IT之家 9 月 30 日消息，美国 BlueQubit 研究团队利用 IBM 最新的 Nighthawk r2 量子处理器完成了一项随机量子线路采样实验，仅用 19 秒就生成了 100 万个样本。BlueQubit 研究人员估计，用 Frontier 超级计算机复现同等任务约需 110 年。
该实验由美国量子计算软件与云平台公司 BlueQubit 理论凝聚态物理学家 Tigran Sedrakyan 领衔。
Nighthawk r2 是一款具备 120 个量子比特的超导量子处理器，可借助 IBM 云平台访问。研究团队选取其中 61 个量子比特，运行复杂度逐步增加的随机量子线路，最多执行 40 个周期。最终，他们在 36 个周期、918 个双量子比特门的条件下找到实验所需的平衡点。
在这一规格下，Nighthawk r2 仅用了 19 秒就完成 100 万次采样。这里的关键并不是量子处理器可以快速生成大量 0 和 1，而是这些结果需要符合特定量子线路产生的概率分布，这正是经典计算机难以复现的部分。
随机量子线路采样（RCS）是量子计算领域用于测试量子处理器能力的一项基准任务。研究人员让量子比特按照随机挑选的操作开展多轮演化并产生纠缠，随后反复测量，得到大量由 0 和 1 构成的结果字符串。随着量子比特数量和操作轮数增加，经典计算机计算相应概率分布的难度会迅速提高。
为了估算经典计算机完成同一任务所需的计算量，研究团队使用了张量网络收缩方法。计算结果表明，复现这 100 万组样本大约需要 1.2 × 10²⁷ 次计算操作。
研究人员随后以 Frontier 超级计算机作为参考开展估算。Frontier 峰值表现超过每秒 10¹⁸ 次运算，依据研究团队使用的持续表现估算，完成同等任务约需 110 年。当然，这里的 110 年是基于特定经典模拟算法得出的估算值，并非经典计算机完成该任务所需时间的理论上限。
为了验证量子处理器产生的结果，研究人员还使用了两种交叉验证方法。第一种方法是构建经过“裁剪”的量子线路，将 61 个量子比特拆分为三个或四个较小平台，使经典计算机可以精确模拟部分线路，再以此评估完整实验的结果。
第二种方法则使用镜像线路，让量子处理器执行一组操作后立即执行完全相反的操作。如果量子处理器理想运行，最终状态应回到初始状态；偏离程度可以反映实验过程中累积的噪声和错误。两种方法随着线路深度增加仍给出了较为接近的结果。
在 36 个周期的实验条件下，剩余量子信号的保真度约为 0.23%。这一数值不能直接理解为传统意义上的“正确率”，因为随机量子线路采样过程中，大部分量子信息会受到噪声影响而逐步消失，实验需要确认的是理想量子概率分布是否仍留下可测量的痕迹。
Nighthawk r2 的硬件外观也协助缩短了实验时间。与 IBM 上一代 Heron 处理器相比，该处理器使用专用复位机制，可以主动释放量子比特中的能量，将连续运行实验之间的等待时间缩短至最低约 1 微秒。
IBM 表示，Nighthawk r2 每秒可执行超过 10 万个量子线路，相比之下 Heron 约为每秒 4000 个。这使得需要大量重复执行的随机量子线路采样实验可以更快完成。
不过，研究团队并未宣称量子计算机已经让经典超级计算机失去意义。“110 年”这一数字建立在特定的张量网络模拟方法和相关假设之上，更高质的算法、重复利用中间计算结果以及其他近似方法，都可能显著降低经典计算所需的时间。
类似情况此前已经浮现在谷歌 2019 年的 Sycamore 实验中。当时研究人员估算，经典超级计算机完成对应任务需要约 1 万年，但随后经典计算领域浮现了更高质的模拟方法，大幅缩短了这一时间。因而，量子长处并不是一个固定不变的指标，而是随着量子硬件和经典算法的发展不断变化。
此次实验的意义在于，相关量子长处测试并非依赖仅存在于实验室中的专用设备，而是在研究人员可以借助云平台访问的商业量子处理器上完成。研究团队还公开了实验所使用的量子线路、样本和分析代码，供其他研究人员复现和验证。
需要注意的是，随机量子线路采样本身并不是面向实际生产应用的计算任务，而是一项用于测试量子计算机能力的基准实验。
相关研究当下以预印本形式发表于 arXiv，尚未经过同行评审。IT之家附论文地址：
https://doi.org/10.48550/arXiv.2609.28657
本文来源：https://www.ithome.com/1/008/789.htm
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Thanks to IT Home netizen "Unique Experience" for the tip!
On September 30, IT Home reported that a US-based BlueQubit research team completed a random quantum circuit sampling experiment using IBM's latest Nighthawk r2 quantum processor, generating one million samples in just 19 seconds. BlueQubit researchers estimate that reproducing the same task on the Frontier supercomputer would take approximately 110 years.
The experiment was led by Tigran Sedrakyan, a theoretical condensed matter physicist at US quantum computing software and cloud platform company BlueQubit.
The Nighthawk r2 is a superconducting quantum processor with 120 qubits, accessible via IBM's cloud platform. The research team selected 61 qubits from it and ran random quantum circuits with increasing complexity, executing up to 40 cycles. Ultimately, they found the optimal balance required for the experiment at 36 cycles with 918 two-qubit gates.
Under these specifications, the Nighthawk r2 completed one million samplings in just 19 seconds. The key here is not that the quantum processor can quickly generate large amounts of 0s and 1s, but that these results must conform to the probability distribution produced by specific quantum circuits, which is precisely the part that is difficult for classical computers to replicate.
Random Circuit Sampling (RCS) is a benchmark task used in the field of quantum computing to test the capabilities of quantum processors. Researchers allow qubits to undergo multiple rounds of evolution and generate entanglement according to randomly selected operations, then repeatedly measure them to obtain large quantities of result strings composed of 0s and 1s. As the number of qubits and the number of operation rounds increase, the difficulty for classical computers to calculate the corresponding probability distribution rises rapidly.
To estimate the computational load required for classical computers to complete the same task, the research team used the tensor network contraction method. Calculation results indicated that reproducing these one million sets of samples would require approximately 1.2 × 10²⁷ computational operations.
The researchers subsequently used the Frontier supercomputer as a reference for their estimation. Frontier's peak performance exceeds 10¹⁸ operations per second. Based on the sustained performance estimates used by the research team, completing an equivalent task would take approximately 110 years. Of course, this 110-year figure is an estimate based on specific classical simulation algorithms and is not the theoretical upper limit of time required for classical computers to complete the task.
To verify the results generated by the quantum processor, researchers also employed two cross-validation methods. The first method involved constructing "pruned" quantum circuits, splitting the 61 qubits into three or four smaller subsystems so that classical computers could accurately simulate partial circuits, and then using this to evaluate the results of the complete experiment.
The second method used mirror circuits, where the quantum processor executes a set of operations followed immediately by the exact opposite operations. If the quantum processor operates ideally, the final state should return to the initial state; the degree of deviation can reflect the noise and errors accumulated during the experiment. Both methods yielded relatively close results even as circuit depth increased.
Under the experimental conditions of 36 cycles, the fidelity of the remaining quantum signal was approximately 0.23%. This figure cannot be directly understood as "accuracy rate" in the traditional sense, because during random quantum circuit sampling, most quantum information gradually disappears due to noise. What the experiment needs to confirm is whether measurable traces of the ideal quantum probability distribution still remain.
The hardware design of the Nighthawk r2 also helped shorten the experiment time. Compared to IBM's previous-generation Heron processor, this processor uses a dedicated reset mechanism to actively release energy from qubits, reducing the wait time between continuous experiments to a minimum of approximately 1 microsecond.
IBM states that the Nighthawk r2 can execute more than 100,000 quantum circuits per second, compared to approximately 4,000 per second for the Heron. This allows random quantum circuit sampling experiments that require a large number of repetitions to be completed faster.
However, the research team did not claim that quantum computers have already rendered classical supercomputers meaningless. The "110 years" figure is built upon specific tensor network simulation methods and related assumptions. Higher-quality algorithms, reuse of intermediate calculation results, and other approximation methods could all significantly reduce the time required for classical computation.
A similar situation has previously been seen in Google's 2019 Sycamore experiment. At that time, researchers estimated that classical supercomputers would need approximately 10,000 years to complete the corresponding task, but the classical computing field subsequently produced higher-quality simulation methods, which greatly shortened this time. Therefore, quantum advantage is not a fixed metric but rather changes continuously with the development of quantum hardware and classical algorithms.
The significance of this experiment lies in the fact that the relevant quantum advantage tests were not conducted on specialized equipment existing only in laboratories, but were completed on commercial quantum processors accessible to researchers via a cloud platform. The research team also published the quantum circuits, samples, and analysis code used in the experiment for other researchers to reproduce and verify.
It should be noted that random quantum circuit sampling itself is not a computational task aimed at practical production applications, but rather a benchmark experiment used to test the capabilities of quantum computers.
The related research is currently published as a preprint on arXiv and has not yet undergone peer review. IT Home attaches the paper address: https://doi.org/10.48550/arXiv.2609.28657