19 SECONDS VERSUS A CENTURY OF SUPERCOMPUTING

Quantum advantage means showing that a quantum computer can do a task no classical supercomputer can finish in reasonable time. The standard test is random-circuit sampling: apply a sequence of randomly chosen quantum gates to dozens of qubits, then measure them. The result is a stream of strings of 0s and 1s drawn from a probability distribution that classical computers struggle enormously to reproduce.

An important point: this task has no practical use. It is a benchmark — a stopwatch for raw computing power.

A race on both sides

The Sycamore processor opened the race with 53 qubits, followed by Zuchongzhi 2.0 and 2.1, Sycamore at 67 qubits, and Zuchongzhi 3.0 with 83 active qubits. On the other side, classical algorithms kept improving and cut the cost of reproducing these experiments by orders of magnitude. So every claim has to say exactly which classical task it is being compared with.

Until now, all these demonstrations ran on dedicated laboratory machines, carefully tuned for the occasion.

A processor rented through the cloud

A team from the startup BlueQubit (San Francisco), with colleagues at EPFL and the XPRIZE Foundation, used an IBM Nighthawk r2 processor: 120 superconducting qubits on a 12 × 10 square grid. They accessed it through the cloud, with IBM’s standard Qiskit tools, and without any special calibration. The authors are not from IBM: they used the machine through IBM’s startup programme and state that their views are not IBM’s.

They used 61 qubits — three were left out because of their calibration data — linked by 102 couplers.

Checking the result directly is impossible: it is exactly what classical computers cannot do. So the team used two independent estimates of the circuit’s fidelity:

a “mirror” test: run a circuit, then its exact inverse, and check whether the qubits return to where they started — no simulation needed;
a “patch” test: cut the circuit into 3 or 4 pieces small enough to simulate, and check each piece.

The numbers

The two estimates agree at every depth measured.
Fidelity per cycle: 0.872, against 0.836 on the first-generation Nighthawk r1 — more than ten times better after 32 to 40 cycles.
The reference point: 36 cycles, 918 two-qubit gates, a fidelity of about 0.0023. One million samples were collected in 19 seconds.
Computing just one output probability classically costs about 10²² operations — about two days of the Frontier supercomputer.
Producing the same million samples classically: about 1.2 × 10²⁷ operations, or around 110 years on Frontier.
The whole experiment used about 11 minutes of quantum computer time.
Known classical “cheats” — cutting the circuit into pieces to mimic the score — fall at least ten times short of the measured score.

The fine print

The 110 years is an estimate for one specific classical algorithm, not an absolute limit. It assumes unlimited memory, which favours the classical side. Other methods — reusing work across many probabilities, approximate simulations — could lower the cost. And, the authors write, the history of this benchmark says some of them will.

A benchmark anyone can rerun

According to the authors, this is the first demonstration of quantum advantage in standard random-circuit sampling on a commercial, widely accessible processor that most non-expert users can reproduce. The circuits, the data — including the million samples — and the code are public.