主题演讲：进入量子实用时代
本页标题、字幕、摘要和阅读笔记为已保存的 AI 译文，可能有误。时间戳与来源一致，可展开查看原始字幕，并对照原视频核对。
IBM 的 2023 量子峰会介绍了 System Two、处理器里程碑和 Qiskit 工具链，概述了该公司实现量子实用化的方法。
文字记录来源：YouTube上创作者上传的字幕。Vidleaf添加了AI标点；保留原始措辞和时间戳锚点。原始文字记录可在下方获取。
AI生成的摘要
IBM 的 2023 量子峰会主题演讲展示了其量子计算的硬件、软件和系统路线图。演讲者介绍了模块化的 Quantum System Two，并讨论了 Condor 和 Heron 处理器的不同目标。Qiskit、执行模式和可复用的编程模式将硬件进展与构建和运行应用的工作联系起来。演讲还涵盖了错误抑制、缓解以及通往错误纠正的长期路径。性能数据和未来里程碑是 IBM 在活动当时的声明和计划，因此链接的章节提供了它们的原始背景和限定条件。
要点
- [9:50] Quantum System Two 被介绍为一种为持续扩展而设计的模块化架构。
- [27:02] Qiskit 连接了量子电路的创建、优化和执行。
- [39:06] 路线图讨论将关注点从量子比特数量扩展到有用电路的质量和规模。
根据文字记录生成。请对照原始视频核查所链接的段落。简要说明和要点由 Vidleaf 于 2026-10-21 审核。
完整概述
IBM Research 高级副总裁兼总监 Dario Gil 在 2023 量子峰会上开幕，将计算——从 AI 到半导体再到量子——视为再次驱动市场并催生行业，并指出 IBM 快速的商业化步伐，包括 watsonX。演讲追溯了 IBM 的量子里程碑：从 2016 年将处理器上云和 2017 年的 Qiskit，到 IBM 量子网络、量子创新中心和 2019 年的 IBM Quantum System One，以及从 Falcon 到 Hummingbird、Eagle、Osprey 和现在拥有 1121 个量子比特的 Condor 的年度处理器节奏。IBM 在约克敦高地的 TJ Watson 研究中心推出了模块化的 IBM Quantum System Two，并描述了将 Watson 和生成式 AI 引入以使量子系统和 Qiskit 更易于使用，包括一个早期原型，可根据自然语言提示生成 Qiskit 代码。峰会宣布了一个效用新时代，得到了大约 2,500 篇已发表应用论文和超过 100 个量子比特及数百或数千个门的效用实验的支持，这些实验通过错误抑制和缓解方法实现，使能力在四年内提高了一千倍。随后 Rajeev 介绍了硬件：Condor 通过超过一英里的高密度柔性 IO 和与 Osprey 相似的相干性推动了单芯片扩展，而 Heron 是一款具有可调耦合器的 133 量子比特芯片，其门错误率是最好 Eagle 的一半，保真度达到三个九，并且几乎消除了串扰。Eagle 在过去几年中将分层门错误从约 10% 降低到 2 到 3% 的范围，而 Heron 在长链中每个量子比特的错误率提高了 3 倍，目标是在下一个版本中使 80 量子比特和 100 量子比特链看起来相同；Heron 更短的门时间为其提供了通往 5000 门及更远的清晰跑道，并且通过类似的错误缓解，比 Eagle 提高了 3 到 6 倍。IBM 宣布其首个客户 Heron 系统 IBM Torino，作为纽约数据中心的探索性设备提供，对于 100 个量子比特的长链错误率低于 0.8%，并宣布明年的 Qiskit 1.0 以及执行模式和 Qiskit 模式、一种新的总低密度奇偶校验码和新的耦合器，路线图现在跟踪每个电路的门数而非量子比特数量，并延伸到 2033。IBM Quantum System Two 被描述为模块化并设计为持续扩展，第一个系统已经包含三个 Heron；Heron 被称为高性能系统，据说已经使效用工作负载运行速度提高了五倍，而 Condor 被认为解决了单量子比特扩展问题。并行化和执行模式预计将使批处理和迭代工作负载运行得更快，IBM 简化了量子算法并引入了 Qiskit 模式以及 Quantum Serverless，现已进入测试阶段，并被描述为更加稳定。Qiskit 将成为 Qiskit 1.0，被描述为快速且稳定，IBM 正在引入 AI 以简化运行量子计算机，包括通过自然语言方式创建量子代码以及现已提供 Alpha 版的基于 AI 的转译器服务。IBM 表示其十年路线图以错误纠正为承诺，称所有里程碑均已实现，并宣布量子效用时代已经到来。
带时间戳的章节摘要
2023 量子峰会开幕致辞
- [0:17] Dario Gil 欢迎与会者参加 2023 量子峰会。
- [0:24] Dario Gil 自我介绍为 IBM 高级副总裁兼 IBM Research 总监。
- [0:38] 计算，从 AI 到半导体再到量子计算，再次驱动市场并催生行业。
- [0:56] Gil 认为这是计算领域最激动人心的时代，可能是自 1940 年代数字计算机或五十年代末晶体管以来最激动人心的时代。
- [1:23] 今年是 Gil 在 IBM 的 20 周年，他说公司甚至与五年前相比也非常不同。
- [1:49] Gil 说他从未见过 IBM 以比今天更快的速度将技术商业化。
- [2:06] 研究部门自称是公司的有机增长引擎，专注于将 R&D 创新带入产品。
- [2:44] IBM 在几个月内将 watsonX 商业化，这是一个将生成式 AI 带给企业和政府的平台。
IBM 量子里程碑与效用时代
- [3:12] 演讲者回忆起几十年前阅读核心量子信息科学思想并看到它们成为现实。
- [3:29] 2016 年，IBM 将第一个量子处理器上云，一周内成员数量超出预期五倍。
- [3:43] 这标志着量子计算变得可访问的时刻，最终覆盖近五十万人。
- [4:02] 2017 年，IBM 推出了 Qiskit，一个开源量子 SDK，成为最广泛使用的量子软件开发环境，受到超过 81% 的量子开发者和超过五十万用户的青睐。
- [4:22] 同年，IBM 启动了 IBM 量子网络，现在拥有超过 280 个成员。
- [4:35] 2018 年，IBM 建立了量子创新中心，现已发展到全球 38 个。
- [4:45] 2019 年，IBM 推出了 IBM Quantum System One，被描述为世界上第一台完全集成且商业可用的量子计算机，现在在德国、日本和加拿大，更多正在安装中。
- [5:12] IBM 称过去几年为量子计算涌现的时代，开发的量子计算能力比业内任何人都多，并将其交到更多人手中。
- [5:45] 进步得益于在硬件和软件上持续快速创新的理念。
- [5:54] IBM 每年至少推出一款量子处理器：2019 年的 Falcon（27 量子比特）、2020 年的 Hummingbird（65）、2021 年的 Eagle（突破 100）、去年的 Osprey（433），以及现在拥有 1121 量子比特的 Condor。
- [6:35] IBM 表示，通过 Condor 已经解决了量子比特扩展的问题，并正在将这一知识应用于更大的系统。
- [6:50] IBM 还专注于提升门质量和增加电路深度，并推出了 Heron，被称为世界上性能最高的量子处理器，采用可调耦合器技术。
IBM Quantum System Two 与 AI 集成
- [9:50] IBM 推出了 IBM Quantum System Two，一种模块化且灵活的架构。
- [9:54] System Two 采用主冷却，在纽约约克敦高地的 TJ Watson 研究中心运行 100+ 量子比特的问题。
- [10:09] 它是一个模块化系统，可以连接成更大、更强大的量子系统，被称为量子中心超级计算的构建模块。
- [10:26] 实用时代预计将巩固 IBM Quantum 作为科学和商业重要工具的地位。
- [10:49] IBM 正在引入 Watson 和生成式 AI，使量子系统和 Qiskit 更易于使用。
- [11:00] 一个早期原型，尚未完全准备好投入主流使用，允许用户通过自然语言提示生成 Qiskit 代码。
- [11:47] IBM 的 CEO 表示 IBM 是一家混合云和 AI 公司，而演讲者表示 IBM 将成为一家混合云、AI 和量子计算公司。
- [12:22] Jay Gambetta 被邀请上台分享进一步的工作和进展。
IBM 量子峰会：实用时代
- [12:53] IBM 的使命是将有用的量子计算带给世界，并使世界具有量子安全性。
- [13:19] 欢迎新的量子计算中心，包括 BasQ、RPI 和 RIKEN。
- [13:24] 欢迎新的量子创新中心，包括 NQCC、KQC 和哥本哈根大学。
- [13:36] 新的商业合作伙伴包括 KPMG、T-systems 和 SAP，新的行业客户包括 Banco Bradesco、Dow、Hyundai、Israel Aerospace Industries、Itau、Moderna、Mitsubishi Electronic、Pfizer 和 Truist。
- [14:08] 我们已经进入了一个新的量子计算时代，称为实用时代。
- [14:26] 社区已经发表了约 2.5 千篇关于量子计算应用的论文，例如模拟自然、结构化数据、搜索和优化。
- [14:55] 实用论文表明，量子可以解决超越经典暴力计算方法的问题，这就是量子实用的含义。
- [15:07] 实用被视为通往量子优势之路的第一个里程碑，这得益于新颖的错误抑制和错误缓解方法，在四年内将能力提升了一千倍。
- [15:34] 在实用时代，量子开始被视为真相的来源，经典方法仅用于验证结果。
- [15:54] 去年夏天至少还有六个实用实验运行了超过 100 个量子比特和数百甚至数千个门的量子电路。
- [16:14] 量子计算现在是材料科学、凝聚态物理和粒子物理等科学领域发现的工具。
- [16:45] 如果你不使用 100 个量子比特，你就不是在研究量子。
IBM 量子硬件路线图和 Condor
Condor 1121 量子比特芯片扩展
Heron 量子比特芯片性能
- [20:42] Heron 是一款 133 量子比特芯片，基于固定频率量子比特和可调耦合器构建，以实现灵活性。
- [20:57] 在 2019 年，团队回到绘图板探索新的门架构。
- [21:15] 第一款 Heron 芯片，代号 Monte-Carlo，显示出比最好的 Eagle 有显著改进。
- [21:24] Heron 的门错误率减半，一些保真度达到三个九，并且几乎消除了串扰。
- [21:39] Heron 在门时间上有显著改进，为用户在实用时代提供了优势，并为扩展模块化处理器奠定了基础。
- [22:04] 在量子比特间串扰测量中，最好的 Eagle 信号在多量子比特操作中随时间明显退化，而 Heron 几乎不受影响。
- [23:02] Eagle 开始时每层门的错误率约为 10%，在过去几年中降至 2 到 3% 的范围。
- [23:21] Heron 在长量子比特链中显示每个量子比特的错误率提高了 3 倍。
- [23:34] 蓝色和灰色点反映了80量子比特链与100量子比特链之间的差异，目标是在下一个版本中使两者看起来相同。
- [23:58] Heron更短的门时间允许运行更多的门，为实现5000个门及以上的目标提供了清晰的路径。
- [24:35] 使用类似的错误缓解技术，Heron相比实用论文中使用的Eagle显示出3到6倍的改进，并且这种改进在更长的实验中变得更加明显。
IBM Torino和Heron系统发布
Qiskit 1.0和执行模式
- [27:02] Qiskit被呈现为创建、优化和执行量子电路和算子的事实标准。
- [27:19] Qiskit的成功部分归功于其“一次编写，到处运行”的模式，支持大多数主要的量子硬件供应商和架构。
- [27:28] Qiskit 1.0宣布将于明年推出，在性能、稳定性和可靠性方面有更多改进。
- [28:03] 第一步将问题映射到量子电路和算子；Qiskit 1.0增加了对动态电路的原生支持，并将内存使用量减少多达55%。
- [28:32] 第二步优化输入；Qiskit转译器比一年前快16倍，并且与另一个知名工具包相比，可以生成少23%的双量子比特门。
- [29:12] 使用强化学习的AI转译器在电路深度和c-not计数方面比启发式算法显示出20到50%的改进，alpha版本今天可用。
- [29:45] 第三步在原语中执行优化后的电路；Qiskit 1.0简化了采样器和估计器接口，Qiskit运行时支持单作业、批处理和会话模式。
- [30:52] 批处理执行模式允许一次进行多个非迭代查询，并且通过线程并行化，可以将工作负载运行速度提高多达五倍。
- [31:28] 路线图承诺了线程化原语，而交付的是一套执行模式：单作业、批处理和会话。
量子算法的Qiskit模式
- [32:33] 实用时代引入了一类新用户，称为量子计算科学家。
- [32:39] 量子计算科学家对量子硬件本身不感兴趣，而是对使用机器解决特定的计算任务感兴趣。
- [32:49] 这些用户最看重性能、兼容性和易用性。
- [33:05] Qiskit模式是所有量子算法和应用必须遵循的四步序列。
- [33:11] 这四个步骤是：生成电路和算子，为执行优化它们，使用Qiskit原语执行，以及对原语输出进行后处理。
- [33:32] Qiskit模式为大规模编写算法和应用提供了逻辑框架，并且它们允许容器化和从量子电路和算子中抽象出来。
- [34:21] 在量子化学模式中，每个步骤由执行单一任务的块组成，块可以由IBM、第三方提供商或开源贡献者制作。
- [34:53] Qiskit模式旨在利用已有的软件框架，例如Scipy的优化器，并且只构建有针对性的量子加速所需的核心功能。
- [35:35] Qiskit模式设计为大规模使用，可以通过资源管理信息增强，上传到异构基础设施（如量子无服务器），并由对量子计算一无所知的最终用户无人值守运行。
- [36:12] 这些组件将从明年开始陆续推出，但量子无服务器已经处于测试阶段。
新的Gross纠错码
用于量子纠错的新耦合器
新的量子路线图聚焦于门质量
- [38:30] 需要为面向客户的系统和服务制定更大的路线图。
- [39:06] 路线图现在展示的是未来五年内处理器在单个电路中可以运行的门数量，而不是量子比特数量。
- [39:20] 未来五年，重点将放在质量而非规模上。
- [39:25] 他们相信单芯片扩展已通过Condor解决，因此更大的挑战是实用工具和提升质量。
- [39:48] 未来五年，他们计划将处理器可运行的门数量增加两倍。
- [39:53] Flamingo包含量子通信，并将连接至少七个QPU，以创建一个拥有数千量子比特的系统。
- [40:08] 在2029年，Starling预计将带来跃升至100百万个门，以及首个具有纠错功能的系统。
- [40:49] 路线图延伸至2033年，并详细计划扩展到1十亿个门。
IBM量子创新路线图
用户工作负载与量子中心超级计算
IBM量子路线图回顾
Qiskit 1.0 与量子计算中的人工智能
量子计算路线图与里程碑
[0:06] 早上好。让我们花一分钟邀请所有朋友和同事就座，以便我们开始。但首先，欢迎来到2023量子峰会。我们非常高兴您今天能加入我们。我叫Dario Gil。我是IBM的高级副总裁兼IBM Research总监。哇，从技术的角度来看，2023年真是非凡的一年。
[0:38] 看，从人工智能到半导体再到量子计算，计算再次驱动市场、催生行业，并真正点燃了全世界人们的热情。以及我们对技术将能做什么的想象力。我真的相信，我们正生活在计算领域最激动人心的时代，可能是自1940年代数字计算机或五十年代末晶体管问世以来最激动人心的时代。
[1:10] 而如今，凭借这些新技术，我们真正开始有机会着手解决那些我们甚至认为上一代技术不可能解决的问题。我很自豪地说，今年是我在IBM的20周年，谢谢大家，我希望你们都能看到，这家公司与仅仅五年前相比已经大不相同。这在很大程度上归功于我们在IBM内部R&D（研发）社区所开展的基础性工作。
[1:49] 坦率地说，还有我们将技术商业化的速度和节奏。看，在我于IBM的20年里，我从未见过我们以比今天更快的速度将技术商业化。我们喜欢说，在研究部门内，我们是公司的有机增长引擎，我们真正不懈地专注于如何将我们在R&D（研发）社区中开发的创新带入产品。
[2:21] 我要举一个例子来突出这种速度和节奏，那就是我们今年在人工智能方面所做的工作。我的意思是，我们都知道，在人工智能领域，就公众而言，这项技术所能实现的可能性无疑已经进入了公众的想象。多年来，我们在研究部门内部对基础模型的作用产生了深厚的热情。但今年，我们以不同于以往任何时期的强度将其商业化，并创建了一个新的平台，将生成式人工智能带入企业和政府领域，名为 watsonX。
[2:56] 我们能够在短短几个月内做到这一点。回到这个主题，即以前所未有的速度将技术商业化，人们确实注意到了。但当然还有更多。还有量子领域。看，我记得读过，你知道，几十年前关于量子信息科学的一些核心思想，并思考这些思想将会变得多么强大。如果你现在看那段历程，它们正在成为现实。这段历程令人惊叹。
[3:29] 回想一下，2016 年我们在量子行业背景下制造了新闻，当时我们将第一个量子处理器放上云端，一周之内加入的成员和用户数量就达到了我们预期的五倍。那确实标志着量子计算向世界开放的时刻。它首先向数百人展示，然后向数千人展示，最终向近五十万人展示，量子计算是值得考虑、拥抱和培育的事物。
[4:02] 现在，在 2017 年，我们推出了 Qiskit，我们的开源量子 SDK，它已成为世界上使用最广泛的量子软件开发环境。事实上，让我给你们一个统计数据。超过 81% 的量子开发者更喜欢它，用户也超过五十万。同年我们启动了 IBM Quantum Network，它爆发式增长。我们今天有超过 280 名成员，你们中的许多人今天就在这里。
[4:32] 所以我感谢你们。在 2018 年，我们还建立了量子创新中心，现在已在全球发展到 38 个。在 2019 年，我们推出了 IBM Quantum System One。它是世界上第一台完全集成且可商用的量子计算机。从那时起，我们一直在世界各地部署 Quantum System One。现在德国、日本和加拿大都有 System One。
[5:05] 更多正在韩国、西班牙、日本以及美国这里安装。所有这些活动确实非常奇妙且令人震惊，这就是为什么我们把这些过去几年称为量子计算涌现的时代。在此期间，我们开发了更多的量子计算能力，并将其交到比业内任何人都多的人手中。现在，看到我们共同建立的社区能够解决的问题，确实非常了不起。
[5:43] 这确实并不容易。但使之成为可能的是我们在硬件和软件方面持续快速创新的理念。在硬件方面，我们每年至少推出一款具有标志性突破的量子处理器。在 2019 年，我们推出了 Falcon，一款 27 量子比特处理器。一年后，Hummingbird 拥有 65 量子比特。而在 2021 年，我们用 Eagle 突破了 100 量子比特的屏障。
[6:19] 就在去年，我们推出了 Osprey，拥有 433 量子比特。而今天，正如所承诺的，我们宣布 Condor，拥有高达 1121 量子比特。借助 Condor，我们解决了扩展问题，即量子比特扩展的问题。现在我们正在运用这些知识来构建更大的系统。当然，我们并没有止步于此。
[6:50] 我们还致力于提高门质量和增加电路深度。这就是为什么尽管我们对 Condor 感到兴奋，但我们对 Heron 更加兴奋。这就是 Heron，世界上性能最高的量子处理器，也是真正能够扩展量子计算的处理器。Condor 消除了扩展的障碍，而 Heron 通过其可调耦合器技术提升了电路深度和质量，我今天非常自豪地介绍这项技术。
[7:29] 哇，如果你看看硬件、软件和生态系统方面的集体成就，让我给你们一个惊人的统计数据。这些是在 IBM Quantum 的机群上执行的电路数量，超过 3 万亿个。所以如果我们回顾过去几年所取得的成就，那确实是非凡的。稍后我们将庆祝，我们会说那个阶段标志着量子计算的涌现，而现在正是迈向一个新时代的时候，我们谦逊地称之为实用时代。
[8:10] 实用是本次量子峰会的主题。我们认为有一项关键进展标志着向这个激动人心的新时代的过渡。六月，团队发表了在扩展量子计算方面的突破，登上了《自然》杂志的封面。论文的标题是《在完全容错之前实用性的证据》，在其中，我们的团队展示了我们现在可以运行超过 100 量子比特的大型电路，电路规模接近 3000 个门，并从中提取无噪声的估计值。
[8:51] 简而言之，这超越了你可以用暴力经典计算模拟的范围。其影响确实巨大。我知道你们很多人都抓住了这个机会。这就是标志着实用时代的东西。现在，社区正在量子系统中运行实验，这些实验的规模足够大，可以研究量子计算超越暴力经典计算的实用性。
[9:24] 在这个新时代，我们还带来了一个新系统。因此，我也有幸向你们展示一个现存独一无二的系统，一台架构足够强大、模块化和灵活到足以与我们共同成长的计算机，随着我们在这段旅程中继续前进。所以，向IBM Quantum System Two问好吧。
[9:54] 它采用顶级冷却技术，在我们北边纽约州约克敦高地的TJ Watson研究中心运行100+量子比特的问题。看，这不是典型的独立系统。这是一个真正的模块化系统，允许我们将它们连接成更大、更强大的量子系统。这就是为什么我们称它为量子中心超级计算的构建块。我真心相信这个实用时代将巩固IBM Quantum作为科学和商业重要工具的地位，我们正以创新的速度将这项技术带给你们所有人。
[10:40] 但是看，一个系统最终的力量和 usefulness 取决于我们如何使其简单易用。这就是为什么我们带来Watson和生成式AI的力量，使我们的量子系统和Qiskit更加易于使用。我在这里向你们展示一个早期原型，还未完全准备好投入主流使用，但我想让你们今天看到未来。在这里你可以看到，你可以使用自然语言提示来生成Qiskit代码。
[11:14] 而这只是开始。想象一下未来几年平台的感觉，当我们将把AI的全部力量带到我们的量子平台上。所以这就是今天的IBM，比以往任何时候都更专注、更快速、更开放。这真是一个不可思议的时刻，感觉与我们以往所做的非常不同。
[11:44] 我们的CEO表示，IBM是一家混合云和AI公司。但毫无疑问，当一切尘埃落定，IBM将是一家混合云、AI和量子计算公司，一家致力于完善这些强大计算平台的公司，同时也致力于开发工具，能够将它们结合起来并扩展它们，以实现和解决超出我们所知可能的问题。
[12:17] 而你们，我的朋友们，将成为这段旅程和使命的一部分。说到这里，我想把话筒交给伟大的Jay Gambetta，他将上台与大家分享所有其他令人惊叹的工作和伟大进展，这些将我们今天聚集在一起。谢谢。谢谢Dario。谢谢，Dario。
[12:47] 大家早上好，欢迎来到IBM Quantum Summit。非常感谢你们加入我们的使命，将有用的量子计算带给世界，并使世界量子安全。很高兴看到我们的许多客户和合作伙伴在这里。没有你们，IBM Quantum就不可能实现。你们中的许多人已经和我们在一起一段时间了，但也有很多新面孔。所以我想花点时间感谢我们的新客户和合作伙伴。
[13:19] 我们欢迎与BasQ、RPI和RIKEN的量子计算中心。我们也欢迎与NQCC、KQC和哥本哈根大学的新量子创新中心。我们的新商业合作伙伴是EY，抱歉，KPMG、T-systems和SAP，我们的新行业客户是Banco Bradesco、Dow、Hyundai、Israel Aerospace Industries、Itau，抱歉，Moderna、Mitsubishi、Electronic、Pfizer和Truist。
[13:53] 谢谢大家。那么现在让我们开始吧。正如Dario所说，我们已经进入了量子计算历史上的一个新时代。我们称这个时代为实用时代。我们做了很多重要的工作。我们学到了很多关于量子计算应用的知识，例如，模拟自然数据与结构、搜索和优化。我们作为社区已经发表了数千篇论文。
[14:26] 我认为有2.5千篇论文正在发表，但我们还能做更多。直到最近，我们可以在笔记本电脑上运行大多数实验，或者如这张图所示。甚至一台1981 IBM PC。如果我们想从量子中提取效用，我们需要颠覆性的改变。今年夏天早些时候，我们发表了一篇论文，颠覆了这一点。我们称之为效用论文。
[14:57] 我们展示了量子可以解决超越暴力经典计算方法的问题。这就是我们所说的量子效用。我认为效用是通往量子优势之路的第一个里程碑。我们之所以能做到这一点，要归功于新颖的错误抑制和错误缓解方法。在四年里，我们能够将其提高一千倍。我们现在可以运行数千个门。我希望你们在接下来的一天里记住这一点。
[15:31] 更重要的是，在这个实用时代，我们开始将量子视为真相的来源。而经典方法只用于验证我们的结果。不要只听我说。这些是许多模拟这篇自然论文结果的论文中的话。但更重要的是，今年夏天，我们看到了至少六个更多的效用实验，运行了超过100个量子比特的量子电路，并使用了数百甚至数千个门。
[16:04] 这些实验正在推进超越量子计算的领域。他们使用量子计算，因为量子计算现在是发现的工具。在材料科学、凝聚态物理和粒子物理等科学领域进行发现。我们开始看到我所说的颠覆性变化。社区正在使用量子计算机做超越经典模拟所能做的事情。
[16:36] 我希望你们加入这场颠覆性变化的浪潮。你可以开始使用量子来探索以前无法访问的科学领域。我喜欢说，如果你不使用100个量子比特，你就不是在 doing 量子。所以我要把Rajeev请上台，告诉你们硬件的下一步。
[17:13] 谢谢你，Jay。大家早上好。那么，我将和大家稍微谈一谈我们为这个实用时代所构建的一些系统。好的。所以你们可能已经见过这张路线图，我们过去谈到过它，并且过去几年一直在执行。从我们几年前发布的 127 量子比特 Eagle 处理器，到去年宣布的 433 量子比特 Osprey 系统。
[17:48] 我们一直在持续交付这些创新，但我今天非常激动要分享的是我们今年在量子系统方面所做工作取得的这些令人兴奋的成果。我们有两个要谈。首先，在这个领域，我们一直在把规模推向极限，我们通过 Condor 做到了这一点，它是世界上最大的量子计算处理器。
[18:19] 不要鼓掌。好的。但 Condor 是我们探索单芯片扩展的结果。而在制冷机容量方面，它推动了单个芯片上能放置多少量子比特以及让这些量子比特良率达标的能力极限。所以它们日复一日地工作，我可以告诉你们这并非易事。借助 Condor，我们问的是，把超过一千个量子比特装入一个芯片和一个制冷机中会带来哪些挑战？例如，在左侧，你们会看到我们如何添加多层超导金属布线来解决密度问题。
[19:05] 在右侧，你们会看到一个关键指标，即相干时间，Condor 和 Osprey 之间非常相似，Osprey 是我们去年的 433 量子比特处理器。因此，我们在这块芯片上所要获得的规模学习已经开局顺利，我们对此非常兴奋。所以总结一下，Condor 是我们的 1121 量子比特芯片。它具有 50% 的量子比特密度提升，以及超过一英里，一英里，想想看，的超高密度柔性 IO 用于信号传输，全部容纳在一个制冷机内。哇。
[19:54] 因此，我们投入 Condor 的技术帮助我们找到了解决扩展问题的关键一块拼图。那已经过去了。但从我们的实用实验中，我们知道门深度和质量很重要。而这正是我们旅程的下一步所指向的。不要误会，我们对 Condor 非常兴奋，但我们真正、真正兴奋的是 Heron。
[20:28] 它就在这里，我可爱的助手 Jay。他们将向你们展示一个真实的 Heron 处理器，专门为你们带到这里。是的。谢谢。好的。Heron 是一个 133 量子比特芯片。借鉴我们使用固定频率量子比特的经验，并加以变化，使耦合器可调，从而赋予我们所需的灵活性。稍微回顾一下历史。早在 2019，当我们意识到必须做出改变时，团队确实回到绘图板，探索新的门架构。
[21:07] 你们即将看到的结果反映了这项工作的成果。这是我们的首块 Heron 芯片，代号 Monte-Carlo，它已经显示出比我们最好的 Eagle 显著提升。它的门错误率减半，一些保真度达到三个九。这太惊人了。但更重要的是，我们几乎消除了串扰。好的，我会稍微谈一下。
[21:39] 而且它在门时间上有显著改善。这将为我们的用户在实用时代带来巨大优势，并为我们继续扩展模块化处理器提供坚实基础。太棒了。现在，我将通过接下来几张小幻灯片向你们展示 Heron 到底提升了多少。所以这是一个评估量子比特间串扰的测量，对吧？你们会在右侧看到两条曲线，灰色的是隔离量子比特操作。
[22:15] 较深的线是当你寻址多个量子比特时。你们会看到我们最好的 Eagle 信号在多量子比特操作中随时间明显退化。那不是我们想要的。现在你们看 Heron 在非常类似的操作中，几乎不显示任何影响，几乎没有变化。这很重要。那么这对我们的用户意味着什么？
[22:46] 好的。这张图显示了一个我们认为能反映实用规模处理器价值主张的指标。我们不想只看我们最好的量子比特。我们想看在一条链中的所有量子比特。使用 Eagle，你们可以回到两年前，我们从每层门约 10% 的错误率开始。在过去几年里，我们已能将这个数字降到 2 到 3% 的范围，但这真的不是我们希望达到的位置。
[23:18] 接下来，就在 Heron 上，我们展示了这些长量子比特链中每个量子比特错误率 3 倍的提升。这将是巨大的。现在你们会看到两者的一些差异，蓝色和灰色点反映了 80 量子比特链和 100 量子比特链之间的差异。我们的目标是让两者看起来都与灰色条所在的位置相同，而这正是我们正在改进的过程。
[23:51] 以及我们的下一个版本。现在让我们看看性能。Heron 的性能再次明显优于 Eagle。这使我们能够运行更多的门，得益于更低的门时间。你们在那里看到一些数字。那些是相当不错的估计，但它为我们实现 5000 个门及以上的目标提供了一条非常非常清晰的跑道。你们今天剩下的时间会看到很多这样的内容。
[24:22] 那么你如何把这一切整合起来？对吧。你有了这些漂亮的数字。当用户使用这些处理器时，这意味着什么？你们会看到 Heron 在我们的用户所关心的地方显示出真正的性能提升。你们已经听 Dario 和 Jay 谈过那篇实用论文，我们在其中使用了类似的错误缓解技术，它显示出比我们在那篇论文中使用的 Eagle 有 3 到 6 倍的提升。这很巨大。对于更长的实验，这种提升变得更加明显。
[24:59] 现在是最精彩的部分。我很高兴地宣布，我们的首个客户 Heron 系统，名为 IBM Torino，现已在我们的纽约数据中心作为探索性设备提供。它的长链错误率在 100 个量子比特上低于 0.8%。这是我们迄今为止最好的系统。我们非常期待看到你们如何使用它。
[25:30] 看，这是一个绝佳的例子，展示了四年的研究与工程如何为我们的路线图其余部分提供动力。我们坚信我们拥有正确的量子比特和正确的门来实现这一目标。感谢大家的聆听，我们将把时间交还给 Jay。再见！是的，谢谢，Rajeev。
[26:01] 所以硬件看起来确实很棒。有些门已经达到了三个九的水平，你可以看到我们的重点是让整个设备真正大规模地工作。但你们都在这里，是因为我们通常有勾选路线图的传统。那么来看路线图。现在我们可以勾选 condor 和 heron。如果我点击按钮，我们就可以勾选 heron。现在我们进入了实用时代。
[26:33] 重点必须放在性能、稳定性和可靠性上。实用时代意味着我们的软件需要支持实用规模的工作负载。这意味着我们必须提升 Qiskit。我真的很兴奋能请 Jessie 上台来向大家讲解这意味着什么。谢谢。谢谢 Jay。大家早上好。我在这里向大家介绍 Qiskit。
[27:06] Qiskit 一直是创建、优化和执行量子电路与算子的事实标准。事实上，Qiskit 是绝大多数量子程序员首选的 SDK。Qiskit 的成功部分归功于其“一次编写，随处运行”的模式，该模式支持大多数主要的量子硬件供应商和架构，现在我们自豪地宣布，明年 Qiskit 将迈出重大一步，发布 Qiskit 1.0。
[27:44] 这个新版本将在性能、稳定性和可靠性方面提供更多改进。从现在起，我们认为运行量子应用遵循四步模式。现在我想告诉你们 Qiskit 如何帮助你们实现这一模式。第一步是将问题映射到量子电路和算子。Qiskit 拥有丰富的工具包来高效构建这些电路和算子，而 Qiskit 1.0 增加了对动态电路的原生支持，例如循环、分支和经典表达式。
[28:22] 它还在内存使用方面做出了显著改进，与一年前相比减少了多达 55%。第二步是优化输入。Qiskit 转译器是业界领先的工具，用于转换电路以符合目标硬件的约束。此外，Qiskit 转译器有一个传递管理器，让你能够灵活地进一步自定义要应用于电路的优化，而在 Qiskit 1.0 中，我们看到转译器在速度和质量上都有大幅提升。
[29:00] Qiskit 转译器现在比一年前快 16 倍，并且与另一个知名工具包相比，可以生成少 23% 的双量子比特门。我们还开始尝试使用 AI 来进一步改进转译过程。这可以轻松地与现有的 Qiskit 转译器结合。得益于其传递管理器的灵活性，使用强化学习方法，与 Qiskit 中的启发式算法相比，我们在电路深度和 c-not 计数方面都看到了 20 到 50% 的改进。
[29:38] 这个新 AI 转译器的 alpha 版本今天就可用了。最后，第三步是在原语中执行这些优化后的电路。我们今天拥有的两个原语，采样器和估算器，封装了对量子处理器最常见的查询。在 Qiskit 1.0 中，我们调整了它们的接口，使其在 Qiskit 运行时中更简单、更一致、更高效。我们现在支持三种不同的执行模式：单作业、批处理和会话。
[30:17] 运行独立作业是在小型设备上执行量子电路的传统方式，它对于测试和调试仍然有用。去年我们引入了会话，允许迭代工作负载完成而无需为每次迭代排队等待。然而，在当今的实用时代，单个会话可能不足以支持现实生活中的工作负载。因此，会话很快就可以通过多个活动窗口进行扩展，以适应长时间的计算。
[30:52] 今年我们还添加了批处理，允许你一次提交多个非迭代查询。现在，借助批处理执行模式，由于线程并行化，你的工作负载运行速度可提高多达五倍。谢谢大家。现在把时间交还给 Jay。谢谢。开始吧。
[31:23] 所以我们的软件将迎来很多更新。但同样，我们必须回到路线图。我们在路线图上承诺的是线程化原语。我们交付的是一套执行模式。我们称之为单作业、批处理和会话。因此，我们可以勾选执行模式。现在我们已经进入了这个新时代，人们正在使用量子计算来做更多的发现工具。
[31:58] 这意味着我们看到了更多具有额外需求和要求的用户。我很高兴请 Paul 上台来谈谈我们如何支持这些新用户。谢谢你 Jack。谢谢。大家早上好。随着我们的硬件和软件不断改进，它们为用户将量子计算集成到其平台中开辟了新的机会。
[32:33] 进入实用时代后，我们迎来了一批新的用户，我们称之为量子计算科学家。量子计算科学家并不关心量子硬件本身，而是关心如何利用机器来解决特定的计算任务。这些量子计算科学家最看重性能、兼容性和易用性，而思考如何在量子算法和应用方面满足这些需求，让我们提出了 Qiskit pattern 这一理念。
[33:05] Qiskit pattern 是所有量子算法和应用都必须遵循的四步序列。第一步，我们生成量子电路和算子，通常来自经典输入数据。第二步，我们拿这些量子对象，针对量子硬件上的执行进行优化。第三步，使用我们之前提到的 Qiskit primitives 执行我们的实验，最后是第四步，我们对 primitive 的输出进行后处理。好了，现在 Qiskit pattern 不仅仅是步骤的集合。
[33:36] 它们提供了一个逻辑框架，我们可以从中开始探索大规模地编写算法和应用。没错。首先，当你以步骤来思考时，你就突出了量子工作流所构建的基础构件，以及如何利用这些组件来创建多样化的量子例程。它们还支持容器化。一个 Qiskit pattern 构成一个完整的量子程序，我们可以开始用量子组件来增强已有的工作流。
[34:06] 最后，它们允许从量子电路和算子中进行抽象，使用户不必在量子汇编代码的层面上工作。好了。为了看看这一切是如何运作的，让我们看一个为量子化学设计的 pattern。没错。就是这里。你再次看到，四个步骤，现在每一步都由一组块组成，每个块执行单一任务。这些块可以由 IBM、第三方提供商，甚至开源贡献者制作。例如，如果我们想改变这个问题中电路和算子的构建方式，只需交换两个块，pattern 的其余部分保持不变。
[34:46] 我们可以更进一步，比如说让我们改变第三步中使用的经典优化器，这也很容易做到。然而，在这个特定的例子中，优化器本身并不是一个构建块。这是因为 Qiskit pattern 旨在利用已有的软件框架，比如本例中来自 Scipy 的优化器，并且只构建实现目标量子加速所需的核心功能。这种有针对性的方法在审视已有的企业工作流时非常有用，因为整个例程中只有一小部分适合量子解决方案。
[35:23] 使用与原始工作流相同的输入和输出。Qiskit pattern 可以针对管道中的特定用例进行定制，简化开发和集成过程。最后，Qiskit pattern 设计为可大规模使用。它们可以通过资源管理信息得到增强。从而优化其执行。一个 pattern 可以上传到异构计算基础设施，例如 quantum serverless，从而实现简化的执行。
[35:55] 最后，我们的 pattern 可以无人值守地运行，并由对量子计算一无所知的最终用户使用。好了，这就是我们通过 Qiskit pattern 以及可扩展算法和优化设计所要前进的方向。我们将从明年开始陆续推出这些组件。不过，随着 Quantum serverless 已经进入测试阶段，你今天就可以开始为此做准备。非常感谢。
[36:29] 谢谢 Paul。所以如你所见，我们真的在努力让量子变得无摩擦。在路线图上，我们说过会引入原型量子函数，现在你看到我们开始通过 Qiskit pattern 的含义来定义它。所以，在路线图上勾选进展，我们现在有了一种原型方式来创建我设想的未来软件函数，以及 quantum serverless。
[37:00] 到目前为止，实用性的论文一直是本次演讲的主要焦点。但今年我们还有另一个公告。我们发布了一种新的纠错码。这是一种新的低密度奇偶校验码，我们实际上称之为 gross code。这种新码所需的量子比特数量比表面码少几个数量级，而且扩展效率更高。但实现这种码需要新的创新。我们需要一种新型耦合器来连接芯片上相距较远的量子比特。
[37:31] 我们称之为 c 耦合器。它要求我们的量子比特具有更高的连接性。度数为六。所以这是我们的最终目标。我们需要一个具有 c 耦合器的系统，以实现芯片上的长距离量子连接，从而实施该码。我们需要 l 耦合器来创建大规模系统并在芯片之间传输量子信息。我们需要 m 耦合器在芯片之间进行短距离信息传输，以便用更小的芯片制造更大的芯片。
[38:06] 我希望你已经看到，我们已经证明在完全容错之前就有大量实用性可用。现在我们有一种比表面码扩展性更好的码，这意味着纠错比我们想象的更近了。但我们目前的路线图上还没有实现纠错的路径。这意味着我们需要一个更大的路线图。
[38:38] 这是我们的新路线图。让我们来参观一下。让我们参观一下我们的开发路线图。它描绘了我们面向客户的系统和服务的前进道路。在这里你可以看到，我们实际上展示了未来五年我们的处理器在单个电路中可以运行的门的数量。
[39:13] 抱歉。现在我们展示的是单个电路上可以运行的门的数量，而不是量子比特的数量。而在接下来的五年里，我们将专注于质量而非规模。这是因为我们相信我们已经通过 condor 解决了单芯片扩展的问题。我们拥有构建更大系统所需的工具。更大的挑战在于我们需要用于实用性的工具，并持续提升质量。因此，我们将其置于核心位置。
[39:46] 在接下来的五年里，我们将使我们处理器能运行的门数量增加两倍。此外，Flamingo 还包括量子通信，我们将能够连接或协同工作至少七个 QPU，从而创建一个拥有数千量子比特的系统。但随后在 2029 年发生了重大事件。我们有了 Starling。你可以看到直接跃升至 100 百万个门。
[40:18] 对我来说，这将是我们能够编码到电路中的大量算法精度。Starling 将是一件非常、非常重大的事。它将需要许多新技术，我们将把所有这些技术整合到一个可部署的系统中。是的，那就是纠错。我们要说在 2029 年我们将拥有第一个具备纠错功能的系统。
[40:49] 我们还将一路展示到 2033 年，向您展示我们如何扩展到 1 亿个门的详细计划。但我们意识到我们需要的不只是一份开发路线图。因此，我们在 IBM 推出了所谓的创新路线图。如您所知，我们致力于对进展保持透明。我们希望您相信我们正在取得进展，并希望您与我们同行。
[41:21] 所以正如我之前所说，要实现 starling，我们需要开发 m 耦合器、L 耦合器和 C 耦合器。我们将把 L 耦合器构建到 Flamingo 中。我们将把 M 耦合器构建到 crossbill 中，并将 C 耦合器构建到 kookaburra 中。结合我们的软件创新，这将使我们继续沿着通往实用量子计算的道路前进。实际上，我们在 kookaburra 中投入了更多。这包括 c 耦合器、六度耦合、用于解码的软件，清单还在不断延长。
[41:51] 因此，我希望您能接受，我们不是在计划的 2025 年交付 Kookaburra，而是将其推迟到 2026 年，以便我们能够整合所有这些效应。路线图中的创新将涌现到开发路线图中，并最终提供给我们的客户。我们与 Flamingo、Crossbill 和 Kookaburra 一起构建的耦合器将使我们能够引入纠错。与 starling 一起。
[42:23] 至此，扩展量子实用性的道路已经清晰。谢谢。从我们的用户那里，我们了解到了很多关于人们想要运行的工作负载的信息。为此我们感谢您。我们现在知道，运行单个电路是不够的。您希望并行运行多个电路，并进行并发的经典计算。
[42:54] 这就是 Dario 在演讲中提到的以量子为中心的超级计算。它驱动着您今天看到的所有更新背后的愿景。那么回顾一下，第一，我们有 IBM 量子系统二号。在我看来，它看起来很棒。它是模块化的，并将持续扩展我们走向未来。第一个系统已经内置了三个 heron，第二，heron，我们的高性能系统已经使实用工作负载的运行速度提高了五倍。我迫不及待想看到那些结果。
[43:24] 第三，通过 Condor，我们解决了单量子比特扩展问题。第四，支持并行化。通过执行模式，批处理和迭代工作负载将运行得更快。第五，我们简化了量子算法，并引入了称为 Qiskit patterns 的概念。第六，我们创建了 quantum Serverless，目前处于测试阶段。这意味着它更加稳定，我非常期待您看到它。但所有这些都在我们的路线图上。我希望您也喜欢我们向您展示的路线图之外的新增内容。
[43:56] 第一个是 Qiskit 将变成 Qiskit 1.0。它将快速且稳定。正如 Dario 提到的，人工智能领域经历了疯狂的一年，我们正在利用人工智能的全部力量来简化量子计算机的运行方式。我们正在引入一些方法，比如使用自然语言来创建量子代码。对我来说，这是一件大事。第九，我们实际上已经介绍并展示了如何使用人工智能来推动生成更好的电路，这将作为 Alpha 阶段的转译器服务推出。
[44:31] 您现在就可以开始使用它。最后，第 10 个公告是，我们现在为所有人制定了一个十年路线图，你们知道我们多么认真对待路线图，制定十年路线图是一件大事，将纠错纳入其中是我们对如何将实用量子计算带给世界的承诺。至此，我想最后说，我们正处于量子实用性时代，正如我们在 2021 年所说，2023 年将是大有可为的一年。
[45:02] 我们有一个清晰、详细的扩展量子计算的路线图。我们已经实现了所有里程碑，现在我们拥有能够探索超越经典暴力计算的问题的系统。现在，我希望您享受今天剩下的会议，您将深入了解我谈到的事情以及更多、更多、更多的内容。量子实用性之年已经到来，我希望您能与我们一同踏上这段旅程。谢谢。
查看英文原始字幕
[0:06] Good morning. Let's take a minute to invite all our friends and colleagues to sit down so that we can get started. But first, welcome to the 2023 Quantum Summit. We could not be happier that you have joined us today. My name is Dario Gil. I am a senior vice president at IBM and the director of IBM Research. And wow, what a year 2023 has been from the perspective of technology.
[0:38] Look, computing from AI to semiconductors to quantum computing, it is once again driving markets and birthing industries and really igniting people all over the world. And the imaginations of what we're going to be able to do with technology. And I really believe that we are living in the most exciting times in the world of computing, probably since the advent of either the digital computers in the 1940s or the transistor in the late fifties.
[1:10] And really now with these new technologies, we begin to have an opportunity to start tackling problems that we didn't even think were possible to be addressed with the previous generation of technology. I'm proud to say that this year marks my 20th anniversary at IBM and thank you and I hope you all are getting to see that the company is just so different than it was even five years ago. And this is in no small part because of the foundational work that we carry out in the R&D community inside IBM.
[1:49] And frankly, the rate and pace at which we are commercializing technologies. Look, in my 20 years at IBM, I have never seen us commercialize technology at a faster pace than we're doing it today. We like to say that within the research division that we are the organic growth engine of the company and we're really relentlessly focused on how do we bring the innovations that we're developing in the R&D community and bring them into products.
[2:21] And I highlight as an example of that rate and pace, what we've done this year with AI. I mean, we all know that in the world of AI, you know, as far as the public is concerned, certainly has entered the public imagination of what is possible with that technology. We had become deeply passionate about the role of foundation models for quite a few years inside the research division. But this year we commercialize and created with an intensity that is unlike no other, a new platform to bring generative AI to the world of enterprises and government called watsonX.
[2:56] And we were able to do this in a matter of months. Back to this theme of commercializing technology faster than ever before, and people are really noticing. But of course there is more. There is the world of quantum. And look, I remember reading, you know, a few decades back some of the core ideas around quantum information, science and thinking how powerful those ideas were to become. And if you look at that journey now, they are becoming a reality. And the journey has been amazing.
[3:29] Recall that we made news in the context of the quantum industry in 2016 when we put the first quantum processor on the cloud and within a week we had five times more members that had joins and users than we had expected. And really that marked the moment where quantum computing became accessible to the world. First it showed to hundreds, then to thousands, and eventually to close to half a million people that quantum computing is something to be considered, embraced, and nurtured.
[4:02] Now, in 2017, we introduced Qiskit our Open Source Quantum SDK, and it has become the most widely used software development environment in the world for quantum. In fact, let me give you a statistic. Over 81% of quantum developers prefer it as do over half a million users. That same year we launched the IBM Quantum Network and it exploded. We now have over 280 members today and many of you are here today.
[4:32] So I thank you for that. In 2018, we also established the Quantum Innovation Centers, which have now grown to 38 worldwide. In 2019, we launched IBM Quantum System one. It was the world's first fully integrated and commercially available quantum computer. And since then we've been applying Quantum System Ones all over the world There are now system ones in Germany, Japan and Canada.
[5:05] More are being installed in South Korea, Spain, Japan and here in the US. All this activity is really fantastic and staggering and it is why we have come to call these past few years the era of the emergence of quantum computing. And during this time we have developed more quantum computing power and put it in the hands of more people than anyone else in the industry. Now, it's really been remarkable to see the problems that the community that we build together has been able to tackle.
[5:43] It really hasn't been easy. And but what's made it possible is our philosophy of continued rapid innovation in both hardware and software. In hardware we bought, we brought at least one quantum processor every year with a signature breakthrough. In 2019, we introduced Falcon, a 27 qubit processor. A year later, Hummingbird boasting 65 qubits. And in 2021 we broke the 100 qubit barrier with Eagle.
[6:19] And just last year we introduced Osprey with 433 qubits. And today, as promised, we're announcing Condor with a soaring 1121 qubits. With Condor, we have solved the problem of scaling, of qubit scaling. Now we're taking that knowledge and applying it as we build larger systems. And of course, we didn't stop there.
[6:50] We were also dedicated to improving gate quality and increasing the circuit depth. And that is why as excited as we are about Condor, we're even more excited about Heron. This is Heron, the most performant quantum processor in the world, and the one that will truly scale quantum computing. While Condor removed the roadblock to scale Heron pushes circuit depth and quality with its tunable coupler technology that I'm so proud to introduce today.
[7:29] And wow, if you look at the collective achievements on both on the hardware and the software and the ecosystem, let me give you a statistic that is remarkable. These are the number of circuits that have been executed on IBM Quantum's fleet, over 3 trillion of them. So if we recap what we have accomplished over the last few years, it really has been phenomenal. And in a moment to celebrate and we would say that that phase marked the emergence of quantum computing and that it is time now to move to a new, new era that we're humbly calling the era of utility.
[8:10] Utility is the theme of this quantum summit. And there was one key advancement that we believe marked the transition to this new and exciting era. In June, the team published a breakthrough in scaling quantum computation that was featured on the cover of Nature. The title of the paper is Evidence of Utility Before Full Tolerance in it, our team showed that we can now run large circuits with over 100 qubits and a circuit size of almost 3000 gates and extract noise free estimates from them.
[8:51] So simply put, this goes beyond what you can simulate with brute force. Classical computation and the implications are really huge. And I know so many of you jumped on it. This is what Marked the utility era. And now the community is running experiments in quantum systems today that have both a scale that is large enough to investigate the utility of quantum computing beyond brute force classical computation.
[9:24] And with this new era, we also bring a new system. So it is also my privilege to reveal to you a system like no other in existence a computer with an architecture that is powerful enough, modular and flexible enough to grow with us as we continue to march forward in this journey. So say hello to the IBM Quantum System Two.
[9:54] It is prime cooled and running 100 plus qubit problems just north of us at the TJ Watson Research Center in Yorktown Heights, New York. Look, this is not your typical stand alone system. This is truly a modular system that allows us to connect them into larger, more powerful quantum systems. That's why we call it the building block of quantum centric supercomputing. I truly believe that this era of utility will cement IBM Quantum as an important tool for science and business, and we're bringing this technology to all of you as fast as we're innovating it.
[10:40] But look, a system in the end is only as powerful and as useful as we make it, driven by simplicity. And that is why we are bringing the power of Watson and generative AI to make our quantum systems and Qiskit ever easier to use. I'm showing you here an early prototype, still not quite ready for prime time, but I wanted you to see the future today. Here you see that you can use natural language prompts to generate Qiskit code.
[11:14] And this is just the beginning. Imagine what the platform is going to feel like in the next few years as we bring and we will the full power of AI to our quantum platform. So this is today's IBM, more focused, faster, more open than ever before. And it really is an incredible moment and it feels very different than we've ever done before.
[11:44] Our CEO expresses that IBM is a hybrid cloud, and AI company. But make no mistake, when all of this is said and done, IBM will be a hybrid cloud. AI and Quantum Computing Company, a company that is dedicated to perfecting each of these powerful computing platforms, but also the tools to be able to combine them and extend them to achieve and solve problems beyond what we know is possible.
[12:17] And you, my friends, will be part of that journey and that mission. And on that note, I want to pass it on to the great Jay Gambetta, who's going to come on stage and share with you all the other amazing work and great advancements that have brought us all together today. Thank you. Thank Dario. Thanks, Dario.
[12:47] Good morning, everyone, and welcome to the IBM Quantum Summit. Thanks so much for joining us on our mission to bring useful quantum computing to the world and make the world quantum safe. It's great to see many of our clients and partners here. Without you, IBM Quantum would not be possible. Many of you have been with us for a while, but there are many new faces. So I wanted to take a moment and just acknowledge our new client clients and partners.
[13:19] We welcome quantum computational centers with BasQ, RPI and RIKEN. We also welcome new Quantum Innovation Centers with NQCC, KQC and the University of Copenhagen. Our new commercial partners are EY, Sorry, KPMG, T-systems and SAP and our new industry clients are Banco Bradesco, Dow, Hyundai, Israel Aerospace Industries, Itau, sorry, Moderna, Mitsubishi, Electronic, Pfizer and Truist.
[13:53] Thank you all. So now let's get started. As Dario said, we've entered a new era in the history of quantum computation. And we call this era the era of utility. We've done a lot of important work. We've learned a lot about quantum computing applications, for instance, and simulating nature data with structure, search and optimization. And we as the community have published thousands of papers.
[14:26] I think it's 2.5 thousand papers are being published, but there's a lot more we can do. Until recently, we could run most of the experiments on a laptop as a or as this graph shows. Even a 1981 IBM PC. We need a disruptive change if we want to extract utility from quantum. And earlier this summer we published a paper that shakes this up. We call it the utility paper.
[14:57] We showed that Quantum can tackle problems beyond brute force classical computing methods. And that's what we mean by quantum utility. I think of utility as the first milestone to the road to quantum advantage. We got here thanks to the novel error suppression and error mitigation methods. And in four years, we've been able to increase it by a thousand times. We can now run thousands of gates. And I want you to keep this in mind for the rest of the day.
[15:31] And what's more, in this utility era, we're starting to treat Quantum as the source of truth. And classical methods only serve to verify our results. And don't take my words for this. These are the words of many papers that simulated the results of this nature paper. But what's even more important is this past summer, we've seen at least six more utility experiments that have run quantum circuits with more than 100 qubits and using hundreds or even thousands of gates.
[16:04] These experiments are advancing fields that go beyond quantum computation. They're using quantum computing for quantum computing is now a tool for discovery. Discovering in scientific domains like material science, condensed matter, and particle physics. And we're starting to see this disruptive change. I was talking about. The community is using quantum computers to do things beyond what we can do with the classical simulation.
[16:36] And I hope you'll join this groundswell of disruptive change. You can start using quantum to explore scientific realms that weren't accessible before. I like to say, if you're not using 100 qubits, you're not doing quantum. So I'm going to bring Rajeev to the stage to tell you about what's next in hardware.
[17:13] Thank you, Jay. Good morning, everyone. So I'm going to talk to you a little bit about some of the systems we've put together for this era of utility. All right. So you've probably seen this roadmap that we've talked about in the past that we have been executing on for the past few years. From our 127 Qubit Eagle processor released a couple of years ago to our 433 qubit Osprey system announced last year.
[17:48] We've continued to deliver on these innovations, but I'm really thrilled to share about today is these exciting results from the work we have done on quantum systems this year. And we have two of them to talk about. First, in this area, we've been pushing scale to the limit, and we have done that with Condor, the world's largest quantum computing processor.
[18:19] No claps. All right. But Condor is a result of our exploration into single chip scaling. And fridge capacity, it pushes the limits of how many qubits can be put on a single chip and the ability to yield those qubits. So they work day in and day out, which I can tell you is not a trivial task. With Condor, we asked, what are our challenges that would come fitting over a thousand qubits into one chip and into one fridge, For example, on the left, you'll see how we've added layers of superconducting metal wiring to solve the density problem.
[19:05] And on the right you'll see a key metric of coherence Time is very similar between Condor and Osprey, which was our 433 qubit processor last year. So the learning on scale that we set out to get on this chip is off to a flying start and we're really excited about it. So just to summarize, Condor is our 1121 qubit chip. It features a 50% increase in qubit density and over a mile, a mile, Think about that, of super high density flex IO for signal delivery all fit within a single fridge. Wow.
[19:54] So the technology we put into Condor helped us figure out a key piece of the puzzle to solve the scaling problem. And that's behind us. But from our utility experiments, we know that gate depth and quality matters. And that's where the next step in our journey takes us. Don't get me wrong, we are super excited about Condor, but what we are really, really excited about is Heron.
[20:28] And here it is, my lovely assistant, Jay. They're going to show you a real life heron processor brought here, especially for you. Yes. Thank you. All right. Heron is a 133 qubit chip. Learn from our experience from fixed frequency qubits with a twist, making the couplers tunable to give us the flexibility we need. A little bit of a history lesson. Back in 2019, when we realized we had to do something different, the team really went back to the drawing board to explore new gate architectures.
[21:07] The results that you're about to see reflect the fruits of that labor. This is our first heron chip codenamed Monte-Carlo, and it's already showing significant improvements over our best eagle. It has half the gate error rate and some fidelities in the three nines. It's amazing. But what's more important, we virtually eliminated crosstalk. All right, I'll talk a little bit about it.
[21:39] And it has a significant improvement in gate time. This is going to give our users a huge advantage in the utility era and give us a solid foundation for where we can continue to scale our modular processors. Amazing. And now I'm going to show you over the next few little slides how much heron has really improved. So this is a measurement to assess qubit to qubit crosstalk, right? You'll see two curves on the right, the gray is the isolated qubit operation.
[22:15] And the darker line is where you address multiple qubits. You'll see that our best eagle signal clearly degrades over time. In the multi qubit operation. That's not what we wanted. And now you look at Heron in a very similar operation, it hardly shows any effect, barely a change. This is a big deal. So what does this all mean for our users?
[22:46] All right. The chart shows a metric we believe reflects the value statement of a utility scale processor. We didn't want to really look at just our best qubits. We wanted to look at all our qubits in a chain With Eagle, you can go back two years ago, we started at about 10% error per layered gate. In the last couple of years, we've been able to take that number down to the 2 to 3% range, but that's really not where we want it to be.
[23:18] Coming up, right with Heron, we show a 3x improvement in the error per qubit for these long chains of qubits. This is going to be huge. Now you'll see some difference in the two, the blue and the gray dots that reflect the difference between the 80 qubit chain and 100 qubit chain. Our goal is to get both of them looking the same next to where the gray bar is, and that's what we are in the process of improving.
[23:51] And our next version. Now let's look at performance. Herron's performance again is markedly better than Eagles. That allows us to run a lot more gates enabled by the lower gate time. You see some numbers there. Those are pretty good estimates, but it gives us a very, very clear runway to our goal of 5000 gates and beyond. And you're going to see a lot of that coming for the rest of the day.
[24:22] So how do you put it all together? Right. You have these nice numbers. What does it mean for when the users are using these processors? You'll see Herron is showing real performance improvements, where our users care. You've already heard about the utility paper from Dario and Jay Where we use similar techniques of error mitigation, it shows a 3 to 6x improvement over the eagle we leveraged in that paper. That is huge. This improvement becomes even more noticeable for longer experiments.
[24:59] And now the best part. I'm excited to announce that our first client Heron system called IBM Torino, is now available as an exploratory device in our New York data center. It has a long chain error rate of 100 qubits of less than 0.8%. Our best system ever. We are super excited to see how you use it.
[25:30] Look, this is a great example of four years of research and engineering that will power the rest of our roadmap. We strongly believe we have the right qubits and the right gates to make this a reality. Thanks for listening and we'll turn it over back to Jay. See ya! Yeah Thanks, Rajeev.
[26:01] So the hardware is truly looking great. Some of the gates are as good as three nines and you see our focus is focusing on getting the whole device to work really large. But you're all here because we generally have this tradition of ticking off the roadmap. So to the roadmap. Now we can tick off condor, and heron. If I click the button and we can tick off heron. Now we're in this utility era.
[26:33] The focus has got to be on performance stability and reliability. Utility era means we need our software to support utility scale workloads. And this means we've got to level up Qiskit. I really am excited to bring Jessie to the stage to talk you through what this means Thank you. Thank you Jay. Good morning, everyone. I'm here to tell you about Qiskit.
[27:06] Qiskit has been the de facto standard for creating, optimizing and executing quantum circuits and operators. In fact, Qiskit is the preferred SDK by the vast majority of quantum programmers. Qiskit's success is party because of its write once and run everywhere model That supports most major quantum hardware vendors and architectures, and now we are proud to announce that next year Qiskit will take a major step forward with the release of Qiskit 1.0.
[27:44] This new release will offer even more improvements in performance, stability and reliability. From now forward, we think of running a quantum application with a four-step pattern. And now I want to tell you how Qiskit can help you implement that pattern. Step one is to map problems to quantum circuits and operators. Qiskit has rich toolkits to efficiently construct these circuits and operators, and Qiskit 1.0 has added native support for dynamic circuits such as loops, branches and classical expressions.
[28:22] It has also made significant improvements on memory usage, reducing it by as much as 55% compared to a year ago. Step two is to optimize the inputs. The Qiskit transpiler is an industry leading tool for converting circuits to respect the constraints of the target hardware. In addition, the Qiskit transpiler has a pass manager. that gives you the flexibility to further customize the optimization you want to apply to your circuits and Qiskit 1.0, We see a substantial improvement in both speed and quality for the transpiler.
[29:00] The Qiskit Transpiler is now 16 times faster than a year ago and can generate 23% fewer two-qubit gates. Than another well known toolkit. We also started experimenting, using AI to further improve the transpilation process. This can be easily coupled with the existing Qiskit transpiler. Thanks to the flexibility of its pass manager using a reinforcement learning approach, we've seen a 20 to 50% improvement in both circuit depth and c-not counts compared to heuristic algorithms in Qiskit.
[29:38] The alpha release of this new AI transpiler is available today. And finally, step three is to execute these optimized circuits in primitives. The two primitives we have today sampler and estimater encapsulate the most common queries to a quantum processor. In Qiskit 1.0. We treat their interfaces to make them simpler, more consistent and more efficient in Qiskit runtime, We now support three different execution modes single job, Batch, and session.
[30:17] Running a standalone job was the traditional way of executing quantum circuits on small devices, and it can still be useful for testing and debugging. Last year we introduced sessions which allowed iterative workloads to complete without queuing delays for each iteration. A single session, however, may not be long enough for a real life workload in this utility era. So a session can soon be extended with multiple active windows to accommodate lengthy computations.
[30:52] This year we also added batch, which allowed you to submit multiple non-iterative queries at once. And now with batch execution mode, your workloads can run up to five times faster thanks to parallelism with threading. Thank you everyone. And now in the back to Jay Thanks. There to go.
[31:23] So there's a lot of updates coming to our software. But again, we have to go to the roadmap. What we promised on the Roadmap was threaded primitives. What we've delivered is a suite of execution modes. We call them single job, batch, and sessions. And so with that, we can tick off execution modes. So now that we've entered this new era, people are using quantum computation for doing more tools with discovery.
[31:58] This means we're seeing additional users with additional needs and requirements. And it's my pleasure to bring Paul to the stage to talk about how we're supporting these new users. You Jack. Thanks. Good morning, everyone. So as our hardware and software continue to improve, they open up new opportunities for users to integrate quantum computing into their platform.
[32:33] Having entered the Era of utility has opened up a new set of users that we call quantum computational scientists. Now, a quantum computational scientist is not interested in the quantum hardware itself, but rather in utilizing a machine to solve a distinct computational task. These quantum computational scientists value performance compatibility and ease of use over everything else, and thinking of how to try to satisfy these needs when it comes to quantum algorithms and applications has brought us to the idea of a Qiskit pattern.
[33:05] A Qiskit pattern is the sequence of four steps that all quantum algorithms and applications must follow. In step one, we generate quantum circuits and operators, most likely from classical input data. In step two, we take those quantum objects and we optimize them for execution on quantum hardware. Step three executes our experiment using the Qiskit primitives that we heard about and finally is step for, we post process the primitive output. Okay, now Qiskit patterns are more than just a collection of steps.
[33:36] They provide a logical framework from which we can begin to explore writing algorithms and applications at scale. Right. First, when you think in terms of steps, you highlight the foundational building blocks of which quantum workflows are built and how to leverage those components to create a diverse set of quantum routines. They also allow for containerization. A Qiskit pattern forms an entire quantum program, and we can begin to enhance pre existing workflows with quantum components.
[34:06] And finally, they allow for abstraction away from quantum circuits in operators, alleviating in users from working at the level of quantum assembly code. Okay. To see how this all works, let's take a look at a pattern designed for quantum chemistry. Right. So here it is. You see, again, the four steps were now each Step is comprised of a collection of blocks where each block performs a singular task. These blocks could be made by IBM, third party providers, or even open source contributors. If, for example, we want to change how the circuit and operator in this problem are constructed its a simple exchange of two blocks, leaving the rest of the pattern unchanged.
[34:46] We can take it a step further and we can say let's change the classical optimizer used in step three, and that's also easy to do. However, in this particular example, the optimizer itself is not a building block. This is because Qiskit patterns aims to leverage preexisting software frameworks such as the optimizes from Scipy in this case, and only build the core functionality needed for targeted quantum acceleration. This targeted approach comes in handy when looking at preexisting enterprise workflows, where only a small portion of the of the overall routine is amenable to a quantum solution.
[35:23] Using the same inputs and outputs as the original workflow. A qiskit pattern can be tailored to specific use cases in pipelines, streamlining the development and integration process. And finally, Qiskit patterns are designed to be used at scale. They can be enhanced with information about resource management. Thus optimizing their execution. A pattern can be uploaded to a heterogeneous computing infrastructure such as quantum serverless, allowing for streamlined execution.
[35:55] And finally, our pattern can be run unattended and utilized by end users with no knowledge of quantum computing whatsoever. Okay, so now this is a flavor of where we are going with Qiskit patterns and scalable algorithm and optimization design. And we're going to start rolling these components out beginning next year. However, with Quantum serverless already out in beta, you can begin to prepare for this today. So thank you very much.
[36:29] Thanks Paul So as you see, we're really trying to make quantum frictionless. On the roadmap We said we'd introduce prototype quantum functions and now you see we're starting to define that with what we mean by Qiskit patterns. So to the roadmap to tick off the progress, we now have a prototype way of creating what I envision to be software functions of the future and quantum serverless.
[37:00] So the utility paper has been the main focus of this talk far. But we made another announcement this year. We released a new error correcting code. It's a new low density parity check code that we actually call the gross code. This new code requires orders of magnitude fewer qubits than the surface code, and it scales more efficiently. But implementing this code requires new innovation. We need a new type of coupler to connect qubits that are further apart on the chip.
[37:31] We call these c couplets. It requires more connectivity on our qubits. Degree six. So here's our ultimate goal. We need a system that has c couplers to enable long range quantum connections on the chip in order to implement the code. We need l couplers to create large scale systems and transfer quantum information across the chips. And we need m couplers to transfer information short range between the chips to make bigger chips out of smaller chips.
[38:06] I hope you've seen that We've already proved there's plenty of utility to be had before full tolerance. And now we have a code that that scales better than the surface code, which means error correction is closer than we thought. But we don't currently have a path for error correction on our roadmap. And this means we're going to need a bigger roadmap.
[38:38] Here's our new roadmap. Let's take a tour. Let's take a tour on our development roadmap. It charts the path forward for our client-facing systems and services. Here you see, we are actually showing the number of gates our processors can run in a single circuit for the next five years.
[39:13] Sorry. Here we now show the number of gates that you can run on a single circuit rather than the number of qubits. And for the next five years, we're going to be tackling exactly that quality as opposed to scale. That is because we believe we've solved ssingle chip scaling with the condor. We have the tools we need to build larger systems. The bigger challenge is the tools we need for utility and to continue to improve the quality. So we're putting that front and center.
[39:46] over the next five years We're going to triple the number of gates our processors can run. What's more, Flamingo also includes that quantum communication, and we'll be able to bring at least seven QPU's connected or working together to create a system of thousands of qubits. But then something big happens in 2029. We have Starling. You can see a stark jump all the way to 100 million gates.
[40:18] To me, this is going to be a lot of algorithmic accuracy that we can code into our circuits. Starling is going to be a big a really big deal. It's going to require a lot of new technologies and we're going to take all those technologies and put them into a deployable system. Yes, that's error correction. We're going to say that in 2029 we'll have the first system with error correction.
[40:49] We're also going all the way to 2033 to show you our detailed plans for how we're going to scale to 1 billion gates. But we've realized we need more than a development roadmap. So we're introducing what we call the innovation roadmap. at IBM As you know, we're committed to being transparent about our progress. We want you to trust that we're making progress and we hope that you go on this journey with us.
[41:21] So as I said before, to achieve starling, we're going to need to develop m couplers, L couplers and C couplers. We're going to build the L couplers into Flamingo. We're going to build the M couplers into crossbill and we're going to build the C couplers in the kookaburra. combined with our software innovations this is going to continue us along the path towards useful quantum computing. We're actually putting a lot more in kookaburra. This includes a c couplers, the degrees six couplings, the software for decoding and the list just goes on and on.
[41:51] So I hope you accept that rather than delivering Kookaburra in the planned 2025, we're moving it out to 2026 so we can incorporate all of these effects. Innovations from the roadmap will bubble up into the development roadmap and eventually become available to our clients. The couplers we build with Flamingo, Crossbill and Kookaburra will let us introduce error correction. With starling.
[42:23] And with that, the road is clear to extending quantum utility. Thanks. From our users, we've learned We've learned a lot about the workloads people want to run. And we thank you for this. We now that running a single circuit is not enough. You want to run multiple in parallel and with concurrent classical computations.
[42:54] This is the quantum centric supercomputing that Dario mentioned in his talk. It is driving the vision behind all the updates you saw today. So just to do a recap, one, we have the IBM quantum system two. To me, it looks amazing. It's modular and it'll keep scaling us to the future. And the first one already has three herons inside it, two, heron, our performant system is already making utility workloads run five times better. I can't wait to see those results.
[43:24] Three, with Condor, we've solved single qubit scaling. four with support for parallelization. With the execution modes, batch and iterative workloads are going to run faster. Five, We've simplified quantum algorithms and introduced this concept called Qiskit patterns. Six We've created quantum Serverless and it's in beta. This means it's a much more stable and I'm really looking forward to you seeing it. But all of these were on our roadmap. I hope you also like the additions that were not on our roadmap that we've revealed to you.
[43:56] The first being Qiskit is going to become Qiskit 1.0 And it'll be fast and stable. As Dario mentioned, it's been a wild year in A.I. and we're bringing the full force of A.I. to simplify how quantum computers will be run. And we're introducing ways that we can use, like natural language to create quantum code. To me, this is a big deal. Nine We've actually introduced and showing how we can use A.I. to power coming up with better circuits, and this would be rolled out as a transpiler service in Alpha.
[44:31] You can start using it now. And finally, the 10th announcement is that we now have a ten-year roadmap for all of you that know how much we take the roadmap seriously, Putting a ten-year roadmap is a big deal and putting error correction on it is our commitment to how we're going to bring useful quantum computing to the world. And with that I would like to end by saying we're in the era of quantum utility, just like we said back in 2021, 2023 would be a big year.
[45:02] We have a clear, detailed roadmap for scaling up quantum computing. We've hit all our milestones and now we have systems capable of exploring problems beyond brute force classical computing. Now, I hope you enjoy the rest of the sessions today where you'll see a deep dive on the things I talked about and much, much, much more. The year of quantum utility is here, and I hope you'll come along with us on this journey. Thank you.
在Vidleaf工作台中打开
搜索转录文本，选择行，复制带时间戳的引文，翻译。
署名
IBM Research 的“主题演讲：进入量子实用性时代”（https://www.youtube.com/@ibmresearch），根据 CC BY 3.0（https://creativecommons.org/licenses/by/3.0/）许可。源视频：https://www.youtube.com/watch?v=De2IlWji8Ck。本页面是视频的文本转录，添加了段落分隔和时间戳；创作者与 Vidleaf 无关，也不为其背书。Vidleaf 添加了 AI 标点和 AI 生成的摘要；原始措辞仍可在本页面上找到。
您是创作者或权利持有人吗？请求更正或移除：copyright@vidleaf.app（参见关于这些页面）。
最后更新于