The Landauer Limit in Numbers

We have proved that erasing a bit costs at least Landauer's kT \ln 2 of heat. That is a beautiful sentence, but it is a symbol. This lesson turns the symbol into a number — a real, weighable quantity of joules — and then holds it up against the actual energy budgets of the machines around you: the chip in your laptop, the neurons in your head, the enzyme copying your DNA, the exotic superconducting demos in the lab. Only once you see where each of them sits on the ladder does the phrase "we are nowhere near the limit" acquire teeth.

The number itself

Plug room temperature into the formula. Boltzmann's constant is k = 1.381 \times 10^{-23}\ \mathrm{J/K}, \ln 2 = 0.693, and take T = 300\ \mathrm{K}:

kT \ln 2 \;=\; (1.381 \times 10^{-23})(300)(0.693) \;=\; 2.87 \times 10^{-21}\ \mathrm{J}.

That is 2.87 zeptojoules — a zeptojoule being 10^{-21}\ \mathrm{J}, one of the smallest energy prefixes anyone ever has cause to use. In the currency physicists prefer for single particles it is

2.87 \times 10^{-21}\ \mathrm{J} \;=\; \frac{2.87 \times 10^{-21}}{1.602 \times 10^{-19}}\ \mathrm{eV} \;\approx\; 0.018\ \mathrm{eV},

about \tfrac{1}{40} of the thermal energy kT \approx 0.026\ \mathrm{eV} that jiggles every atom at room temperature — indeed exactly \ln 2 \approx 0.69 of it, since kT\ln 2 = 0.693\,kT. To erase a gigabyte (8 \times 10^{9} bits) even perfectly costs 8 \times 10^{9} \times 2.87 \times 10^{-21} \approx 2.3 \times 10^{-11}\ \mathrm{J} — 23 picojoules, less than a mosquito's wingbeat. The floor is fabulously low. That, not its height, is the surprise: real hardware misses it by a factor you are about to meet.

The energy ladder

Here is the landscape of "joules to flip one bit", plotted on a logarithmic axis because it spans fifteen powers of ten — no linear axis could hold it. Each rung is roughly where that technology dissipates per elementary information-changing event.

Read the vertical gaps as multiplications. A 1990s microprocessor spent about 100\ \mathrm{pJ} = 10^{-10}\ \mathrm{J} per operation — a hundred billion Landauer quanta. Today's CMOS gate switch has fallen to roughly 10^{-15}–10^{-16}\ \mathrm{J}, still some 10^{4}–10^{5} above the floor. Biology, astonishingly, runs closer: a synaptic event costs a neuron on the order of 10\ \mathrm{fJ}, while a single base added by DNA polymerase dissipates only about 20–100\,kT — within a factor of a hundred of Landauer. The lowest experimental rungs of all are reversible superconducting logic demonstrations, which have flipped bits dissipating just tens of kT.

How many erasures could a laptop sustain?

Numbers this small invite a vertigo-inducing thought experiment. Suppose you had a 100\ \mathrm{W} laptop that spent every joule at the Landauer floor — no engineering waste at all, just the irreducible cost of forgetting. How many bits could it erase per second? Power over energy-per-erasure:

\frac{100\ \mathrm{W}}{2.87 \times 10^{-21}\ \mathrm{J/bit}} \;\approx\; 3.5 \times 10^{22}\ \text{erasures per second}.

Thirty-five sextillion. For comparison, a real high-end chip performs on the order of 10^{18} transistor switches per second — around ten thousand times fewer events, each of them 10^{4}–10^{5} times more expensive. Multiply those two shortfalls together and you recover the whole gap: today's silicon dissipates something like 10^{8}–10^{9} times more heat than a Landauer-limited machine of the same throughput would. The thermodynamic floor is not the ceiling we keep bumping into — the ceiling is engineering, and it is a very long way above the floor.

Does the cold help? Yes — and no

The bound is linear in temperature. Cool the environment from 300\ \mathrm{K} to liquid-helium 4\ \mathrm{K} and the cost per erased bit divides by 300/4 = 75:

kT\ln 2\big|_{4\,\mathrm{K}} \;=\; \frac{2.87 \times 10^{-21}}{75} \;\approx\; 3.8 \times 10^{-23}\ \mathrm{J}.

A tempting free lunch — until you ask who pays for the fridge. To dump a joule of waste heat out of a 4\ \mathrm{K} cold stage into a 300\ \mathrm{K} room, the second law charges a Carnot refrigeration surcharge of at least

\frac{T_{\text{hot}} - T_{\text{cold}}}{T_{\text{cold}}} \;=\; \frac{300 - 4}{4} \;=\; 74\ \text{joules of room-temperature work per joule pumped}.

So the erasure itself got 75\times cheaper, but every one of those cheaper joules now needs {\sim}74\times its own energy just to be carried upstairs to the warm world. The two factors very nearly cancel: measured in real, room-temperature electricity from the wall, computing cold buys you almost nothing on the erasure account. (Cold helps for other, genuine reasons — lower leakage, superconductivity, less thermal noise — but "cheaper Landauer erasures" is essentially a mirage once the compressor is on the bill.)

A single green photon carries about 4 \times 10^{-19}\ \mathrm{J} — so the Landauer quantum is roughly one hundred and forty times smaller than the energy of one particle of visible light. Put differently: catch a single photon and you have, in one grab, more than enough energy to erase a hundred bits at the thermodynamic floor. Or scale up: a housefly doing a single push-up expends around a microjoule, which at 2.87\ \mathrm{zJ} apiece would pay to erase some 3 \times 10^{14} bits — forty thousand gigabytes. The reason your computer is warm has nothing to do with this floor; it is warm because it insists on paying 10^{5}-fold over the odds for each of its bits.

Because the bound falls linearly with temperature, people reach for the obvious dodge: run the computer in a deep cryostat and the per-bit cost plummets. The arithmetic of the bound is correct — but it is the wrong ledger. The heat you deposit at low temperature does not vanish; it must be pumped back up to room temperature, and the Carnot cost of that pumping grows precisely as (T_{\text{hot}}-T_{\text{cold}})/T_{\text{cold}}, which climbs as fast as the erasure cost falls. Account for the whole system — computer plus its refrigerator, measured at the wall — and the net saving on erasure is negligible. The honest lesson is the same one Landauer keeps teaching: you cannot cheat the second law by changing where you draw the boundary. Real progress comes from erasing less (reversibility), not from making each erasure colder.