Three studies in China show that classical computers will completely disintegrate Google's quantum hegemony
An important milestone in the nisq era is the Google 53 qubit "sycamore" quantum processor, which can perform random circuit sampling tasks in 200 seconds, and the same task is estimated to run on Summit for 10000 years. Two recent experiments on Zuchongzhi 2.0 (56 qubits) and Zuchongzhi 2.1 (60 qubits) quantum processors have set this record.
On the other hand, the classical simulation algorithm and underlying hardware are also improved. For example, in 2020, Alibaba quantum computing team used tensor network method to estimate [1], and it only took 20 days to simulate 53 qubits and 20 layer loop circuit sampling of "sycamore" on summit supercomputer. Recently, Zu Chongzhi's 2.0 paper estimated that it would take 16 days to use summit to carry out classical simulation sampling for the "sycamore" experiment, while it would take 8 years to carry out classical simulation sampling for the "Zu Chongzhi 2.0" experiment.
In an arXiv paper in March 2021, Zhang pan of the Chinese Academy of Sciences and his students proposed a big head tensor network method [2], using a small computing cluster composed of 60 NVIDIA GPUs to complete Google's quantum hegemony experiment in 5 days. The linear cross entropy benchmark fidelity (fxeb) is 0.739, which is much higher than Google's result.
If, according to John Preskill's earliest definition, the quantum hegemony can only be realized when the computing speed of quantum computer reaches more than 1010 times that of supercomputer, then Google's quantum hegemony has actually been broken long ago. However, in terms of absolute time, no study has proved that the classical computer can exceed 200 seconds of Suzuki. Three recent studies in China show that there is no obstacle to reducing the sampling time of classical computer simulation Google random circuit to less than 200 seconds.
In an arXiv paper [3] on October 27, the National Supercomputing Center (Wuxi) developed a tensor based high-performance random quantum circuit simulator on the new Shenwei supercomputer, which shortened the simulation sampling time of Google's "sycamore" from 10000 years previously claimed to 304 seconds.
Another study by the same team [4] shows that it takes a week to perform the sampling task on Shenwei supercomputer based on the custom tensor network contraction algorithm, and uses the new generation Shenwei to redefine the baseline of "quantum hegemony".
Figure 1 simulates the total running time (circled green dotted line) of the same sampling task (generating 1 million bit strings with the same xeb fidelity as the quantum processor) on the circuits of (a) sycamore, (b) Zuchongzhi 2.0 and (c) Zuchongzhi 2.1.
In these two studies and the study of the Chinese Academy of Sciences in March this year, relevant samples were obtained. However, if the goal of the simulation is not only to pass the xeb test, but also to meet the constraints of obtaining irrelevant samples, as in the "sycamore" experiment, the tensor network needs to shrink thousands of times, making the calculation cost unbearable in practice.
Therefore, Zhang pan of the Chinese Academy of Sciences led his team to propose a new method to solve this problem, which only needs to shrink the corresponding tensor network once, and is much more efficient than the existing methods in obtaining a large number of irrelevant samples with target fidelity. For the "sycamore" quantum circuit with 53 qubits and 20 layer loops, they spent about 15 hours on a computing cluster with 512 GPU. The authors estimate that if their algorithm can be efficiently implemented on a modern supercomputer with 10 billion times of performance, the simulation will ideally take tens of seconds, which is faster than Google's quantum hardware.
Link：
[1]https://arxiv.org/abs/2005.06787
[2]https://arxiv.org/abs/2103.03074
[3]https://arxiv.org/abs/2110.14502
[4]https://arxiv.org/abs/2111.01066
[5]https://arxiv.org/abs/2111.03011
