Of all the technologies chasing the Landauer floor, one has already crept within shouting distance of it — and it does so by throwing away the one thing every other computer relies on: resistance. Cool certain metals below a critical temperature and they become superconductors, carrying current forever with no voltage drop and no heat. A wire that never warms up is a tantalising place to try to compute without dissipating.

In this technology a bit is not a voltage held on a wire but a single quantum of magnetic flux, \Phi_0 = h/2e \approx 2.07 \times 10^{-15}\ \mathrm{Wb}, trapped in a loop of superconductor. Flux comes only in these indivisible lumps — you cannot have half a \Phi_0 — so the loop is a naturally digital device. Switching is done by Josephson junctions: thin breaks in the superconductor across which the flux quanta tunnel, one at a time, in picoseconds. This is the world of single-flux-quantum (SFQ) logic, and it runs at tens of gigahertz while sipping femtojoules.

The adiabatic quantum-flux-parametron

Raw SFQ logic is fast but still irreversible — it dumps a flux quantum's worth of energy every time it switches. The reversible frontier lives in a cousin: the adiabatic quantum-flux-parametron, or AQFP, developed by the Yokohama group of Yoshikawa, Takeuchi and colleagues. The trick is exactly the trick of a power-clock: instead of a fixed supply voltage, the gate is driven by an AC excitation current that gently shapes a potential landscape.

Picture the same double well from module 1. At the start of a clock cycle the AQFP's potential is a single flat basin — no bit yet. As the AC excitation rises, the basin bifurcates into a double well, and the flux settles left or right depending on a tiny input current present at the moment of splitting. The logic is encoded in which well the state falls into, and because the excitation rises and falls smoothly (adiabatically), almost no energy is lost. Measured switching energies land around 10\text{–}100\,kT per operation at 4\ \mathrm{K} — orders of magnitude below room-temperature CMOS, and the closest any working logic family has come to the kT\ln 2 wall. Working prototypes exist: the MANA microprocessor is an AQFP design demonstrated at GHz clock rates.

From adiabatic to genuinely reversible: RQFP

AQFP is adiabatic — low-loss — but not yet logically reversible: a two-input AND-like gate still merges states, and that merge carries the irreducible kT\ln 2. The reversible quantum-flux-parametron (RQFP) closes the gap. Its building block is a three-in, three-out majority gate: a bijective, information-preserving cell whose outputs are a reversible function of its inputs, so no state is ever destroyed inside the logic.

In simulation, RQFP gates dissipate below kT\ln 2 per logic operation — the first logic family to cross that line even on paper — and laboratory measurements have come near it. A related device, the nSQUID (used by Semenov and collaborators to build ballistic, reversible shift registers), moves flux along a chain with almost no loss by exploiting the symmetric coupling of two SQUID loops. These are proofs of principle, not products, but they establish something important: reversibility is not just a theorem, it is measurable in a real cryostat.

The honest accounting: you have to pay the fridge

Here is where enthusiasm meets thermodynamics. All of this runs at 4\ \mathrm{K}, and getting to 4\ \mathrm{K} costs energy — a lot of it. Removing one joule of heat from a cold stage and dumping it into a room-temperature world is governed by the Carnot limit. The ideal ratio is

and real cryocoolers are perhaps 3–10× worse than Carnot, so the practical overhead is roughly 300\text{–}1000\times. Every joule saved on the chip must be paid back several hundred times over at the wall socket. The comparison that matters is therefore system-level: the on-chip energy per operation must beat room-temperature CMOS by more than the refrigeration multiplier before the whole machine wins. AQFP's ~1000× on-chip advantage over CMOS is exactly in the range where this becomes a real, close contest — not a slam dunk.

So who should care? Two natural customers. First, exascale and datacentre-scale computing, where cooling is already the dominant cost and a bulk cryogenic plant amortises well. Second — and most compelling — the control electronics for quantum computers, which already live inside a dilution refrigerator. If you must be cold anyway, the fridge tax is already paid, and superconducting reversible logic becomes almost free by comparison.

How far below CMOS, really?

The chart puts the numbers on one log scale: energy per operation for room-temperature CMOS, for AQFP and RQFP at 4 K, and the two Landauer floors — kT\ln 2 evaluated at 300\ \mathrm{K} and at 4\ \mathrm{K}. Notice how low the 4 K floor sits: cold makes the logical floor tiny. But remember the previous card — that tiny floor is what the fridge is fighting to maintain.