跳过 Sycamore:利用 1432 个 GPU 实现量子随机电路抽样 7 倍加速
量子物理
2024-06-28 v1
摘要
随机量子电路抽样是检验量子计算优势的基准测试方法。近期基于张量网络方法的经典算法显著缩短了经典仿真时间,挑战了第一代量子优势实验的主张。然而就生成无关联样本、求解时间和能源消耗而言,之前的经典仿真实验仍不及 \textit{Sycamore} 处理器。本文报告了一种能耗高效的经典仿真算法,使用 1432 个 GPU 对量子随机电路抽样进行仿真,其生成的无关联样本具有更高的线性交叉熵得分,且比 \textit{Sycamore} 53 位量子比特实验快 7 倍。我们提出一种后处理算法以降低总体复杂度,并集成最新的高性能通用 GPU,实现比以往工作低两个数量级的能源消耗。我们的工作提供了对 \textit{Sycamore} 声称的量子优势的首个明确实验性证据,重新定义了基于随机电路抽样的量子计算优势边界。
引用
@article{arxiv.2406.18889,
title = {Leapfrogging Sycamore: Harnessing 1432 GPUs for 7$\times$ Faster Quantum Random Circuit Sampling},
author = {Xian-He Zhao and Han-Sen Zhong and Feng Pan and Zi-Han Chen and Rong Fu and Zhongling Su and Xiaotong Xie and Chaoxing Zhao and Pan Zhang and Wanli Ouyang and Chao-Yang Lu and Jian-Wei Pan and Ming-Cheng Chen},
journal= {arXiv preprint arXiv:2406.18889},
year = {2024}
}
备注
This work was completed on August 2023. A further 50x improvement has been achieved and will be posted on arXiv shortly