fall? Is there a floor, or is this an engineering race with no finish line?

There is a floor, it is a consequence of the second law of thermodynamics rather than of any manufacturing limit, and it is not where most people put it. It does not attach to computing. It attaches to forgetting.

Statement

Landauer’s principle (1961). Erasing one bit of information in an environment at temperature $T$ requires the dissipation of at least

$$E \ge k_B T \ln 2$$

joules of energy as heat, where $k_B = 1.38 \times 10^{-23}$ J/K.

More generally, any logically irreversible operation — one whose output does not determine its input — that reduces the number of possible logical states by a factor $2^n$ dissipates at least $n k_B T \ln 2$.

At room temperature, $T = 300$ K:

$$k_B T \ln 2 = 1.38\times10^{-23} \times 300 \times 0.693 \approx 2.9 \times 10^{-21}\ \text{J}.$$

That is about 0.018 eV per bit, or 2.9 zeptojoules. A current CMOS switching event costs on the order of $10^{-15}$ J, so production hardware sits roughly six orders of magnitude above the limit. The bound is real and it is not close.

The physical-law tag is the honest one: this is not a theorem about computation but a consequence of statistical mechanics applied to physical devices, and its status depends on physics rather than on proof.

Argument

The phase-space argument, which is Landauer’s own. A physical bit is a device with two distinguishable macrostates, 0 and 1, each corresponding to some volume of microstates in phase space. Before erasure, the device may be in either, so the accessible volume is $V_0 + V_1$. After erasure — “set to 0 regardless of prior state” — only $V_0$ is accessible.

The volume has halved, so the entropy of the device has fallen by $k_B \ln 2$. The second law forbids the total entropy of device plus environment from decreasing, so the environment’s entropy must rise by at least $k_B \ln 2$. Heat $Q$ delivered to an environment at temperature $T$ raises its entropy by $Q/T$. Therefore $Q \ge k_B T \ln 2$. $\blacksquare$

Note what the argument used: that two physically distinct states were mapped to one. It did not use anything about logic gates, semiconductors, or computation. Any many-to-one map on physical states pays.

Szilard’s engine and why the bookkeeping had to be fixed. Maxwell’s demon appears to violate the second law: a demon that observes molecules and sorts fast from slow extracts work from a single heat bath. Szilard’s 1929 single-molecule version makes it quantitative — measure which half of a box the molecule is in, insert a piston on that side, extract exactly $k_B T \ln 2$ of work as it expands.

For decades the resolution was assumed to be that measurement must cost energy. Bennett showed in 1982 that this is wrong: measurement can in principle be done reversibly. The cost is in resetting the demon’s memory for the next cycle. Each cycle extracts $k_B T \ln 2$ and must erase one bit, costing $k_B T \ln 2$. The engine breaks exactly even, and the second law survives. The demon is defeated by its own garbage collection, which is the most satisfying result in this area and the reason the principle matters conceptually.

Reversible computing, which is the escape hatch. If the cost attaches only to irreversible operations, avoid them. AND is irreversible — output 0 does not determine the inputs. The Toffoli gate (CCNOT: flip the third bit iff the first two are 1) is reversible, universal for classical computation, and dissipates no $k_B T \ln 2$ in principle.

The price is garbage. A reversible circuit accumulates intermediate values it cannot discard. Bennett’s trick: compute forward, copy the answer out, then run the computation backwards to un-compute everything else, ending with only the input and the answer. The cost is a space-time trade-off, not an energy one. Adiabatic CMOS demonstrates the real version, recovering charge instead of dumping it, at the cost of running slowly — the dissipation of a real reversible process scales inversely with the time taken, so you buy energy with latency.

The experiments, which is why this is not just theory. Bérut et al. (2012, Nature) built a one-bit memory from a colloidal particle in a double-well optical trap, erased it, and measured the dissipated heat: it approached $k_B T \ln 2$ from above and did not cross it, with the excess falling as the erasure was performed more slowly. Jun, Gavrilov, and Bechhoefer (2014) confirmed it with better statistics. The principle has been measured.

Back to the datacentre. Six orders of magnitude of headroom means Landauer is not your constraint and will not be in your career. What the principle gives you is the shape of the answer: the floor is proportional to $T$, so cooling helps twice over; and the floor applies per erased bit, which is why the long-term research direction is not smaller transistors but circuits that discard less.

Forbids

Erasing a bit at zero energy cost at positive temperature. The second law forbids it, and the experiments have looked.

Maxwell’s demon as a source of free work. Any demon must store measurements, and any cycle must reset that store. The books balance exactly.

Unbounded computation in a finite energy budget when the computation is irreversible. $N$ irreversible bit operations need at least $N k_B T \ln 2$ joules. That is a hard bound on any irreversible architecture.

Beating the limit by making transistors smaller. The bound is thermodynamic and independent of device technology. Smaller devices close the gap to it; they do not move it.

Does not forbid

It does not make computation inherently dissipative, which is the misreading that gets the principle quoted as a fundamental limit on computing speed or efficiency. The cost attaches only to logically irreversible steps. Reversible computation has no Landauer floor at all, and Toffoli gates are universal, so there is no computation that must be performed irreversibly. The limit is on a design choice, not on computation.

It does not say the Bremermann or Bekenstein bounds follow from it. These are often listed together as “the physical limits of computation” and they have entirely different derivations. Bremermann’s $\sim 10^{50}$ operations per second per kilogram comes from mass-energy and the time-energy uncertainty relation. The Bekenstein bound on information in a region comes from black-hole thermodynamics. Neither is a consequence of Landauer, and treating the three as one result is a common and confusing error.

It does not describe where today’s energy goes. Real chips dissipate roughly $10^6$ times the Landauer limit per operation. Your power bill is charging capacitances, leakage current, clock distribution, and data movement. Optimising against Landauer would be optimising against the smallest term by a factor of a million.

It does not depend on any particular definition of information. The derivation uses phase-space volume. The connection to Shannon entropy is that both count distinguishable states in units of $\ln 2$, and this is the point at which information theory and thermodynamics turn out to be the same subject, but the physics does not need the information-theoretic language to go through.

It does not forbid erasure being cheap in practice. $2.9 \times 10^{-21}$ J is a small number. Erasing an exabyte costs about 23 microjoules at the limit. The principle is conceptually enormous and practically negligible.

Boundary

  • It is a lower bound approached only quasi-statically. Erasing quickly costs much more; the excess scales roughly as $1/\tau$ for erasure over time $\tau$. Any real process at real speed pays a large multiple.
  • Temperature must be well-defined. The derivation assumes a thermal environment at equilibrium temperature $T$. Non-equilibrium and strongly-coupled regimes need more care and the corrections are active research.
  • The status has been genuinely disputed. Earman, Norton, and others have argued the principle either assumes the second law or does not follow from it cleanly. The experimental confirmations are the strongest response, and the philosophical argument is not fully settled.
  • Quantum systems have their own version, with coherence and entanglement changing the accounting; the quantum Landauer bound is related but not identical.
  • Nothing here bounds speed. Landauer constrains energy per erasure, not operations per second. Those are different limits with different derivations.

The line to keep: information is physical. A bit is not an abstraction floating above the hardware; it is a configuration of matter, and destroying it is a physical act with a price set by thermodynamics.