Chinese quantum computer's 'quantum advantage' claim undercut as single GPU server matches its scores

Standard benchmarks show classical emulation performs on par with — or better than — the actual quantum machine

A model of a quantum computer. This image is not directly related to the article. [Getty Images Bank]

Claims of "quantum advantage" — the ability of a quantum computer to solve problems that classical machines effectively cannot — have come in rapid succession in recent years. China's photonic quantum computer Jiuzhang, developed at the University of Science and Technology of China, is one such system.

Now, research has found that emulating Jiuzhang's calculations on a single GPU server produced results that matched or outperformed the actual quantum computer on standard benchmarks.

A research team led by Tom Dodd and Raul Garcia-Patron at the University of Edinburgh published their findings in volume 12 of the international journal Science Advances, reporting the development of an algorithm that reproduces Jiuzhang's computational outputs using a conventional computer.

Guessing which exit the light takes

The problem Jiuzhang solves is known as Gaussian boson sampling. The process involves sending photons through an optical circuit of intricately arranged mirrors and lenses, then recording which of 144 exits the light emerges from.

As the number of exits through which photons can emerge grows, the possible combinations increase explosively — making the calculation theoretically impossible for a classical computer under ideal, error-free conditions.

In practice, however, real experiments are difficult to run without errors, such as light leaking out as it passes through the circuit.

The research team argued that because such errors simplify the range of possible outcomes, a conventional computer should be capable of emulating the results well enough.

An illustration of Gaussian boson sampling. [AI-generated image]

Each time light is fired, consistent patterns emerge in which of the 144 exits simultaneously detect it. For example, when light emerges from exit 1, it tends to also emerge from exit 7 at the same time.

The research team calculated these correlations only up to cases where light simultaneously emerged from at most five exits. Correlations involving six or more exits simultaneously were deemed negligible and excluded. Earlier Jiuzhang experiments had shown these higher-order correlations to be statistically insignificant.

Using this approach, the team generated about 10 million sample outputs from Jiuzhang experiments in just 115 minutes on a single server equipped with eight of Nvidia's latest GPUs. The memory required was just 2 gigabytes — about one-thousandth of what earlier research had needed.

Closer to the 'right answer' than the actual quantum computer

The research team compared their algorithm's outputs against the experimental results from Jiuzhang 2.0 in 2021 and Jiuzhang 3.0 in 2023.

Whether a quantum computer has calculated correctly is assessed by how closely its results match the theoretically expected distribution. The team scored both Jiuzhang's actual results and their algorithm's results using the same method.

The algorithm matched or outscored Jiuzhang across the board. The same held true even under more demanding conditions involving higher photon counts.

"This improvement means that what was previously possible only at a handful of facilities worldwide can now be done by anyone," the team wrote in the paper.

This image is not directly related to the article. [Herald DB]

The team said the algorithm can handle increasingly complex experiments and could therefore be used to challenge results from larger quantum computers in the future. They added that it could also be applied to emulating some calculations performed by qubit-based quantum computers, not just photonic ones.

Speed, however, remains firmly in quantum computing's favor. Jiuzhang produces about 10 million samples in roughly 10 seconds.

According to the research team, the comparison measures only how closely each system approximates the correct answer indirectly — a direct experimental comparison is not possible.

The team acknowledged that they were unable to emulate the latest Jiuzhang 4.0 using their current method, adding that doing so would require a new approach involving multiple servers working in tandem.

Reference

DOI: 10.1126/sciadv.aef8431

Tom Dodd et al., "A fast and frugal Gaussian boson sampling emulator." Sci. Adv. 12, eaef8431 (2026).

dbsdn1110@heraldcorp.com