↓跳过正文
没用知识库

人工智能扩展取决于连接

目录

title: “人工智能扩展取决于连接” source: “https://www.youtube.com/watch?v=1Er0n1pJUgg" video_id: “1Er0n1pJUgg” channel: “COMPUTEX TAIPEI” published: “2026-06-02” duration: “01:18:00” language: “en” transcript_source: “speech_to_text” created: “2026-06-27” tags:

  • youtube
  • 视频笔记
  • 人工智能基础设施
  • 光互连
  • Marvell

人工智能扩展取决于连接
#

[!info] 视频信息

  • 原标题：COMPUTEX 2026 CEO Keynote: Marvell
  • 演讲主题：The Future of AI Scaling Depends on Connectivity
  • 主讲人：Matt Murphy（Marvell 董事长兼 CEO）
  • 特邀嘉宾：Jensen Huang（NVIDIA 创始人兼 CEO）、Tien Wu（日月光投控 COO）
  • 频道：COMPUTEX TAIPEI
  • 发布日期：2026-06-02
  • 时长：01:18:00
  • 有效对白：约 00:31:00–01:17:42
  • 原视频：YouTube
  • 文本来源：无可用字幕；使用本地 Whisper small 模型转写
  • 校订说明：视频前约 31 分钟为无对白候场，已删除静音导致的周期性 “You” 幻听；机器翻译已统一关键术语，仍可能存在人名和口语细节误听。

结构化中文摘要
#

  • 核心主题：Marvell 认为，AI 基础设施在先后突破计算与内存瓶颈后，下一阶段的决定性约束将转向连接。AI 工作负载正从单芯片、单机架扩展到数据中心园区乃至跨地域系统，必须以更高带宽、更低时延和更低功耗连接计算、内存、存储和网络资源。
  • Marvell 的战略转型：Matt Murphy 回顾公司从“制程快速跟随者”转向领先制程的过程，包括跳过 7nm、从 14/16nm 直接迈向 5nm，以及通过收购和内部整合建立定制计算、CXL 内存、存储控制器、SerDes、交换芯片、光学 DSP、硅光子和先进封装能力。其核心定位是覆盖“处理、存储、移动数据”，其中数据移动与高速连接已成为数据中心业务的主要收入来源。
  • 瓶颈迁移：生成式 AI、混合专家模型和智能体式 AI 会显著增加跨加速器、跨机架和跨数据中心的数据流量。演讲把连接描述为继 GPU 计算和高带宽内存之后的新“稀缺资源”；超大规模客户因此开始重构网络，把扩展 AI 首先视为连接问题。
  • 与 NVIDIA 的协作：Jensen Huang 将智能体计算描述为分解式、分布式、异构的计算模式，需要计算、长期记忆、短期工作记忆和编排组件之间持续通信。双方围绕 NVLink Fusion、定制芯片、硅光子与光互连合作；NVIDIA 对 Marvell 投资 20 亿美元。Jensen 的判断是：在带宽和距离允许时继续使用成本更低的铜互连，超过物理极限后再转向光学，未来 5–10 年两者会长期并存。
  • 从毫米到公里的连接技术栈：封装内使用 die-to-die 接口和先进封装；机架内的 scale-up 网络目前主要依赖高速铜 SerDes；机架间和数据中心内部 scale-out 网络依赖短距光模块；园区和跨数据中心连接则使用相干光学 DSP。Marvell 以“每一跳、每一距离”概括其端到端产品组合。
  • “铜墙”正在向机架内移动：信号带宽越高，铜缆可达距离越短。演讲举例称，200Gbps/通道铜缆约能覆盖 2.5 米，而升级到 400Gbps/通道后将难以完整覆盖机架布线，因此光互连会从机架间逐步进入机架内部。每向更短距离推进一级，潜在连接数量可能增加一个数量级。
  • 产品与工程路径：演讲展示了 1.6Tbps、2nm 相干光学方案、Teralynx T100 102.4Tbps 交换芯片，以及把光引擎放到交换芯片封装周围的共封装光学（CPO）系统。CPO 通过缩短电连接、让光纤直接进入封装来改善带宽密度与功耗，但同时要求领先 CMOS、硅光子、模拟器件、封装和制造良率协同。
  • 制造生态系统：Tien Wu 强调，先进封装与光互连量产需要提前多年投入产能、设备和人才。台湾的工程人才密度、供应链集群、客户距离和长期资本投入形成难以快速复制的生态优势；Marvell 与日月光等伙伴之间的承诺与产能投资是技术从演示走向大规模部署的关键。
  • 十年愿景：当光互连削弱距离约束，数据中心可从固定服务器与机架演变为计算池、内存池和网络池，按工作负载动态组合。最终目标是让数百万资源像一台机器协同工作，系统架构由模型需求决定，而不再由连接距离决定。
  • 风险与不确定性：这是供应商主题演讲，增长率、性能领先和市场规模等表述具有营销属性，并非独立验证。光互连和 CPO 仍面临成本、功耗、热设计、可靠性、良率、供应链扩产和标准协同等挑战；铜互连不会立即消失，实际迁移速度取决于不同距离与工作负载的经济性。
  • Takeaway：AI 基础设施竞争正在从“谁拥有更多计算芯片”扩展为“谁能以可承受的功耗和成本把更多计算、内存与存储连成一个系统”。Marvell 的押注是：连接将成为 AI 时代与计算同等重要的平台层，而光学技术是突破规模边界的主要路径。

中文全文译文
#

00:31:00
#

首先，我们建立了一个令人难以置信的技术平台。而这一切都始于先进的工艺节点。实际上，我们做出的最重要的决定之一就是成为制程节点领导者。现在，Marvell、Cavium 以及我们收购的一些公司都是快速追随者，这意味着您所做的一切都落后一两个节点。这很大程度上是由于规模不够造成的。这通常就是人们这样做的原因。但当我们整合这些业务时，我们决定，如果我们要在数据基础设施方面竞争，我们就必须处于绝对领先优势。别无选择。现在，这是一个鲜为人知的事实。 Marvell 完全跳过了 7 纳米。当时我们进行了全节点跳跃，从 14 纳米和 16 纳米一直到 5 纳米。我的意思是，没有人这样做。没有人愿意冒这样的风险或打赌。但我们做到了，而且成功了。事实上，它运作得非常好，完美无缺。我们的工程团队在执行这一转变方面表现出色。因此，在 2020 年初，我们发布了第一个世界一流的 IP 平台，配有芯片间接口、定制 SRAM、高速 SerDes 等。现在，SerDes 是我们如何构建这个平台的一个很好的例子。它将 Marvell 自身的核心工程实力与 Avera、Aquancia、Infi 等公司的杰出人才相结合。如今，Marvell 已拥有 1,500 名员工，在工程规模和能力方面堪称首屈一指。因此，为了支持我们使命的过程数据部分，我们构建了一流的定制计算平台，并与世界领先的超大规模企业建立了深入的合作关系。这项业务对我们来说一直表现得很好。在存储数据方面，我们构建了存储控制器、基于 CXL 的内存池和近内存计算的完整产品组合。但这才是我们真正全力以赴的地方。那就是数据移动。这就是我们的高速连接产品组合的所在。当你看看 Marvell 今天的数据中心业务时，我们的绝大多数收入实际上来自连接，从数据中心内部的高速光学互连到数据中心之间的长距离光学，再到高速交换基础设施。所以今天，我们是无可争议的连接领导者。当你退后一步看看我们所打造的产品以及市场最终的走向时，我认为结果不言自明。早在 2016 年，Marvell 就是一家市值 23 亿美元的公司。事实上，当我们开始转型时，在前五年，我们的公司收入就翻了一番，达到 45 亿美元。在接下来的五年里，我们的增长加速了。根据华尔街的一致估计，今年我们的收入将在过去五年中增长约 2 倍半，达到 114 亿美元。但最近几年，如果你真正深入研究，Marvell 的年增长率约为 40%。因此，过去几年的增长速度实际上正在加快。因此，此时此刻，Marvell 已经开始投入竞争了。根据我们上周在财报电话会议上分享的前景，一致预期已经出现，他们预计我们明年将实现 164 亿美元的收入。正如我之前所说，当我们开始这一旅程时，数据中心仅占我们收入的不到 10%。我们把整个农场都押在了上面。上个季度，它占我们收入的 75% 以上，并且增长非常迅速。这是一家与我们以前非常不同的公司，而且这篇论文基本上已经完成了。但我们仍处于早期阶段

00:34:42
#

的基础设施建设。下一阶段就在我们面前。我们将有一组不同的要求，这让我们回到了连接性。所以过去几年，随着人工智能对基础设施提出新的需求，我们看到行业解决了一个又一个重大瓶颈。首先，是计算。该行业需要更多的计算来支持现代人工智能。 NVIDIA 在引领这场革命方面做出了令人难以置信的贡献。并一路成为全球第一家市值达到5万亿美元的公司。祝贺 Jensen 和这里的整个团队。这真是一个非凡、非凡的结果。接下来是内存瓶颈。较大的模型需要大量的内存和带宽。存储器公司现在正在积极扩展规模以满足这一需求。就在最近，我们看到该市场出现了三个新的 1 万亿美元市值公司。但瓶颈再次发生转变。现在，连接性将定义基础设施的限制，就像计算和内存一样。该行业将团结起来应对这一挑战。现在，这不仅仅是我一个人这么说。这是我们从最大的客户那里听到的。全球最大的超大规模企业现在正在重新构想其整个网络架构。他们认识到，扩展人工智能基础设施现在首先是一个连接挑战。随着推理模型、专家架构的混合、生成式人工智能，这一切都在不断发展，更多的数据必须在基础设施上移动，从而需要更高的带宽和更低的延迟。随着工作负载不再适合一个数据中心，你猜怎么着？他们需要建造更大的数据中心，或者充满数据中心的整个园区，以及它们之间的所有高速连接。因此，连接性成为扩展计算的关键推动因素。我们的客户越来越认识到光学是前进的方向。他们正在寻求像 Marvell 这样的领导者来帮助他们大规模地建立更大、更快的网络。因此，当你纵观整个半导体行业，看看支持这种基础设施建设的领先公司时，很明显我们每个人都专注于基础设施的不同部分。这体现在收入组合中。有些公司是计算优先的。这意味着他们的绝大多数收入都与计算、某些类型的连接或大部分计算相关。这显然是堆栈的关键部分，这就是为什么我们在这个组中拥有价值数万亿美元以上的公司。这些公司专注于内存领域，而且目前所有市值达数万亿美元的公司。真是难以置信。然后就是 Marvell。我们不同。我们是独一无二的。如今，我们的绝大多数收入实际上来自连接。因此，我们围绕数据移动建立了这家公司。如今，我们的绝大多数收入实际上来自连接。现在，这涵盖了广泛的技术。您可以看到，甚至我们收入中来自计算的部分从根本上来说也是因为客户将我们的连接嵌入到他们的计算引擎中。因此，这为我们提供了对正在发生的这些技术转型的独特立场和视角。它创造了我们与生态系统其他部分之间非常不同的关系。我们与计算公司深度合作。我们与存储器公司深度合作。这些都是非常具有战略意义的关系。

00:38:23
#

在很多方面，我们都是该行业的瑞士，我们与每个人合作。现在，Marvell 在这个生态系统中发挥的作用的最佳例子之一就是最近宣布与 NVIDIA 建立的战略合作伙伴关系。作为我们几个月前发布的公告的一部分，NVIDIA 向 Marvell 投资了 20 亿美元。我们现在正在多个维度扩展我们的合作伙伴关系，包括光学、光子学、NVLink 融合。我很高兴地宣布詹森本人今天也来到这里。他将和我一起上台。我们将花几分钟讨论合作伙伴关系。我们将看看人工智能基础设施将走向何方。那么，请允许我欢迎黄仁勋登上舞台。哇！音乐 怎么样，詹森？你好吗？孩子，这是一个巨大的舞台。那跑了很远的路。你做得好吗？是的。那是一个巨大的舞台。那跑了很远的路。你气喘吁吁吗？你还好吗？我知道。让我们点火吧。很高兴见到你。达贾豪。就这样吧。恭喜 GTC 昨天取得了精彩的开局。你们这周要去参加比赛了。谢谢。谢谢。听着，也许你听到了我刚才说的一些话。所以我们今天讨论的是连接性。女士们先生们，下一个万亿美元公司。哇！那会很令人兴奋。让我们一起来做吧。但这实际上一切都始于当今更广泛的人工智能基础设施中发生的事情。那么，您如何看待从大局角度来看，我们正处于这个非凡的时刻，客户的需求达到了顶峰。您如何看待连接性在所需的互连中发挥的作用？是的，那真的很棒。你知道，昨天我说过有用的人工智能已经到来。这就是您的需求激增的原因。这就是为什么我的需求如此强烈的原因。是的。这种使之成为可能的新计算模式称为代理。这些代理有一个特定的计算平台，一种分解的、分布式的计算模式。当您将一个计算问题分解为许多部分并分布在整个数据中心时，连接性是必需的。这就是马特表现如此出色的原因。这就是 Marvell 如此重要的原因。我们已经分布式和分解计算，以便它在这些巨大的集群上运行，这样我们就可以聚合我们拥有的总计算、总内存和总带宽。而让这一切成为可能的是连通性。是的。我的意思是，我们正在看到它。然后当你思考的时候。这就是为什么他们将成为下一个万亿美元公司。我们还有一些工作要做，但我们已经上路了。我们正在路上。谢谢你，詹森。好吧，我们来谈谈规模。我们过去谈论的是连接数十个 GPU、CPU 和 XPU，现在是数千个，现在可能是数百万个。因此，当你扩展计算和扩展连接时，我认为我们讨论了诸如代理之类的事情。但您如何看待跨数据中心、数据中心内的这一点？您如何看待连接性在整体上发挥的作用？您认为哪些技术是重要的？嗯，在其基础上，代理计算模式需要一个编排系统，该系统允许大型语言模型、计算能够思考、推理并提出计划。但它也必须使用工具，你知道，浏览互联网、访问内存、访问

00:42:04
#

长期记忆，处理短期工作记忆。所有这些都需要大量的连接。但也确实如此。如果你看看我们介绍 Vera Rubin 的方式，Hopper 就是为训练而设计的。 Grace Blackwell 推出了 NVLink 72，这是我们的第一个纵向扩展互连架构。它引入了对非常大的 MOE 模型以及非常大的专家模型的混合进行极快速推理的想法。Grace Blackwell（Grace Blackwell）是为了推断。 Vera Rubin 是运行代理，这就是为什么 Vera Rubin 系统当然包括 Vera Rubin 思考 AI，但它也包括用于编排的 Vera CPU。它包括用于存储加速和管理长期内存的 Vera CX。我对这些系统的看法是，有时 CSP 可能想要设计自己的定制芯片。我们之间还就 NVLink Fusion 进行了合作，这使得您可以使用相同的系统架构。Vera Rubin (Vera Rubin) 内部还有一些半定制芯片、大量互连硅光子学和光学技术等。从本质上讲，我们可以创建一个分解的、分布式的、异构的数据中心。这就是伟大的想法。但它们的系统架构是相同的。他们的网络技术可以利用很多 NVIDIA 堆栈。 CPU 可能是 Vera。但它可以充分利用您的堆栈。因此，NVLink Fusion 就是采用 NVIDIA 的技术和我们的平台、Marvell 的技术和工厂，然后我们将其融合。这就是为什么它被称为融合。是的。不，我想，你知道，我想到了我们长期以来一直合作的伙伴关系。我想用投资来纪念它，我们真的很感激。我认为这对我们来说意义重大。我们很荣幸拥有它。你知道，谁不喜欢赚钱呢？很高兴给予。自从你投资以来，它做得很好。男生。是的。当Jensen投资时，跟着他就可以了。跟着他就行了。把我所有的钱都给马特，然后看着他赚钱。这就是我每天都在做的事情。我喜欢那个。这就是我每天都在做的事情。但我认为你谈到的这些事情，我们已经实现了，NVLink Fusion，在光学方面的合作。我的意思是，我认为代理时代和现在的新平台非常适合。我的意思是，NVLink Fusion，我们几年前就有了这个想法，对吧？但我认为这有点超前于时代。现在，我想看看您是否同意，当您考虑您的平台类型以及我们的客户拥有的一些自定义网络和计算需求以及互操作和协同工作的能力和需求时，Marvell 和 NVIDIA 现在似乎是时候真正让我们的客户拥有他们所寻求的灵活性，并真正利用代理时代来共同扩展我们的平台。是的。你知道，归根结底，我确实认为，如果你除了 NVIDIA 之外什么都不买，那也没关系。好的。我的意思是，但如果您绝对必须设计自己的 ASIC，我们仍然很高兴 NVIDIA 位于该数据中心内。所以，你知道，你不必从我们这里购买所有东西。只需从我们这里购买一些东西即可。您知道，我们很高兴为您和客户提供支持。所以，我认为，在我们两个人之间，你可以受益于一个通用的、非常高效率的系统，这个系统是从Vera Rubin（Vera Rubin）开始构建的。

00:45:48
#

但任何您想要扩展至专业化的领域，您都可以这样做，这就是为什么您的客户 NVIDIA 位于 AWS 中，Marvell 位于 AWS 中，NVIDIA 位于所有云中，很高兴看到 Marvell 扩展到所有这些不同的云中。是的。伟大的。谢谢。最后一张送给你。给我留点事吧，知道吗？看，我们现在是您最好的销售人员。我们是您最好的销售人员。一起工作。最后一个问题给你。我的很多演讲都是关于一些过渡，特别是当您进入从铜缆到光纤的机架内部时。显然这不会是一零。这需要时间，而且有不同的用例。但您现在如何看待这种情况呢？从铜到光学的转变，也许我们也可以如何在这方面合作？好吧，我们应该尽可能多地使用铜，尽可能长时间地使用铜，但铜有其局限性。铜缆在带宽和距离方面都有其局限性。因此最终，正确的策略是尽可能扩大铜的规模。之后，您可以通过光学进一步扩大规模，通过光学扩大规模，并通过光学扩大规模。因此，您可以在任何需要的地方使用光互连。只要有可能，就使用铜。所以我认为这种交叉将会持续很长一段时间。最重要的是，在未来 5 年、10 年里，我们将使用大量的铜，并且将使用大量的光互连。因此，这些数据中心现在是基础设施的一部分。而我之所以说现在AI有用，有用的AI已经到来，是因为现在AI有利可图，token（词元）也有利可图。当token（词元）生产有利可图时，每个人都想生产更多token（词元），这就是为什么 Marvell 的需求如此高，我们的需求如此高，因为每个人都想生产更多token（词元），因为它被AI 智能体到处使用。绝对地。我想你谈到了我稍后将介绍的很多事情。如果你想完成我的演示的其余部分，你可以。是的，女士们先生们，这些漂亮的幻灯片，你们知道。就坐在那里吧。我会的。你从这里拿走它。好吧，黄仁勋。好吧，很高兴见到你，兄弟。好吧，保重。好吧，你们，谢谢你们。谢谢你，詹森。再见马维尔。好吧，优秀，优秀。 Jensen 一如既往地来到这里，真是太有趣了。好的，我们已经讨论了很多关于连接的问题。 Jensen 和我刚刚介绍了这一点，所以现在让我们深入探讨吧？让我们再深入一层。因此，人工智能基础设施跨越各个距离。它的跨度从数据中心之间的数百甚至数千公里到封装内的毫米级。每一个距离都需要不同的解决方案。这是不同的技术、不同的工程团队。这是一群完全不同的专家。在许多情况下，这是一条不同的供应链。所以这些不是同一问题的变体。这里面临的是根本不同的工程挑战，这就是我们接下来要讨论的内容。好吧，让我们从最长的距离开始吧。詹森提到了这一点。这是跨规模的，将数据中心连接在一起。现在，每个主要的云提供商在全球都拥有数百个数据中心。所有这些数据中心都需要相互通信。这从根本上来说是一个长距离连接问题。我们谈论的是跨越数百甚至数千公里的链路。

00:49:22
#

这需要非常具体、非常复杂的技术，称为相干调制。它的核心是专用数字信号处理器 DSP。它旨在以极高的可靠性在很长的距离内跨光纤电缆传输大量数据。世界上只有少数几家公司能够制造这些相干 DSP，我们就是其中之一。 Marvell 多年来一直是该技术的领导者。我们构建的光学模块包含驱动和调制激光器以及长距离传输数据所需的所有电子设备。所以我的口袋里有一个小节目和讲述。这次没有举起筹码。我拿着一个光模块。这是我们的相干光模块之一。这是一项极其复杂的工程。在 Marvell，我们构建了整个模块。这是我们的。它包括先进节点 CMOS DSP。它是最复杂的芯片之一，仅是我们在 Marvell 设计的 DSP。但它也融入了我们的第四代硅光子技术。那是在里面。十年来，我们一直在硅光子学领域开发生产技术。它还包括我们自己设计的宽带模拟组件，该组件采用硅锗设计。因此，Marvell 率先推出了这项技术，十年前从100Gbps 开始，然后发展到 400 GB，现在出货量达到 800Gbps。今年晚些时候，我们将采样世界上第一个 1.6Tbps、2nm相干光学解决方案。这来得正是时候。对带宽的需求从未如此之大。好吧，现在让我们进入数据中心内部。因此，这些数据中心可能非常大，跨越数百米，并且包含一个又一个计算服务器机架。现在，每个机架的顶部通常都有一个交换机，服务器连接到该交换机。这些机架级交换机连接到主干，然后连接到核心交换机。这将创建将整个数据中心连接在一起的网络结构。所有这些都是通过光纤电缆连接的。现在，光学模块再次通过这些光纤电缆驱动数据传输。但这次的调制方案不同。我们没有使用相干技术，而是使用了一种功率更加优化的调制技术，称为 PAM4。因此，这部分市场的两个关键半导体解决方案是模块内部的 PAM4 芯片组，以及将数据中心连接在一起的云交换基础设施。 Marvell 两者均生产。从PAM4芯片组开始，我们构建了业界领先的PAM4 DSP解决方案。还有它们周围的高速模拟组件，包括跨阻放大器或 TIA 和激光驱动器。顺便说一句，这些也是硅锗。我们引领行业完成了 PAM 技术的每一次重大转型，从 50 GB、100 GB、200 GB、400 GB 和 800Gbps 开始。去年，我们开始升级 Marvell 的 1.6 T 3 纳米 PAM4 解决方案，引领行业向 1.6 T 连接过渡。现在，对于以太网交换，Marvell 拥有同样完整的产品组合，从 12.8 太比特到 51.2 太比特。今天，我们发布了全新的 100T 以太网交换机，专为 AI 数据中心设计，具有业界最低的功耗。 Computex 特别公告。我们等待着。因此，将所有这些放在一起，我们为数据中心内的连接提供完整的解决方案。现在让我们进入机架内部。

00:53:23
#

这里的目标是以完整的任意配置将尽可能多的处理器连接在一起。换句话说，每个处理器都可以直接与其他处理器通信。詹森谈到了这一点。第一家将该架构推向市场的公司是 NVIDIA 的 NVL 72，该名称因在单个机架内连接在一起的 72 个 GPU 而得名。这需要一种完全不同的连接方法。这是一种不同类别的交换机，能够通过机架内的铜背板驱动超高速信号。所以今天，这不是光学的领域。这是铜的领域。这里的核心区别在于电气 SerDes 技术，而不是光学技术。现在，Marvell 还拥有200Gbps的领先电气 SerDes。我们在过去几年中已经展示了未来400Gbps的速度。因此，我们正在将 SerDes 技术构建到我们的客户、定制芯片及其 XPU 以及我们自己的扩展交换机中。好吧，现在让我们深入了解包的内部。这里我们不再谈论米。我们谈论的是毫米。您实际上可能不会认为这是一个连接挑战，但如今，大多数先进芯片的封装内都有多个小芯片。因此，当您拥有两个半 D 或 3D 封装时，它实际上从根本上来说是一种连接技术。它允许这些小芯片在封装内非常靠近地放置在一起，并通过超高速、短距离、芯片间接口进行通信。 Marvell 拥有领先的芯片到芯片 SerDes 和先进封装方面的领先能力，使我们的客户能够构建一些业内最复杂、最独特的多芯片芯片。正如您所看到的，人工智能数据中心的连接需要非常广泛的技术组合。每个距离都需要非常不同的解决方案。 Marvell 拥有业界最完整的产品组合，从毫米到公里、每跳、每距离。事实证明，将所有这些功能集中在一个屋檐下是很不寻常的。它是独一无二的。当我们去竞争时，通常我们在每个类别中跨越不同的距离与一组不同的公司竞争。但这就是我们的独特之处。我们是一站式商店。我们是整个连接堆栈的领导者。这给我们带来了该行业面临的下一个重大挑战。因此，当我在最后几张幻灯片中描述这些不同的解决方案时，您可能会注意到，对于不同的距离有不同的解决方案，并且今天的一些连接是光学的，而今天的一些连接是电气的。它实际上是由距离定义的。因此，该图表左侧的联系是光学辩论。这意味着他们使用光纤电缆来传输光，电缆两侧都有复杂的电子设备来驱动和调制传输光的激光器。右侧的连接是电气连接。因此，他们使用铜电缆或仅印刷在电路板上的铜迹线，甚至封装内的微观铜布线。所以这里的共同主题是铜。在中间，你看到墙，铜墙。墙壁是由铜缆传输信号的最远距离定义的。因此，您必须先使用光纤连接。

00:56:53
#

所以这是一个重要的区别，因为铜很简单并且成本低廉。正如詹森所说，你想尽可能长时间地使用它。非常实用。但光互连就比较复杂了。它需要激光、光子学和复杂的电子学。所以这是一个更大的提升，但它是需要的。还有铜墙，我今天要告诉你们的是，它即将移动。它会再次移动并接管机架本身。因此，这导致了光学行业需求的爆炸式增长。在此过程中，令人难以置信的复杂工程挑战不断出现。那么为什么会发生这种情况呢？因此，这不仅仅是某人的偏好。这是物理学。信号通过铜缆传输的距离与带宽成反比。因此，每次将带宽加倍时，就必须将距离减半。如今，世界上速度最高的生产系统每通道的运行速度为200Gbps，仅举个例子。因此，在该带宽下，电缆长度限制为大约 2.5 米。相比之下，现在以 100G运行的系统可以使用大约 5 米的电缆。并且机架的高度约为2米。因此，一旦考虑到机架内的所有布线，2.5 米就达到了极限。因此，当我们迁移到 400G 时，我们无法再用铜完全连接机架。所以墙正在移动，而且现在正在移动。展望未来，甚至机架内的连接也将变成光学连接，整个行业都知道这即将到来。因此，我们一直在为这一刻做准备，不仅仅是 Marvell，整个行业也是如此。顺便说一下，你在台湾看到了这一点，以及供应链和正在发生的增长。这样做的后果实际上是巨大的，因为墙每向右移动一步，你所拥有的连接数量就会至少增加一个数量级。因此，正如我所提到的，它造成了需求的爆炸式增长，光学供应链需要大规模扩大规模并做好准备。我们以前看过这部电影，好吗？我的意思是，20 年前，我记得这一年，最先进的数据中心速度是每秒 10 吉比特。那是 10 场演出。我们在整个数据中心都使用铜缆。当时的光学只能用于非常非常长的距离。它本质上就像一种电信技术。但当隔离墙移动时，光学行业实际上迎接了挑战。如今，世界上所有的超大规模数据中心都采用光纤连接。正如我们在这一转变中看到的那样，它确实需要新的解决方案。您无法使用同样耗电的电信方法，这就是 Pam 4 的用武之地。它针对功率、密度和覆盖范围以及专门针对数据中心内部的要求进行了优化。 Marvell 是该领域的关键创新者之一。因此，随着光互连在机架内移动，我们将看到同样的创新浪潮。这就是所谓的共封装光互连或 CPO 技术。你现在听到很多关于这个的事情。我要告诉你更多。 CPO 是一种我们将光学连接一直引入封装本身的技术。就在计算旁边，无论是定制计算还是交换芯片。我们用 CPO 解决的根本挑战是密度和功率。现在，请记住，机架内部的连接数量大约是机架之间连接数量的 10 倍。

01:00:11
#

因此，如果您只是尝试在数据中心的各个机架上使用相同的光学技术，您将没有足够的电源，也没有足够的物理空间。您无法像现在一样安装所有这些标准光学模块和电缆。它就是行不通。这是不可能的。因此，业界一直在发明这种共封装光学概念，将光纤直接放入封装中。它将通过光纤直接驱动信号的电子设备与定制计算或交换芯片紧密耦合。所以这是一个巨大的变化。这很困难，因为你要结合芯片行业中一些最先进的技术。领先的 CMOS、硅光子学、先进封装、光学互连，全部在小型紧密集成系统中制造。因此，复杂性非常高，但这是继续扩展带宽并克服我在铜缆中谈到的这一限制，同时降低功耗的唯一方法。这就是该行业的发展方向。这也是 Marvell 十多年来在硅光子、光学 DSP、所有相关模拟宽带组件以及实现这一目标所需的所有先进封装方面进行投资的原因之一。它实际上需要在 CPO 中整合在一起。所以这不是什么未来主义的事情，伙计们。现在正在发生。事实上，我今天带来了几个 Marvell 示例。那么让我们快速演示一下并讲述一下。好的。好的，这里有一个传统的以太网交换机。这是我们今天发布的 102.4Tbps Teralynx T100 交换机。你们是第一个真正看到它的人。每个人都在房间里。您可以看到板中间的开关。 PCB 内的铜迹线将信号传送至此处的前面板。这是所有光学模块插入的地方。现在让我们移到这里。这是一个基于 CPO 的交换机。现在请注意，中间仍然有开关芯片。它位于封装芯片的中心。在本例中，这是我们的 51.2T 交换机。四周都是16个3.2T光学引擎。所以 16 乘以 3.2，得到 51.2。因此，光纤现在直接连接到这些发动机上。它不是到前面板。所以我们已经完全消除了 PCB 上的铜迹线。光线直接从封装中射出。这是一个非常非常复杂的工程。今天能够展示这一点真是太酷了。好的，共封装光互连已经出现，该行业正在扩大规模以应对挑战。正如我们一次又一次看到的那样，每次我们遇到物理障碍时，我们都会通过技术和创新来突破它。在这种情况下，通过用光纤代替铜，因为与在铜线上传播的电子不同，光子可以携带的距离（通过玻璃的信号）在很大程度上与带宽无关。因此，随着人工智能基础设施需要更高的传输速度，并需要扩展到更大、更复杂的系统，跨越现在编织在一起的数百万个处理器，而不是数千或数百个，光导将日益成为事实上的解决方案。因此，真正的问题是，如何在整个人工智能基础设施堆栈中提供光互连？需要什么？嗯，首先要认识到整个数据中心没有单一的技术。事情不是这样的。不存在一刀切的解决方案。没有捷径。没有简单的方法可以到达这里。

01:03:42
#

没有单一的架构调制方案、频段、独特的技术可以完成这一切。天下没有免费的午餐。这就是为什么我们在各个距离上寻找一系列不同的独特光路来到达这里。我在这里拥有的每一项技术都针对不同的设计点进行了优化。每一项都支持基础设施的关键部分，并满足整个堆栈对密度、带宽、功率和集成的不同要求。因此，如果光互连是构建下一代 AI 基础设施的底层技术，那么 Marvell 正在构建业内最广泛、最深入的产品组合。但没有一家公司能够独自实现这一转变。正如詹森之前谈到的，需要一个生态系统才能到达这里。正如我所说，技术创新是伟大的。这是挑战的一部分，但不是全部。但大规模展示这一点确实很重要。此时，如果您只是操作 PowerPoint 或演示、POC、新闻稿，那么它不会帮助您实现目标。客户现在需要现成的解决方案。他们是可靠的。它们需要可制造并准备好大规模部署。 Marvell 和我们的生态系统合作伙伴长期以来一直在这样做。我们已经售出了数亿个 DSP。我们已经在现场积累了数百亿设备小时的数据。这种经验很重要，因为这些产品不仅必须在实验室中运行，而且必须在世界上最大的数据中心中以非常高的容量和非常可靠的状态运行多年。因此，这需要对制造生态系统进行提前投资。您必须在市场到来之前建立供应链基础设施的能力。这就是生态系统如此重要的原因。顺便说一句，这在台湾这里很重要。现在，日月光半导体（ASE） 是 Marvell 在这一历程中最重要的合作伙伴之一。现在ASE是全球领先的半导体制造公司之一。他们在亚洲乃至全球拥有超过 100,000 名员工，拥有数十年帮助实现半导体行业几乎所有重大技术转型的记录。现在领导 ASE 度过这一转型期的是我非常熟悉的人。他花了超过 25 年的时间帮助塑造公司和行业。今天我很高兴迎来我的下一位演讲嘉宾，他是 ASE 首席执行官Tien Wu 博士。Tien Wu 博士，请跟我一起上台吧。谢谢。你好。Tien，你好吗？感谢您邀请我来到 COMPUTEX。很高兴见到你。很荣幸您能与我们一起登上舞台。嗯，这是我的荣幸。瞧，我们已经合作很长时间了。当我成为首席执行官时，我们有一系列的抱负。我们与很多供应商进行了交谈。甚至在我成为最成功的人之前我就认识你了。我在担任 Marvell 首席执行官之前就认识您了。当我回到 Maxim 担任高管时，我们在那里一起工作。但有时人们没有意识到，它也可能向观众解释的部分内容是，作为这个生态系统的关键供应商，你必须下注。你必须在与你合作的公司上下注。你必须押注于你认为谁会成功。我们非常感谢 ASE 很早就押注于 Marvell。我们已经看到基于此的巨大成功。

01:07:24
#

但我只是好奇您能否分享一下您对 Marvell 的看法、您的思维过程以及我们今天在共同旅程中的进展。Tien，很高兴听你分享。好吧，我认为描述这一点的最好方式是一个渐进的过程。第一个决定并不困难。 Marvell 这家无晶圆厂公司拥有非常好的声誉，经历了很多转型。因此，Marvell 的业绩记录已经存在。当您加入时，产品集有点过时了。所以第一个是商业模式需要调整。台湾日月光（ASE）属于制造业。因此，我们正在寻找赌注，不仅押注于您的成功，我们还押注于能够为下一代架构和技术要求提供见解的人。如你所知，这家台湾公司提前10年投资基础设施与资本支出，赌注很大。我们只指望我们投入的容量、将需要的以及将被利用的容量。这就是我们赚钱的方式。因此，押注于我们相信能够让我们对未来有很好洞察力的公司变得非常重要。所以一开始就是这么决定的。在过去的十年里，我真的很高兴。我们谈论的一切，都是十年前的梦想。这是一个梦。今天我们要发货。您刚刚提到未来几年的增长率将达到 40%。我相信你会成为那样的人。所以我们现在正忙着为您准备容量。我们也很高兴在过去10年里，我们进行了大量的战略讨论。您向我们做出承诺，我们为您进行投资，随着时间的推移，我们将为您生产更多的零件。我认为这对于如何做出这个决定来说确实是一个简短的故事。是的，不，这是一个很棒的故事。也许再给你一份。台湾这里的生态系统非常独特。正如您所说，需要十年的投资才能真正看到回报。这里正在发生这样的力量。你如何向这里的人们描述它？而且世界各地都有很多人在观看。那么是什么让这一切成为可能呢？为什么它是独一无二的？那么，是什么使得在世界其他地方难以复制这一点呢？但与此同时，还有全球化。那么我们如何看待这些动态呢？我认为这会是一件有趣的事情。我认为你问这个问题的原因是世界各地存在很多竞争力量和不确定性。所以我认为我的信念是任何企业都需要有远见以及长期的价值一致性。因此，在商业模式上，整个台湾行业都是建立在产能利用率以及领先的创新和技术投资的基础上的。这就是台湾的价值。因此，与商业模式相符的优秀公司或特定 IDM 公司。其之下将是规模经济。台湾积累了40年的基础，从PC过渡到无线、到移动计算、到数据中心，现在我们进入HPC。使40年的经验积累了35万名半导体员工。还积累了110万高科技员工，其中很多人都在这里。结合规模经济和集群效率，这种经验变得极其宝贵。因此，我们考虑的是拥有多年经验的员工队伍。我们考虑集群效率。

01:11:25
#

当你考虑到我们已经投入的产能和规模经济时。但还有一件事，我认为台湾好或坏，我们的选择比美国等其他地区要少。所以大多数工程师出来后，他们几乎没有什么选择。半导体IT产业在台湾成为有吸引力的选择，但在其他地区却不一定。因此，将所有这些结合起来，我认为这个生态系统非常非常难以复制。这并非不可能，但需要数年时间。伟大的。嗯，非常感谢。我非常感谢这种伙伴关系。我们要去比赛了。Tien，谢谢你。Tien Wu，谢谢。好的。因此，正如我们所说，人工智能数据中心的未来都是光连接的基础设施。你也听到他这么说了，对吧？这将推动规模和制造所需的增长和创新浪潮。但这个不可避免的未来实际上是什么样子的呢？我的意思是，如果你退后一步，实际上不考虑现在，那么就考虑一下 10 年后的情况。这是一个许多铜连接都消失的世界。想象一下现在的数据传输在某些时候完全是光纤的世界。事实上，这是一个距离并不重要的世界。这是一个深刻的变化。如今的服务器、机架和整个数据中心架构都是围绕距离的限制而设计的。软件工作负载实际上也围绕这些相同的约束进行了优化。但如果距离不再重要怎么办？架构本身可能会发生怎样的变化？当基础设施不再受距离限制时，哪些新功能成为可能？那么让我们从机架中的扩展网络开始。正如我们之前讨论的，我们可以在此处以完整的任意对任意配置将尽可能多的处理器连接在一起。过去，该域的大小受到铜连接长度的限制，但对于光学距离来说并不重要。因此，现在我们可以将扩展域的大小从 72 或 144 xpus 或 GPU 更改为 1000 或更多，所有这些都通过光学互连。这对工作负载的影响是巨大的。如今，人工智能工作负载必须分解为适合扩展集群的更小的子问题。因为如今集群外部的通信速度较慢，带宽也低得多。但光互连系统可以管理更大数量级的工作负载。顺便说一句，事情并不止于此。当光纤连接位于服务器内部时会发生什么？现代人工智能服务器由一定数量的CPU、xpu、内存和网络接口组成。它们之所以位于同一个系统中是因为距离。 CPU 和 xpu 需要以非常非常高的带宽访问内存，这意味着它们需要在板上彼此相邻，并用铜迹线作为它们之间的连接。但在未来，这些连接都是光纤的，距离实际上并不重要。您可以想象一个完全分解的架构。一个系统中的 Xpu，另一个系统中的内存，另一个系统中的代理 CPU，这解锁了另一种可能性。在当今的系统中，CPU和xpus或gpu的比例是固定的。因此，必须在构建和部署系统时定义这些比率。但没有两个工作负载需要完全相同的比率。事实上，詹森谈到了这一点。

01:15:06
#

这意味着在任何给定时间，对于给定工作负载，计算机内存的某些部分可能未得到充分利用。那要花钱。但是，一旦我们将系统分解为单独的计算池和内存池，并且它们都通过光学互连，我们就可以动态分解专用系统，然后针对任何工作负载进行优化。想象一下未来的数据中心。全球光学互连的数据基础设施。我们今天所拥有的系统中的这些严格界限开始消失。现在计算可以池化，内存也可以池化。基础设施可以大规模动态组合。架构师第一次可以开始围绕模型的需求设计人工智能系统，而不是围绕互连的限制。这就是人工智能基础设施的发展方向。这是一个没有距离的数据中心。计算、内存、网络和光子学作为一个统一的系统运行。数据中心内的数百万资源可以像一台机器一样协同工作。根据工作负载的需求而不是连接的限制定义的架构。我们相信这是计算基础设施的下一个时代。 Marvell 正在帮助建立连接基础，使这一切成为可能。非常感谢您今天抽出时间。谢谢。谢谢。

视频原文全文（英语）
#

00:31:00
#

First of all, we have built an incredible technology platform. And it all starts with the advanced process node. It’s one of the most important decisions we made, actually, was to become a process node leader. Now, Marvell, Cavium, and some of the companies we acquired had all been fast followers, meaning you’re like a node or two behind on everything you do. And that’s largely a result of just not having enough scale. That’s usually why people do that. But as we integrated these businesses, we made the decision that if we’re going to compete in data infrastructure, we had to be at the absolute leading edge. No choice. Now, here’s a little known fact. Marvell skipped 7 nanometer completely. We made a full node jump at that time from 14 and 16 nanometer all the way to 5. I mean, nobody does this. Nobody takes that kind of a risk or a bet. But we did, and it worked. It worked really well, flawlessly, actually. Our engineering team did an outstanding job executing this transformation. So in early 2020, we released our first world-class IP platform, complete with die-to-die interfaces, custom SRAM, high-speed SIRDs, and more. Now, SIRDs is a good example of how we built this platform. It combined Marvell’s own core engineering strength with exceptional talent from Avera, Aquancia, Infi, and others. Now, today, that is a 1,500-person organization at Marvell, second to none in terms of engineering scale and capability. So to support the process data portion of our mission, we built a best-in-class custom compute platform, working in deep partnerships with the world’s leading hyperscalers. And that business has been doing very well for us. In store data, we built a whole portfolio of storage controllers, CXL-based memory poolers, and near-memory compute. But here’s where we really went all in. That was in data movement. And this is where our high-speed connectivity portfolio. And when you look at Marvell’s data center business today, the vast majority of our revenue actually comes from connectivity, from high-speed optical interconnect inside the data center to long-reach optics, between data centers, to high-speed switching infrastructure. So today, we are the undisputed connectivity leader. And when you step back and look at what we built, and where the market ultimately went, I think the results speak for themselves. So back in 2016, Marvell was a $2.3 billion company. As we embarked on the transformation, actually, in the first five years, we doubled the company, $4.5 billion in revenue. Over the next five years, our growth accelerated. And according to consensus estimates on Wall Street, for the current year we’re in, we’re set to grow about 2 and 1 half times over the last five years to $11.4 billion. But in the recent couple of years, if you actually drill down, Marvell has been growing like 40% a year. So the growth rate is actually accelerating in the last few years. So at this point, Marvell is off to the races. And based on the outlook that we shared in our earnings call last week, consensus estimates have come up, and they expect us now to deliver $16.4 billion in revenue next year. So as I said earlier, when we started this journey, data center represented less than 10% of our revenue. And we bet the farm on it. Last quarter, it was over 75% of our revenue and growing very rapidly. This is a very different company than we used to be, and the thesis is largely played out. But we’re still in the early innings

00:34:42
#

of this infrastructure buildout. The next phase is all in front of us. We’ll have a different set of requirements, and that brings us back to connectivity. So for the past several years, as AI has created new demands on the infrastructure, we’ve seen the industry solve one major bottleneck after another. And first, it was compute. And the industry needed dramatically more compute to enable modern AI. And NVIDIA did an incredible job leading that revolution. And along the way became the world’s first $5 trillion market cap company. Congratulations to Jensen and this whole team that’s here. It was just a phenomenal, phenomenal result. Next came the memory bottleneck. Larger models required enormous amounts of memory and bandwidth. And the memory companies are scaling aggressively now to meet that demand. And just recently, we’ve seen three new $1 trillion market cap companies emerge in that market. But the bottleneck is shifting again. Now, it’s connectivity that will define the limits of the infrastructure, just like with compute and memory. The industry will rally to meet this challenge. Now, this isn’t just me saying this. This is what we’re hearing from our largest customers. The world’s largest hyperscalers are now reimagining their entire network architectures. They recognize that scaling AI infrastructure is now, first and foremost, a connectivity challenge. As reasoning models, mixture of experts architectures, a genetic AI, it all continues to evolve, more data has to move across the infrastructure, demanding higher bandwidth, and lower latency. And as workloads no longer fit within one data center, guess what? They need to build larger data centers, or full campuses full of data centers, and all the high speed connectivity between them. Thus, the connectivity becomes a critical enabler of scaling compute. And increasingly, our customers recognize that optics is the way forward. And they’re looking to leaders like Marvell to help them build larger, faster networks and at scale. So when you look across the semiconductor industry, at the leading companies supporting this infrastructure build out, it becomes clear each of us is focused on a different part of the infrastructure. That shows up in the revenue mix. Some of the companies are compute first. It means the vast majority of their revenue is tied to compute, with some of the type of connectivity, or most of its compute. And it’s obviously a critical part of the stack, and that’s why we have several trillion dollar plus companies in this group. And you have the companies focused on memory, and again, all trillion dollar market cap companies at this point. It’s unbelievable. And then you have Marvell. We’re different. We’re unique. Today, the vast majority of our revenue actually comes from connectivity. So we built this company around data movement. And today, the vast majority of our revenue comes actually from connectivity. Now, this spans a broad range of technologies. And even the portion of our revenue that’s from compute, which you can see, is fundamentally because customers embed our connectivity in their compute engines. So this gives us a unique position and perspective on these technology transitions that are happening. It creates a very different relationship that we can have with the rest of the ecosystem. We partner deeply with the compute companies. We partner deeply with the memory companies. These are very strategic relationships.

00:38:23
#

And in many ways, we are the Switzerland of the industry, and we work with everybody. Now, one of the best examples of the role that Marvell plays in this ecosystem is the recently announced strategic partnership in expansion with NVIDIA. And as part of this announcement that we made a few months back, NVIDIA invested $2 billion into Marvell. And we’re expanding our partnership now across multiple dimensions, including optics, botanics, NVLink fusion. And I’m thrilled to announce that Jensen himself is here today. He’s going to join me on stage. We’re going to spend a few minutes chatting about the partnership. And we’re going to see where AI infrastructure goes from here. So with that, let me please welcome to the stage Jensen Wong. Woo! MUSIC What’s up, Jensen? How you doing? Boy, that’s a huge stage. That run a long way. Are you doing a good job? Yes. That’s a huge stage. That run a long way. Are you out of breath? Are you OK? I know. Let’s fire up. Good to see you. Dajahaw. There you go. Congrats on a great kickoff yesterday, GTC. You guys are off to the races this week. Thank you. Thank you. Look, maybe you heard some of what I just said. So we’re talking about connectivity today. The next trillion dollar company, ladies and gentlemen. Whoa! That would be exciting. Let’s do it together. But it really all starts with what’s happening today in AI infrastructure kind of more broadly. So how do you see that like just from the big picture standpoint, we’re at this extraordinary moment, customer demands through the roof. How do you see connectivity playing into this in the interconnect that’s required? Yeah, that’s really great. You know, yesterday I said that useful AI has arrived. It’s the reason why your demand is going through the roof. It’s the reason why my demand is going through the roof. Yeah. And this new computing pattern that makes it possible is called agents. And these agents has a particular computing platform, a computing pattern that is disaggregated and distributed. When you take a computing problem and you disaggregated into a lot of parts and you distributed across the entire data center, what’s necessary is connectivity. That’s the reason why Matt’s doing so well. That’s the reason why Marvell is so essential. We’ve distributed and disaggregated computing so that it runs across these enormous clusters so that we could get aggregating the total compute, the total memory, the total bandwidth that we have. And what makes it possible is connectivity. Yeah. I mean, we’re seeing it. And then as you think about it. That’s why they’re going to be the next trillion dollar company. We got a little work to do, but we’re on our way. We’re on our way. Thank you, Jensen. Well, let’s talk about scale. We used to talk about tens of GPUs and CPUs and XPUs connected, now thousands, now maybe millions at some point. So as you scale the compute and you scale the connectivity, I think we talked about things like agents. But how do you think about that across data centers, within data centers? How do you think about connectivity at large playing that role? And what kinds of technologies do you think are important there? Well, at the foundation of it, the agent computing pattern requires an orchestration system that allows the large language models, the computing, to be able to think and reason and come up with plans. But it also has to use tools and, you know, browse the internet, access memory, access

00:42:04
#

long-term memory, deal with short-term working memory. All of that requires a lot of connectivity. But it’s also the case. And if you look at the way we introduced Vera Rubin, Hopper was designed for training. Grace Blackwell introduced NVLink 72, our first scale-up fabric. It introduced the idea of extremely fast inference for MOE models that are very large, mixture of expert models that are extremely large. And so Grace Blackwell was for inference. Vera Rubin is to run agents, which is the reason why the Vera Rubin system includes, of course, the Vera Rubin thinking AI, but it also includes Vera CPUs for orchestration. It includes Vera CX for storage acceleration, for managing long-term memory. And the way that I think about these systems, you know, sometimes maybe the CSP wants to design their own custom chip. And between us, we also partnered together on NVLink Fusion, which makes it possible for you to use the same system architecture. And with Vera Rubin inside, some of your semi-custom chips, a lot of your interconnect silicon photonics and optics and technology such. And we can create, essentially, a disaggregated, distributed, and heterogeneous data center. And so that’s the big idea. And yet their system architecture is identical. Their networking technology can leverage a lot of NVIDIA stack. The CPU could be Vera. And yet it can leverage a lot of your stack. So NVLink Fusion is about taking NVIDIA’s technology and our platforms, Marvell’s technologies and plant, and we fuse it. That’s why it’s called Fusion. Yeah. No, I think, you know, I think about the partnership that we’ve been working together a long time. I think memorializing it with the investment, which we really appreciate. I think it’s been, it’s been huge for us. We’re honored to have it. You know, who doesn’t love making money? It’s nice to give. It’s done well since you invested. Boy. Yeah. When Jensen invests, just follow him. Just follow him. Give Matt all my money and just watch him make money. That’s what I’m doing every day. I love that. That’s what I’m doing every day. But I think these things you talked about, which we’ve brought to fruition, NVLink Fusion, working together on optics. I mean, I think the era of agents and kind of your new platform now, I think it’s ideally suited. I mean, NVLink Fusion, we had this idea years ago, right? But I think it was a little ahead of its time. And now, and I wanted to see if you agree when you think about kind of your platform and then some of the custom networking and compute needs that our customers have and the ability and the need to interoperate and work together, it seems like the time is now between Marvell and NVIDIA to really go enable our customers to have that flexibility that they’re looking for and really use the era of agents to scale our platforms together. Yeah. You know, ultimately, I do think that if you buy nothing but NVIDIA, it’s okay. Okay. I mean, but if you absolutely must design your own ASICs, we’re still happy having NVIDIA be inside that data center. And so, you know, you don’t have to buy everything from us. Just buy something from us. You know, we’re happy to support you and support the customer. And so, I think that between the two of us, you have the benefit of a general purpose, very high efficiency, you know, a system that is very well built starting with, you know, of course, Vera Rubin.

00:45:48
#

But anything that you want to extend to specialize, you can do so as well, which is the reason why your customers of mine, NVIDIA is in AWS, Marvell is in AWS, NVIDIA is in all of the clouds, and it’s wonderful to see Marvell expand into all of these different clouds. Yeah. Great. Thanks. One last one for you. Just leave some business for me, you know? Look, we’re your best salespeople right now. We’re your best salesperson. Working together. Final question for you. A lot of my talk is about some of the transition, especially as you go to inside the rack from copper to optical. It’s obviously not going to be a one-zero. It’s going to take, you know, there’s time and there’s different use cases. But how do you see that playing out right now? The transition from copper to optics and maybe how we can work together there too? Well, we should use copper as much as we can for as long as we can, but copper has its limits. Copper has its limits with bandwidth and also with distance. And so ultimately, the right strategy is to scale up with copper as long as you can. After that, you scale up further with optics and you scale out with optics and you scale across with optics. And so you use optics wherever you must. You use copper wherever you can. And so I think that that intersection is going to continue for a long time. The bottom line is in the next 5, 10 years, we’re going to use a ton of copper and we’re going to use tons and tons of optics. And so these data centers are part of infrastructure now. And the reason why I say that AI is now useful, useful AI has arrived is because now AI is profitable and tokens are profitable. When token production is profitable, everybody wants to make more tokens, which is the reason why, you know, Marvell’s demand is so high, is our demand is so high because everybody wants to produce more tokens because it’s used all over the place by agents. Absolutely. I think you touched on a bunch of things I’m going to cover later. If you want to do the rest of my presentation, you can. Yeah, so ladies and gentlemen, these beautiful slides, you know. Just sit right there. I’ll be. You take it from here. All right, Jensen Wong. All right, good to see you, brother. All right, take care. Okay, you guys, thank you. Thank you, Jensen. Bye Marvell. All right, outstanding, outstanding. Super fun to have Jensen here as always. All right, so we’ve been talking a lot about connectivity. Jensen and I just covered this, so let’s dive in now, right? Let’s go one level deeper. So AI infrastructure spans every distance. It spans from hundreds or even a thousand kilometers between data centers to just millimeters inside the package. Every one of those distances, it requires a different solution. It’s a different technology, different engineering team. It’s a completely different set of experts. And in many cases, it’s a different supply chain. So these are not variations of the same problem. What you have here is fundamentally different engineering challenges, and that’s what we’re going to walk through next. All right, so let’s start with the longest distance. Jensen referred to this. This is scale across, connecting data centers together. Now, every major cloud provider has hundreds of data centers around the world. And all of those data centers need to communicate with each other. This is fundamentally a long distance connectivity problem. We’re talking about links that can span hundreds or even a thousand kilometers.

00:49:22
#

This requires very specific, very complex technology called coherent modulation. And at the heart of it is a specialized digital signal processor, DSP. It’s designed to push enormous amounts of data, cross fiber optic cables over very long distances with extremely high reliability. There’s only a few companies in the world that build these coherent DSPs, and we’re one of them. Marvell has been a leader in this technology for many generations. We build optical modules that contain all the electronics needed to drive and modulate the laser and transmit data over long distances. So I’ve got a little show and tell here in my pocket. Not holding up a chip this time. I’m holding up an optical module. This is one of our coherent optical modules. This is an incredibly complex piece of engineering. At Marvell, we build the entire module. This is ours. It includes the advanced node CMOS DSP. It’s among the most complex chips, just the DSP alone that we design at Marvell. But it also incorporates inside our fourth generation silicon photonics technology. That’s inside here. We’ve been developing that technology in production for a decade on silicon photonics. It also includes our own broadband analog components that we designed, which is designed in silicon germanium. So Marvell pioneered this technology starting with 100 gigabits per second a decade ago, then moving to 400 gig, and now shipping 800 gig in volume. And later this year, we’ll be sampling the world’s first 1.6 terabit 2 nanometer coherent optical solution. And that couldn’t come at a better time. Demand for bandwidth has never been greater. All right, now let’s go inside the data center. So these data centers can be very large, spanning hundreds of meters, and they contain racks and racks of compute servers. Now, each rack typically has a switch at the top with servers connected into that switch. Those rack level switches connect to the spine and then the core switches. This creates the network fabric that ties the entire data center together. And all of that is connected through fiber optic cables. Now, once again, optical modules drive data transmission over those fiber optic cables. But this time the modulation scheme is different. Instead of coherent technology, we use a more power optimized modulation technology, which is called PAM4. So the two key semiconductor solutions for this part of the market are the PAM4 chip set inside the module, and then the cloud switching infrastructure that ties the data center together. Marvell builds both. Starting with the PAM4 chip set, we build the industry’s leading PAM4 DSP solution. And also the high speed analog components that go around them, including transimpedance amplifiers or TIAs and laser drivers. These are also in Silicon Germanium, by the way. And we’ve led the industry through every major transition of PAM technology, starting at 50 gig, 100 gig, 200, 400 and 800. Then last year, we began ramping Marvell’s 1.6 T 3 nanometer PAM4 solutions, leading the industry’s transition to 1.6 T connectivity. Now for ethernet switching, Marvell has a similarly complete portfolio of products from 12.8 terabits to 51.2 terabits. And today, we announced our new 100T ethernet switch, specifically designed for AI data centers with the industry’s lowest power. Special announcement for Computex. We waited. So you put it all together, we provide a complete solution for connectivity inside the data center. Now let’s move inside the rack.

00:53:23
#

The goal here is to connect the largest possible number of processors together in a full any to any configuration. In other words, every processor can communicate directly with every other processor. And Jensen talked about this. The first company to bring this architecture to market was NVIDIA with NVL 72, named for the 72 GPUs connected together inside a single rack. And this required a completely different approach to connectivity. It was a different class of switch and the ability to drive very high speed signals over copper back planes inside the rack. So today, this is not the domain of optics. This is the domain of copper. And the core differentiator here is the electrical SIRTES technology, not the optical. Now Marvell also has leading electrical SIRTES at 200 gigabits per second today. And we’ve demonstrated already over the last couple of years, 400 gigabits per second for the future. So we’re building this SIRTES technology into our customers, custom silicon and their XPUs and also into our own scale up switches. All right, now let’s go all the way inside the package. Here we’re not talking about meters anymore. We’re talking about millimeters. And you might not actually think about this as a connectivity challenge, but today, most advanced chips have multiple chiplets inside the package. So when you have two and a half D or 3D packaging, it’s fundamentally a connectivity technology actually. And it allows these chiplets to sit very close together inside a package and communicate through ultra high speed, short reach, die to die interfaces. And Marvell has leading die to die SIRTES and leading capability in advanced packaging, allowing our customers to build some of the most complex, unique multi die chips in the industry. So as you can see, connectivity for AI data centers requires a very broad portfolio of technologies. Each distance requires a very different solution. And Marvell has the industry’s most complete portfolio from millimeters to kilometers, every hop, every distance. And it turns out having all of those capabilities under one roof is unusual. It’s unique. When we go and compete, normally there’s a different set of companies that we compete against in each one of these categories across these different distances. But this is what makes us unique. We’re the one stop shop. We’re the leader across the entire connectivity stack. And that brings us to the next major challenge facing the industry. So what you probably notice as I describe these different solutions in the last couple of slides is there’s different solutions for different distances and that some of those connections today are optical and some of those connections today are electrical. And it’s actually defined by distance. And so the connections on the left side of this chart are optical debate. That means they use fiber optic cables to transmit light with complex electronics on either side of the cable to drive and modulate the laser that’s transmitting that light. Connections on the right side of this are electrical. So they use copper cables or just copper traces that are printed on the circuit board or even microscopic copper routing inside the package. So the common theme here is copper. And in the middle, you see the wall, the copper wall. And the wall is defined by the longest distance you can transmit a signal over copper. So before you have to move to an optical connection.

00:56:53
#

So this is an important distinction because copper is simple and it’s low cost. And as Jensen said, you want to use it for as long as you can. It’s very practical. But optics and optics is more complicated. It requires lasers, photonics, complex electronics. So it’s a bigger lift, but it’s going to be needed. And the copper wall, what I’m here to tell you today is it’s about to move. It’s going to move again and it’s going to take over the rack itself. So this is creating an explosion in demand for the optical industry. Incredibly complex engineering challenges are coming in along the way. So why is this happening? So it’s not just somebody’s preference to go do this. This is physics. The distance a signal can travel over a copper cable is inversely proportional to the bandwidth. So every time you double the bandwidth, you have to cut the distance in half. Today, the highest speed production systems in the world run at 200 gigabits per second per lane, just to give you an example. So at that bandwidth, the cable length is limited to roughly 2.5 meters. Now by comparison, systems running at 100 gain could use about 5 meter cables. And the height of the rack is about 2 meters. So once you account for all the routing inside the rack, 2.5 meters is right at the limit. So when we move to 400 gig, we can no longer fully connect the rack with copper. So the wall is moving and it’s moving now. Going forward, even the connections within the rack will become optical and the whole industry knows this is coming. So we’ve been preparing for this moment, not just Marvell, but the industry. And you see this in Taiwan, by the way, and the supply chain and the ramp up that’s happening. The ramifications for this are actually enormous because each time the wall moves one step to the right, the number of connections that you have goes up by at least an order of magnitude. So it’s creating this explosion in demand, as I mentioned, and the optical supply chain needs to scale up massively and be ready. We’ve seen this movie before, okay? I mean, 20 years ago, and I remember this one, State of the Art was 10 gigabits per second inside the data center. It was 10 gig. And we use copper cables all across the data center. Optics back then was reserved for just very, very long distances. It was essentially like a telecom technology. But when the wall moved, the optics industry actually rose to the challenge. And today, all the hyperscale data centers in the world, they’re all optically connected. And as we saw in that transition, it did require new solutions. You couldn’t use the same power hungry kind of telecom approach, which is where Pam 4 came in. It was optimized for power, density and reach and requirements specifically tuned to inside the data center. And Marvell was one of the key innovators there. So we’re about to see the same wave of innovation needed as optics moves inside the rack. And that’s what the technology called co-package optics or CPO. You hear a lot about this now. I’m going to tell you more. CPO is a technology where we bring the optical connections all the way to the package itself. Right next to the compute, either the custom compute or the switching silicon. And the fundamental challenge we’re solving with CPO is density and power. Now, remember, the number of connections inside the rack is like 10x the number of connections between the racks.

01:00:11
#

So if you just try to use the same optical technology used across the racks in the data center, you wouldn’t have enough power, you wouldn’t have enough physical space. You cannot fit all these standard optical modules and cables as they are today. It just doesn’t work. It’s not possible. So the industry has been inventing this co-package optics concept, which brings the optical fiber right to the package. It tightly couples the electronics that drive the signal over the fiber directly with the custom compute or switching silicon. So this is a massive change. And it’s hard because you’re combining some of the most advanced technologies in the chip industry. Leading edge CMOS, silicon photonics, advanced packaging, optical interconnect, all manufactured in a small tightly integrated system. So the complexity is very high, but it’s the only way to continue scaling bandwidth and overcome this limitation that I talked about with copper while reducing power at the same time. So this is where the industry is headed. And this is one of the reasons that Marvell has invested for more than a decade in silicon photonics, optical DSPs, all the analog broadband components around it and all the advanced packaging you need to pull this off. It needs to all come together actually in CPO. So this isn’t some futuristic thing, guys. It’s happening now. And in fact, I brought a couple of Marvell examples with me today. So let’s do a quick show and tell. OK. OK, so over here you have a traditional Ethernet switch. This is our 100T TeraLink switch that we announced today. And you guys are the first to see it actually. Everybody here in the room. You can see the switch in the middle of the board. Copper traces inside the PCB carry the signal to the front panel, which is here. And this is where all the optical modules plug in. Now let’s move over here. This is a CPO based switch right here. Now notice that there’s still the switch silicon in the middle. That’s right in the center of the die of the package. In this case, this is our 51.2T switch. And all around the edges are 16 3.2T optical engines. So the 16 times 3.2, you get 51.2. So the fiber is directly attached now to these engines. It’s not to the front panel. So we’ve completely eliminated the copper traces on the PCB. Light comes directly out of the package. This is a very, very complex piece of engineering. And it was very cool to be able to show this off today. OK, so co-package optics is here, and the industry is scaling up to meet the challenge. And as we’ve seen time and time again, every time we reach a physical barrier, we break through it with technology and innovation. In this case, by replacing copper with fiber, because unlike electrons traveling over copper wires, the distance that photons can carry, a signal through glass, is largely unrelated to the bandwidth. So as AI infrastructure demands even higher transmission speeds and needs to scale to larger and more complex systems, spanning millions of processors woven together now, not thousands or hundreds, optical conductivity will increasingly become the de facto solution. So the real question becomes, what does it take to deliver optics across the full AI infrastructure stack? What’s it going to take? Well, it starts with recognizing there is no single technology for the entire data center. It’s not how this works. There’s no one-size-fits-all solution. There’s no shortcuts. There’s no easy way to the end here.

01:03:42
#

There’s not a single architecture modulation scheme, frequency band, unique technology that’s going to do it all. There’s no free lunch. That’s why we are pursuing a bunch of different unique optical paths across every distance to get here. Each one of these technologies that I have up here is optimized for a different design point. Each one enables a critical part of the infrastructure and addressing different requirements for density, bandwidth, and power and integration all across the stack. So if optical interconnect is the underlying technology for which next generation AI infrastructure is built, then Marvell is building the broadest portfolio with the deepest bench in the industry. But no company can deliver this transformation alone. And as Jensen talked about earlier, it takes an ecosystem to get here. So like I said, technology innovation is great. It’s part of the challenge, but not all of it. But demonstrating this at scale is really what matters. And at this point, if you’re just operating on a PowerPoint or a demo, POC, press release, it’s not going to get you there. Customers need solutions now that are ready. They’re reliable. They need to be manufacturable and be ready to deploy at scale. So Marvell and our ecosystem partners have been doing this for a long time. We’ve already shipped hundreds of millions of DSPs. We’ve accumulated through our volumes tens of billions of device hours of data in the field. This experience matters because these products have to work not just in the lab, but in the world’s largest data centers at very high volume and very reliably for years. So that requires investing ahead in the manufacturing ecosystem. You’ve got to build the capacity in the supply chain infrastructure before the market arrives. This is why the ecosystem matters so much. And it matters a lot here in Taiwan, by the way. Now one of our most important partners at Marvell in this journey has been advanced semiconductor engineering or ASC. Now ASC is one of the world’s leading semiconductor manufacturing companies. They’re more than 100,000 employees with operation in Asia and actually all around the globe with the decades-long track record of helping enable pretty much every major technology transition we’ve gone through in the semiconductor industry. Now leading ASC through this period of transformation is someone that I know quite well. He spent more than 25 years helping shape both the company and the industry. Today I’m thrilled to have my next guest speaker come up, which is ASC CEO Dr. Tian Wu. Tian, please join me on the stage. Thank you. Hi. Tian, how are you? Thank you for inviting me to Computec. Great to see you. It’s an honor to have you on stage with us. Well, it’s my honor. Look, we’ve been working together a long time. And when I became the CEO, we had a set of ambitions. We talked to a lot of our suppliers. I’ve known you even before I was the most successful. I’ve known you even before I was the Marvell CEO. When I was executive back at Maxim and we worked together there. But part of what it may be explained to the audience too that sometimes people don’t realize is that as a key supplier into this ecosystem, you have to make bets. You’ve got to make bets on the companies you work with. You’ve got to make bets on who you think is going to be successful. And we really appreciate that ASC bet on Marvell very early, very early. And we’ve seen great success actually based on that.

01:07:24
#

But I’m just curious if you could share your perspective maybe on where Marvell was, what your thought process is, and then where are we today in our journey together. So it would be great to hear from you, Tian. Okay, I think the best way to describe this is a gradual process. The first decision was not difficult. Marvell’s fabulous company has a very good reputation, has gone through a lot of transition. So the track record of Marvell has already been there. The product set was a little bit obsolete at the time when you joined. So the first one is the business model needs to be aligned. Taiwan ASC is in the manufacturing sector. So we’re looking for bet, not only on betting on your success, we’re also betting on somebody who can provide the insight for the next generation architecture and also the technology requirement. As you know, the Taiwan company invests infrastructure and cat packs 10 years ahead of time, big bet. We’re only counting on whatever capacity we put it in, will be needed and will be utilized. That’s how we make money. So betting on company that we believe will give us very good insight well into the future becomes very important. So that’s how the decision was made at the very beginning. And for the last 10 years, I’m just really happy. Everything that we talk about, it was a dream 10 years ago. It was a dream. And today we are going to ship it. And you just mentioned that you’re going to have 40% growth for the next few years. I believe you’re going to be that. So we’re busy now preparing the capacity for you. We also appreciate that over the last 10 years, we have gone through a lot of strategic discussion. You make commitment to us, we make investment for you, and over time we’re going to produce more of your parts. I think that’s really a short story for how that decision will come. Yeah, no, it’s been a great story. Maybe one more for you. The ecosystem here in Taiwan is so unique. And like you said, it takes like a decade of investment before you really can see the return. And there’s just such a power that’s happening here. How do you describe it to the people here? And also there’s a lot of people around the world watching. And then what makes it possible here? Why is it unique? And then what also makes it difficult to replicate this in the rest of the world? But at the same time, there’s globalization. So how do we think about those dynamics? I think that’d be an interesting one. I think the reason why you’re asking the question is there’s a lot of competing forces and also uncertainty across the world. So I think my belief is any business needs to have vision as well as long-term alignment on value. So in the business model, the whole Taiwan sector is built on capacity utilization and also innovation and technology investment way ahead of the curve. That’s what Taiwan’s value. So with the fabulous company or with specific IDM company that business model aligns. Beneath that will be the economy of scale. Taiwan accumulated 40 years based on the PC transition to the wireless, to the mobile computing, to the data center, now we’re into HPC. So that 40 years of experience accumulated 350,000 semiconductor employees. Also accumulated 1.1 million high-tech employees and many of them are here. That experience becomes extremely valuable combined with the economy of scale as well as the cluster efficiencies. So we think about the workforce with years of experience behind it. We think about the cluster efficiency.

01:11:25
#

When you think about the capacity, economy of scale we already put it in. But one more thing, I think Taiwan good or bad, we had fewer choices than the other region like the United States. So most of the engineer when they come out, they have few choices to make. Semiconductor IT industry becomes an attractive choices in Taiwan, not necessarily in the other region. So with all of this combined, I think this ecosystem is very, very difficult to replicate. It is not impossible, but will take years. Great. Well, thank you so much. I appreciate the partnership so much. We’re off to the races. Tian, thank you. Tian Wu. Thank you. Okay. So, like we said, the future of AI data centers is all optically connected infrastructure. And you heard him say it, right? This is going to drive a tidal wave of growth, innovation that’s needed in scale and manufacturing. But what does that inevitable future actually look like? I mean, if you just take a step back for a minute and you actually don’t think about right now, think about 10 years in the future. And it’s a world where a lot of the copper connections are gone. And just think about a world where data transmission now at some point is all optical. This is a world where then distance doesn’t matter, actually. And that’s a profound change. Servers, racks and overall data center architectures today have all been designed around the constraints of distance. And software workloads actually have also been optimized around those same constraints. But what if distance no longer matters? How might the architecture itself change? And what new capabilities become possible when the infrastructure is no longer constrained by distance? So let’s start with the scale up network in the rack. As we discussed earlier, this is where we can connect the largest possible number of processors together in a full any to any configuration. Now in the past, the size of this domain was limited by the length of the copper connection, but with optics distance doesn’t matter. So now we can change the size of the scale up domain from 72 or 144 xpus or GPUs to 1000 or more, all optically interconnected. The implications for workloads are enormous. Today, AI workloads must be broken down into smaller sub problems that fit within the scale up cluster. Because communicating outside the cluster today is slower, much lower bandwidth. But optically interconnected systems can manage workloads on an order of magnitude larger. And it does not stop there, by the way. What happens when the optical connectivity comes inside the server itself? Modern AI servers are composed of a certain number of CPUs, xpus, memory and network interfaces. And the reason they’re all in the same system is because of distance. CPUs and xpus need to access memory at very, very high bandwidth, which means they need to sit right next to each other on the board with copper traces serving as the connections between them. But in a future where these connections are all optical, distance actually doesn’t matter. You can imagine a completely disaggregated architecture. Xpus in one system, memory in another, agentic CPUs in another, which unlocks another possibility. In today’s systems, the ratio of CPU and xpus or gpu, it’s fixed. So these ratios have to be defined at the time the system is built and deployed. But no two workloads require exactly the same ratio. Jensen talked about this, actually.

01:15:06
#

Which means at any given time, some portion of the computer memory could be underutilized for a given workload. That costs money. But once we decompose the system into separate pools of compute and memory, and they’re all optically interconnected, we can then decompose dedicated systems on the fly, which are then optimized for whatever the workload is. So imagine future data centers. A globally optically interconnected data infrastructure. These rigid boundaries we have today in the systems we have, they begin to disappear. Compute can now be pulled. Memory can be pulled. An infrastructure can be composed dynamically at scale. For the first time, architects can begin designing AI systems around the needs of the model, not around the limits of the interconnect. So this is where AI infrastructure is headed. It’s a data center without distance. Where compute, memory, networking and photonics operate as one unified system. Where millions of resources across the data center can work together as if they were one machine. An architecture defined by the needs of the workload, not the limits of the connectivity. We believe this is the next era of computing infrastructure. And Marvell is helping build the connectivity foundation that will make all this possible. Thank you very much for your time today. Thank you. Thank you.

这里还没有任何文章可以列出。