Arm CEO Says Edge AI Demands Threaten Legacy Chip Formats

Arm首席执行官称边缘AI需求对传统芯片架构构成威胁 本地化AI推理呈指数级增长的处理需求将阻止计算停滞在“够用就好”的门槛上，从而巩固市场从传统x86架构永久转向Arm可定制、低功耗生态系统的趋势。 Arm首席执行官Rene Haas探讨了公司从低功耗RISC根基发展至如今1500亿美元估值及架构无处不在的发展历程。他阐述了Arm灵活的授权模式如何战胜传统上僵化的x86生态系统，使超大规模企业能够通过高度定制的芯片实现总拥有成本的大幅优化。Haas还审视了本地化AI边缘计算不断扩大的前景，详述了Arm向预先验证计算子系统迈进的战略转变，并将Nvidia斥资400亿美元收购未果的事件视为对Arm在未来技术中结构性重要地位的终极印证。 在经历了从剑桥起步到2016年被SoftBank私有化的复杂企业历程后，Arm Holdings已重返公开市场，估值约为1500亿美元。 (0:00) Arm采用双重商业模式，既可将其指令集架构（ISA）作为蓝图进行授权，也可授权完整现成的CPU核心设计。 (1:19) Arm在市场上拥有极高的普及度，2024财年芯片出货量接近290亿颗，平均相当于地球上每个人约拥有四颗芯片。 (5:20) Arm的基础设计约束始于Apple Newton，迫使其专注于极致的低功耗与散热优化，以适应廉价的塑料封装外壳。 (6:46) 精简指令集计算机（RISC）架构通过使用精简、简单的指令将内存占用降至最低，相比复杂指令集计算机（CISC）在功耗和经济效益上实现了重大突破。 (12:28)

Arm CEO Rene Haas outlines how the company's energy-efficient RISC architecture and neutral licensing model established it as the dominant compute platform, expanding from low-power mobile devices into cloud infrastructure and artificial intelligence. Foundational Efficiency and the Shared-Success Business Model (6:46) Arm originated over 30 years ago at Cambridge University, achieving its initial commercial design win with the battery-constrained Apple Newton. To cut packaging expenses, the original Arm 1 processor was optimized for low heat generation so it could use inexpensive plastic rather than costly ceramic. Unlike CISC designs such as Intel x86—which ran complex, power-hungry instructions across multiple clock cycles to preserve legacy code compatibility—Arm implemented a streamlined RISC architecture executing basic operations with reduced power and memory footprints. Founders Robin Saxby and Tudor Brown created a shared-success business model to lower barriers to entry. Rather than funding costly fabrication plants like Intel or Motorola, Arm licensed its IP via upfront fees and ongoing per-unit royalties. To maintain universal software portability, Arm strictly prohibited architectural licensees from altering the instruction set architecture (ISA) with proprietary extensions. Early baseband adoption by Texas Instruments for Nokia 2G handsets, followed by Apple's decision to base iOS on an Arm-targeted Mac OS kernel rather than Intel's Atom, turned the architecture into the global standard for mobile devices. Software Stickiness and Industry Consolidation (5:20) Over several decades, alternative microprocessor architectures—including SPARC, MIPS, Motorola 68000, and DEC Alpha—vanished from mainstream compute. Haas notes that a programmable processor relies entirely on the persistence of its software ecosystem, as rewriting code carries prohibitive costs. As venture capital shifted away from semiconductor design toward high-margin software during the dot-com era, new general-purpose CPU development stalled, leaving x86 and Arm as the two primary survivors. This dynamic created an operational moat comparable to the extreme capital expenditures required for leading-edge semiconductor foundries. With more than 20 million developers actively targeting Arm platforms, Haas explains that new entrants face nearly insurmountable software switching barriers. In fiscal year 2024, nearly 29 billion Arm-based chips shipped across consumer electronics, automotive platforms, and enterprise infrastructure, equating to roughly four processors per person globally. Expansion into Data Centers and Custom Silicon Arm has moved beyond low-power edge devices into modern cloud data centers, competing directly with the AMD and Intel x86 duopoly. Haas attributes this shift to the total cost of ownership (TCO) flexibility offered by licensing. Hyperscalers such as AWS, Microsoft, and Google license Arm's architecture to build customized systems-on-chip (SOCs) integrated with tailored memory and interconnects, yielding publicly reported performance improvements of up to 60% on a like-for-like basis. Haas rejects the argument that parallel GPU computing renders CPUs obsolete, emphasizing that system host processors remain essential for operating systems, memory management, and workload coordination. Systems like Nvidia's Grace Blackwell demonstrate this integration, pairing Arm CPUs with GPUs over NVLink to optimize bandwidth. Local AI inference on power-constrained edge hardware similarly depends on CPUs paired with dedicated neural processing units (NPUs), such as Arm's Ethos line delivering 4 to 8 TOPS. Compute Subsystems and Vertical Integration Tensions (63:38) Arm is expanding its portfolio beyond individual IP blocks into verified compute subsystems. Instead of delivering disparate CPUs, GPUs, and Coherent Mesh Networks for licensees to assemble, Arm pre-integrates and validates complete layouts with foundries like TSMC, Samsung, and Intel. Haas explains that these pre-tested virtual chipsets can trim three to nine months from customer development schedules and guarantee specific performance thresholds, such as 4.4 GHz frequency targets. The strategic value of Arm's horizontal positioning was highlighted in 2020 when Nvidia proposed a $40 billion acquisition. The transaction faced broad regulatory and customer resistance out of concern that a single player could restrict neutral industry access. Since returning to the public markets, Arm's quarterly revenues have reached $939 million (around a $4 billion annual run rate) against a market capitalization near $150 billion, supported by structural tailwinds as the semiconductor industry targets a $1 trillion market by 2030. Arm's growth rests on remaining the universal, neutral architecture for computing across edge and data center markets, though its ongoing shift into fully verified compute subsystems leaves open the question of whether it might ultimately move closer to direct chip production.

Arm 首席执行官 Rene Haas 指出，Arm 凭借低功耗 RISC 架构设计、中立开放的 IP 授权模式以及数千万人构建的软件生态，已从早期的移动设备核心跃升为驱动边缘计算和现代数据中心的关键计算标准。 起源与能效核心：从 Newton 到全球普及 Arm 始于 30 多年前的剑桥大学，最初为苹果公司的 Apple Newton 掌上电脑设计 Arm 1 芯片。受限于当时高昂的散热陶瓷封装成本，团队将低功耗和低温运行作为硬性指标，以适配廉价的塑料封装。随着计算环境从传统的插电台式机延伸至移动硬件，精简指令集（RISC）相比包含庞大历史包袱且每周期操作复杂的复杂指令集（CISC，如 Intel x86）展现出极高的能耗比与内存占用优势。 在 2024 财年，基于 Arm 架构的芯片出货量接近 290 亿颗，折合全球人口平均每人约 4 颗，远超年出货量仅约 2 亿台的全球笔记本电脑市场。Haas 指出，无论是奥迪汽车的数字仪表板、Nest 门铃、LG 冰箱、三星电视，还是索尼 PS5、Pixel 手机和 iPad，现代家庭中通常运行着数百个 Arm 处理实例。Arm 将 CPU 定义为设备的数字大脑，并通过提供基础指令集架构（ISA）授权或现成 CPU 核心设计，支持苹果、三星、高通和亚马逊等客户。 开放授权模式与不可动摇的生态粘性 (24:27) 不同于英特尔集设计与物理制造于一体的高利润模式，Arm 在 1990 年代开创了“共享成功”的轻资产商业范式：收取前期研发许可费并在客户量产后按比例提取版税。在 1990 年代中期，德州仪器（TI）为诺基亚生产 2G 及 GSM 手机基带芯片时率先引入 Arm 核心；随后苹果在初代 iPhone 的开发中放弃了功耗过高的英特尔 Atom 处理器，选择围绕 Arm 架构改造 Mac OS 内核以构建 iOS。伴随 Android 阵营的标准同步，Arm 一举锁定移动端统治地位。 更为关键的是，Arm 严格禁止架构授权客户私自添加自定义指令集，以确保底层软件在全产业链的完全兼容。过去数十年间，虽然半导体行业埋葬了 SPARC、MIPS、DEC Alpha 及摩托罗拉 68000 等众多优秀 CPU 架构，且风险投资长期转向高毛利的 SaaS 领域，但 Arm 凭借超过 2000 万名活跃开发者的软件生态形成了类似先进制程晶圆厂（如 TSMC 投资 300 亿至 350 亿美元）的坚固壁垒，使跨架构迁移面临极高的软件重写摩擦。 攻入数据中心：定制算力与软硬件协同 在云计算与数据中心领域，Arm 的开放授权打破了以往受制于 Intel 与 AMD 传统 x86 现成芯片的局限。微软、谷歌及亚马逊（AWS）等超大规模云服务商（Hyperscalers）纷纷基于 Arm 打造定制片上系统（SoC），并整合自有存储、高速互连及机架配置。Haas 透露，客户报告在同等工作负载下实现了最高达 60% 的性能提升，大幅降低了整体拥有成本（TCO）。 以英伟达的 Grace Blackwell 架构为例，基于 Arm 的 Grace CPU 能够突破 x86 的接口限制，通过高速 NVLink 与 GPU 紧密耦合并优化高带宽内存（HBM）的协同表现。Haas 驳斥了 GPU 加速计算将取代 CPU 的观点，他将加速器比作发动机，而 CPU 是不可或缺的方向盘和底盘系统，主控处理器仍然承担着运行完整操作系统与基石软件栈的职责。 AI 时代展望：边缘推断与计算子系统整合 (63:38) 针对半导体行业到 2030 年有望达到 1 万亿美元规模的行业前景，Haas 认为虽然短期市场对 Copilot 等软件存在商业化回报质疑，但人工智能本质上是类似于智能手机和工业革命的底层转型。与过去上网本时代因算力过剩出现的“足够好就好”停滞现象不同，前沿 AI 模型在边缘侧的推断对能效提出了无休止的要求。端侧设备无法运行 100 瓦的大功耗 GPU，促使 Arm 通过架构扩展及其内置的 Ethos NPU（提供 4 至 8 TOPS 算力）来承载本地加速与隐私安全验证。 为应对日益复杂的系统集成挑战，Arm 正在推进“计算子系统”（Compute Subsystems）业务，将 128 核 CPU、GPU、NPU 及相干网状网络（CMN）进行预验证与预集成，并联合台积电、三星等代工厂实现特定频率（如 4.4 GHz）的性能保障。这种做法将原先零散的 IP 模块转变为虚拟芯片组，可帮助客户缩短 3 到 9 个月的开发周期，使其能专注于相机 ISP 或专有 AI 单元的差异化创新。 针对市场曾讨论 Arm 是否会跨过客户直接交由台积电自主生产实体芯片的猜想，Haas 明确表示当前无可奉告；同时，英伟达此前发起的 400 亿美元收购案虽因反垄断审查和生态抵制而终止，但该事件反而印证了 Arm 作为全球中立计算底座不可替代的战略价值。

## Background 该内容由知名科技商业播客频道 Acquired 于 2025 年 4 月 3 日发布在 YouTube 平台（音频原版于 2024 年 12 月上线）。节目简介表明，本期深度访谈了 Arm Holdings 首席执行官 Rene Haas，讲述 Arm 如何从 1990 年在英国剑桥为 Apple Newton 等便携设备研发低功耗芯片起步，跨越 iPod、智能手机与 AI 浪潮，并经历软银收购、二次上市以及险被英伟达收购等重大历史转折，最终蜕变为当今全球计算世界默认芯片架构的非凡历程。 ## TL;DR 访谈全景式拆解了 Arm 依托精简指令集（RISC）的能效基因，从早期诺基亚手机基带与初代 iPod 切入，再借由 iPhone 和 Android 浪潮彻底统治移动计算的演进之路。Rene Haas 阐述了 Arm 依靠前端授权费与出货版税构筑的“共享成功”商业模式，以及凭借坚守指令集兼容性铸就的全球级软件生态护城河。在当前 AI 与算力大爆发的背景下，Arm 凭借超大规模云厂商自研芯片和英伟达 Grace Blackwell 等系统级协同方案，在数据中心领域对封闭垄断的 x86 发起强力颠覆。此外，节目还探讨了 Arm 填补边缘 AI 推理空白的战略，以及从单体 IP 转向预先验证的“计算子系统”（CSS）这一逼近虚拟芯片形态的重大商业跃迁。 ## Atmosphere & style 整期节目散发着半导体工业编年史的厚重感与顶级商业洞察的张力，既有对公用电话亭和“手持设备不融化塑料壳”拓荒岁月的怀旧回眸，也有产业巨擘面对万亿美元 AI 范式转移的从容气度。Acquired 主持人抽丝剥茧的深究式发问，与 Rene Haas 形象生动的“乐高积木”、“底汤之水”及“汽车轮胎”等比喻碰撞出密集的智识火花，谈话氛围严谨扎实却又举重若轻。随着话题层层推进到世纪反垄断博弈以及最终 Arm 是否会亲自涉足实体芯片制造的敏感设问，全场在充满玄机的微笑与巨大悬念中收尾，极具戏剧张力。 ## Bulletin - [05:20](https://www.youtube.com/watch?v=Y8guMe665RE&t=320s) Arm 芯片在 2024 财年出货量接近 290 亿颗，相当于全球平均每人对应近 4 颗新芯片，其单季营收已达 9.39 亿美元，市值约达 1500 亿美元。 - [50:14](https://www.youtube.com/watch?v=Y8guMe665RE&t=3014s) 主流云厂商（AWS、微软、谷歌）自研的 Arm 服务器芯片在同等对比下实现了约 60% 的性能与能效提升，打破了 x86 在数据中心的传统垄断。 - [01:10:14](https://www.youtube.com/watch?v=Y8guMe665RE&t=4214s) Arm 正在推进“计算子系统”（CSS）商业模式升级，将多达 128 核的 CPU、互连网状网络和内存控制器预先集成并完成软硬件验证，可为客户节省 3 到 9 个月的研发周期。 - [01:12:49](https://www.youtube.com/watch?v=Y8guMe665RE&t=4369s) 面对未来是否会跳过授权直接委托台积电代工自有实体芯片的问题，Arm CEO Rene Haas 报以微笑并表示“今天没有任何可以透露的”，为市场留下巨大想象空间。 ## Key takeaways - [05:20](https://www.youtube.com/watch?v=Y8guMe665RE&t=320s) Arm 诞生于剑桥为 Apple Newton 掌上电脑设计低功耗芯片的极端限制，为了能塞进廉价塑料封装散热而将能效刻入底层基因。 - [10:28](https://www.youtube.com/watch?v=Y8guMe665RE&t=628s) RISC 与 CISC 之争的本质是硬件极简与历史包袱的博弈，但 CPU 架构的真正价值并非由理论优劣决定，而是取决于其上编写的软件生态及其寿命。 - [15:23](https://www.youtube.com/watch?v=Y8guMe665RE&t=923s) IBM PC 开放采购 8086 与 DOS 促成了兼容机洪流，绑定海量软件生态，使得理论上更优秀的 RISC 在 PC 时代被沉重的软件惯性锁死在边缘。 - [19:16](https://www.youtube.com/watch?v=Y8guMe665RE&t=1156s) 黄仁勋提出“软件永不死”，架构迁移无法依靠仅 15% 的增量改善，必须依赖颠覆性的功耗红线跨越或开辟全新产品形态。 - [24:27](https://www.youtube.com/watch?v=Y8guMe665RE&t=1467s) Arm 开创了“前期适度授权费 + 后续出货版税分成”的共享成功商业模式，打破了半导体行业以往各家从零自研微处理器的垂直整合旧范式。 - [27:36](https://www.youtube.com/watch?v=Y8guMe665RE&t=1656s) 德州仪器在诺基亚 2G/GSM 手机基带芯片中选用 Arm，成为其历史上首个关键 Design Win，推动 Arm 确立了移动通讯的事实标准并孕育塞班软件生态。 - [33:34](https://www.youtube.com/watch?v=Y8guMe665RE&t=2014s) 在初代 iPhone 选型中，英特尔受限于“创新者的窘境”试图向下阉割 Atom，而苹果 iPod 派系凭借自底向上的极限能效追求胜出，决定了整个移动计算浪潮的技术走向。 - [40:35](https://www.youtube.com/watch?v=Y8guMe665RE&t=2435s) Arm 架构授权有不可逾越的底线：严禁客户擅自增添私有独占指令破坏 ISA 兼容性，确保全球开发者无需面向硬件碎片化寻找“最低公共分母”。 - [44:11](https://www.youtube.com/watch?v=Y8guMe665RE&t=2651s) 随着风投资金长期从半导体芯片涌向 SaaS 与互联网软件，CPU 领域迎来历史大出清，绝大多数历史架构凋零，最终仅剩 x86 与 Arm 两大支柱。 - [47:48](https://www.youtube.com/watch?v=Y8guMe665RE&t=2868s) 先进制程晶圆厂依赖 300 亿至 350 亿美元的 CapEx 资金门槛，而处理器架构则依靠全球 2000 多万名常年开发者的 OpEx 沉淀构筑起软件重力场。 - [54:08](https://www.youtube.com/watch?v=Y8guMe665RE&t=3248s) AI 的爆发彻底终结了“上网本”时代算力“足够好”的停滞梦魇，不论是前沿大模型云端训练还是数十亿边缘设备的推理，都对算力密度与能效提出了永无止境的需求。 - [59:23](https://www.youtube.com/watch?v=Y8guMe665RE&t=3563s) “CPU 已死”是荒谬的论调，正如给车换上 V8 发动机依然需要方向盘，作为系统核心调度的 CPU 在海量边缘推理与端云安全协同中扮演着 GPU 无法替代的角色。 - [01:03:14](https://www.youtube.com/watch?v=Y8guMe665RE&t=3794s) 英伟达 2020 年拟以 400 亿美元收购 Arm 遭全行业强烈抵制，监管机构罕见地基于未来潜在竞争风险叫停，反向证明了 Arm 作为中立科技通用基建的不可替代性。 - [01:09:08](https://www.youtube.com/watch?v=Y8guMe665RE&t=4148s) 英伟达在 Grace Blackwell 架构中通过 NVLink 将 Arm CPU 与 GPU 极速耦合打通显存，释放了传统封闭 x86 系统无法企及的架构定制自由度。 - [01:10:14](https://www.youtube.com/watch?v=Y8guMe665RE&t=4214s) 通过推出“计算子系统”（CSS）完成高度整合与代工厂硬核标定，Arm 正在从单体 IP 供应商演进为提供虚拟芯片组方案的系统级领导者。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=Y8guMe665RE&t=0s) 开场与 Arm 的商业模式：从剑桥走出的“数字大脑” 节目开场介绍了本期重磅嘉宾——Arm Holdings 首席执行官 Rene Haas。作为当今计算世界的隐形基石，Arm 研发的指令集架构与芯片设计方案驱动着从智能手机到汽车的无数设备。Acquired 曾在 2019 年梳理过这家传奇公司的起源：它诞生于英国剑桥大学，经历过公开上市，2016 年被软银（SoftBank）以私有化方式收购，随后在 2023 年重返公开市场，市值如今已达约 1,500 亿美元。Rene 本人在半导体行业履历深厚，在加入 Arm 过去的 11 年前，他曾担任 NVIDIA 副总裁并直接向黄仁勋（Jensen Huang）汇报。 对于资本市场和科技从业者而言，Arm 近年来战略转型的每一步都备受瞩目，但对于普通公众来说，很多人甚至不知道 Arm 究竟是什么、它做出来的东西究竟藏在哪里。面对“Arm 到底在做什么”的基础问题，Rene 给出了一个极为精炼的定义：Arm 做的是 CPU，而 CPU 就是“每一台现代电子设备的数字大脑（digital brain）”，无论是电视、恒温器还是汽车，其智能运转的核心都离不开它。 针对普通消费者的常见疑惑——“我手机里的芯片是苹果 A18，电脑里是 M4，这怎么能说是 Arm 做的呢？”Rene 拆解了 Arm 赋能行业的两种主要商业模式： ```mermaid flowchart TD accTitle: Arm 的核心商业交付模式 accDescr: 展示 Arm 向下游不同客户交付芯片技术与架构的两种主要路径 subgraph ARM["Arm Holdings"] ISA["指令集架构 (ISA) 授权"] CPU_IP["自研 CPU 核心设计 (IP 授权)"] end subgraph PARTNERS["行业合作伙伴与终端产品"] direction TB ARCH_DEV["自研微架构厂商<br/>例如：Apple（A系列 / M系列芯片）"] CORE_DEV["芯片集成商与系统厂商<br/>例如：Samsung、MediaTek、Tesla、Qualcomm、Amazon"] end ISA -->|"授权指令集规则"| ARCH_DEV CPU_IP -->|"直接交付设计好的核心方案"| CORE_DEV ``` - **架构授权（ISA Licensing）**：Arm 将指令集架构授权给合作伙伴，由客户自行研发基于 Arm 指令集的专有微架构与芯片。苹果（Apple）便是这一模式的典型代表，其 iPhone 与 Mac 内部的芯片均基于 Arm 架构自主研发。 - **处理器核心授权（CPU IP Licensing）**：Arm 亲自完成 CPU 核心的电路设计与验证，将现成的核心 IP 授权给芯片公司与系统厂商（例如三星、联发科、特斯拉、高通、亚马逊等），客户可直接将其集成到自己的系统级芯片（SoC）中。 --- ### [02:33](https://www.youtube.com/watch?v=Y8guMe665RE&t=153s) “推开家门全是 Arm”：现代生活中的隐形神经中枢 若要直观感受 Arm 的普及程度，不妨清点一下普通人家中的各类设备。究竟有多少设备内置了 Arm 芯片？或者说，一个家庭环境中到底漂浮着多少颗 Arm 处理器？Rene 直言，这很难给出一个精准的绝对数值，因为如果从“数字大脑”的第一性原理出发——只要设备需要运行复杂的操作系统、控制面板或应用程序，内部往往就跳动着一颗 Arm 核心。 Rene 列举了自己下班回到家时的典型场景，几乎每一个触碰的角落都被 Arm 覆盖： - **车库中的奥迪（Audi）**：驱动数字液晶仪表盘与中控大屏的处理器是 Arm；参与高级辅助驾驶（ADAS）计算的芯片是 Arm；甚至连控制车门电动锁、电动车窗的微控制器中也是 Arm。 - **智能家居入口**：门口安装的 Google Nest 智能可视门铃，负责摄像头视频流处理、门铃交互的核心芯片由 Arm 提供。 - **厨房家电**：走过厨房里的 LG 冰箱和 Wolf 炉灶，无论是门上镶嵌的高清显示屏、精准的温控传感器调节，还是烤箱内部的烘焙曲线控制逻辑，全部由 Arm 芯片在后台运算支撑。 - **客厅娱乐**：打开三星（Samsung）智能电视，其流畅运行的 Android TV 系统底层正是 Arm 架构。 - **游戏机与移动终端**：如果走到地下室玩索尼 PS5，从游戏手柄的交互到主机显示控制芯片，都有 Arm 的身影；而身旁随手翻阅的 Google Pixel 手机与 iPad，更是 100% 的 Arm 计算平台。 现实中如果想避开 Arm，几乎只能退回到没有任何数字化模块、全凭机械杠杆运作的古董机械设备。在任何一个稍具现代互联属性的家庭中，“数以百计的 Arm 实例”并非夸张的修辞，而是早已发生的技术现实。 --- ### [05:20](https://www.youtube.com/watch?v=Y8guMe665RE&t=320s) 年出货 290 亿颗：从 Apple Newton 的低功耗基因到云端 在出货量量级上，Arm 展现出了惊人的规模效应。根据财报披露的数据，仅在 2024 财年，基于 Arm 架构的芯片出货量就达到了近 290 亿颗。全球目前约 80 亿人口，这意味着在短短 12 个月内，地球上平均每个人都对应着近 4 颗新出厂的 Arm 芯片。 为了衬托这 290 亿颗的夸张量级，Rene 对比了人人都熟知且兵家必争的笔记本电脑市场：全球笔记本电脑一年的总出货量大约在 2 亿台左右，在 290 亿面前仅仅是一个微不足道的零头。从机场航显大屏、民航客机机载系统，再到最底层的工业传感器，Arm 真正做到了无所不在。更为颠覆认知的是，如今 Arm 甚至反向渗透进了支撑所有这些终端设备通信的后端云基础设施与数据中心架构中，打破了十年前业界认为“Arm 只能做移动端、搞不了大型服务器”的固有偏见。 追溯这一奇迹的源头，Arm 的底层基因其实形成于 30 多年前剑桥创立之初的一场关键冒险——为苹果早期的掌上电脑 **Apple Newton** 设计芯片： - **前互联网时代的严苛限制**：那是一个既没有移动互联网、也没有成熟语音识别与指纹识别的拓荒时代，作为一款开创性的 PDA，设备必须依靠干电池供电，这就把“极低功耗（low power）”刻进了芯片定义的死命令中。 - **塑料封装（Plastic Package）的成本倒逼**：当时半导体工业散热表现较好的是陶瓷封装（ceramic package），但成本极其高昂且热学特性沉重；为了在控制成本的同时保证散热，Arm 团队必须将处理器的发热与功耗压榨到极致，使其能够塞进廉价的塑料封装中。这使得早期的 Arm 1 处理器从诞生第一天起，就是彻头彻尾围绕“靠电池运转”而量身定制的。 --- ### [08:05](https://www.youtube.com/watch?v=Y8guMe665RE&t=485s) 公用电话亭时代的 FAE 与移动计算的原动力 在那个绝大多数计算都必须插着电缆、固定在办公桌前的年代，“低功耗与移动性”在很多人看来是一件无足轻重的非主流需求。但对于亲历技术演进的人来说，便携性带来的效率跃进宛如魔法。Rene 回忆起早期大哥大和大型卫星电话问世时的震撼：“在当年，如果能拿着一台大型设备在外面走动 20 分钟而不用插电，或者有任何能运行复杂任务的机器能撑上 30 到 40 分钟的电池续航，人们都会觉得这简直是不可思议的奇迹与划时代的变革。” 为了说明当年的“移动荒漠”，Rene 分享了自己职业生涯早期担任现场应用工程师（FAE）时的趣事： - 当时他每天开着车奔波于各大客户之间，整个人与外界处于断联状态。为了向总部汇报并获取客户更新，他必须一路搜寻路边的公用电话亭（payphone）。 - 停好车投币打给办公室后，前台总会念出一长串留言纸条，而上面的内容往往极其简短且令人哭笑不得，比如“速回电”、“没听懂你到底问了什么”或者干脆就是“我很忙”。 - 直到后来车里终于装上了一部早期车载电话，能够实现随时随地远程沟通时，那种震撼感让所有人觉得“这是生产力的终极飞跃”——相比之前在公用电话亭排队联络，过去的日子简直就像生活在 19 世纪的西联汇款（Western Union）电报时代。 这一段真实的历史切片，恰恰折射出 Arm 创办至今的核心叙事主线：对移动与电池供电效率的极致追求，最终从一种边缘的技术偏好，演变成了重塑整个全球计算范式的根本推手。 ### [09:33](https://www.youtube.com/watch?v=Y8guMe665RE&t=573s) 从“极度受限的劣质架构”到计算世界的默认核心 主持人回顾了 ARM 最初诞生的时代背景：它是在极端严苛的散热与低功耗需求下构想出来的，当时锂电池技术远不如今天成熟，整个世界的能效意识也处于初级阶段，处理器必须保证不能迅速耗尽电量。在那时人们的眼中，这样一种能力极其有限、甚至显得有些“糟糕简陋（crappy）”的处理器架构，绝不可能成为世界上最先进、最复杂计算应用的主流架构。然而几十年后，事实却恰恰相反。为了探讨 ARM 是如何一步步走到今天统治地位的，Acquired 秉承深挖历史的传统，将话题拉回到精简指令集计算机（RISC）的诞生之初。主持人邀请身为 ARM CEO 的 Rene Haas 客串一把“计算机科学史教授”，讲解 RISC 与复杂指令集计算机（CISC）之间那段旷日持久的路线之争。 ### [10:28](https://www.youtube.com/watch?v=Y8guMe665RE&t=628s) RISC 对决 CISC：精简与臃肿的底层逻辑 Rene Haas 讲述了 RISC 的概念起源，指出这一理念最初主要由加利福尼亚大学伯克利分校的大卫·帕特森（David Patterson）等学者提出。在早期计算时代，像英特尔 x86 或摩托罗拉 68000 系列这样的架构都被划归为 CISC。CISC 架构的本质在于它拥有海量指令，并且为了让以往编写的软件能够持续运行，不得不背负沉重的历史包袱将所有旧指令一代代沿用下去。在底层实现上，一条复杂指令意味着在时钟周期内必须执行多次微操作，导致晶体管始终处于过度运转状态，白白消耗大量功耗。 主持人用形象的例子补充道：在 CISC 架构下，哪怕只是一个单指令周期，硬件也必须为了应对各种极端复杂的情况预留充足带宽——例如在同一操作里完成“从内存拉取数据、加载到寄存器、执行相加并返回结果”。Rene 赞同并给出了另一个生动的比喻：一条 CISC 指令就像命令一个人“向前走三步，向左走两步，斜对角走两步，再向右走三步”。虽然遇到完全匹配该动作的专门任务时效率很高，但绝大多数程序根本用不上这种复合指令。然而一旦有软件依赖它，芯片架构就不得不永远保留这套笨重繁琐的机制。相比之下，RISC 则倡导“向前走一步、向后退一步、向左走一步、向右走一步”这种原子化的简单动作（如加、减等），通过高效组合简单指令，以极高的能效解决问题。 ```mermaid flowchart TD accTitle: CISC 与 RISC 架构设计哲学对比 accDescr: 对比 CISC 复合成套复杂动作但背负历史债务，与 RISC 依赖原子基础动作从而实现极致能效。 subgraph CISC["CISC 复杂指令集 (如 x86, 68000)"] direction TB C1["复合动作指令<br/>'前进三步+左拐两步+斜走两步'"] --> C2["硬件层背负历史包袱<br/>为满足复杂场景预留庞大带宽"] C2 --> C3["晶体管过度频繁运转<br/>单周期微操作多、功耗极高"] end subgraph RISC["RISC 精简指令集 (如 MIPS, ARM)"] direction TB R1["原子化基础指令<br/>'前进单步' / '基础加减' / '单纯Load'"] --> R2["精简高效的硬件执行通道<br/>编译器组合基础指令构建逻辑"] R2 --> R3["时钟利用率高、无冗余消耗<br/>低功耗、节省运行内存占用"] end ``` - **诞生节点**：20 世纪 70 至 80 年代，以 MIPS 为代表的早期 RISC 处理器开始涌现，将指令集精简理念推向工业实践。 - **能效与成本双赢**：在那个时代，运行程序所需的内存极其昂贵；RISC 架构不仅大幅降低了功耗，还能减少程序在内存中的占用空间（memory footprint），带来了巨大的商业与技术收益。 - **软件编译范式的变革**：随着编程模式从汇编语言等底层开发转向 Fortran、Pascal 以及后来的 C/C++ 等高级语言，编译器开始接管生成底层机器码的工作。如果编译器频繁调用底层繁杂沉重的指令，就会不可避免地编译出体积臃肿、执行低效的代码，进一步放大了 CISC 的固有缺陷。 ### [15:23](https://www.youtube.com/watch?v=Y8guMe665RE&t=923s) IBM PC 的意外开放与 x86 软件护城河的筑起 主机和小型机时代硬件体量庞大，IBM 这类巨头全包软硬件，完全不用顾虑功耗与指令集复杂度。主持人提出疑问：按常理推断，当计算产业向个人电脑（PC）时代迁移时，本该是功耗更低、更优雅的 RISC 架构大展身手的完美窗口，为何 CISC 反而在 PC 时代大行其道？是 RISC 出现太晚，还是当时的 ARM 尚未成型？ Rene Haas 随即回溯了 1981 年 IBM PC 诞生背后的历史反差。彼时的 IBM 在大型机（如著名的 System/360）与小型机领域占据绝对霸主地位，是软硬件全包、服务一体化交付的“一站式商店”。但面对 20 世纪 70 年代末到 80 年代初苹果公司依托摩托罗拉芯片掀起的家庭电脑浪潮，以及 TRS-80、Commodore 等各类新兴机型的夹击，IBM 在 PC 领域已然严重落后。急于入局的 IBM 做出了一个极不符合其巨头基因的决定：既没有自研 CPU，也没有自研操作系统，而是采用完全开放的外部采购策略。操作系统方面，他们没有谈成加里·基尔代尔（Gary Kildall）公司的 CP/M-80，最终敲定了微软（DOS）；处理器方面，他们放弃了当时被视为行业领头羊的摩托罗拉，最终选用了英特尔的 8086。 这一妥协造就了日后席卷全球的“IBM PC 兼容机”洪流。由于 IBM PC 完全由公开零部件拼凑而成（英特尔处理器、微软 DOS、Conner/Maxtor/Seagate 硬盘，以及台湾厂商供应的显示器），任何第三方都可以组装出功能相同的克隆机（clones）。伴随 PC 兼容机市场的爆炸式增长，Lotus 1-2-3 等杀手级软件纷纷基于 x86 指令集编写并进行极致优化。 Rene Haas 指出，芯片架构领域有一个残酷的“肮脏小秘密（dirty little secret）”：历史上曾诞生过无数优秀的 CPU 架构——无论是 SPARC、MIPS、DEC Alpha，还是他曾工作过的 Tensilica，抑或是 29000、68000 和 ARC——**一款 CPU 架构的真正价值，完全取决于在其之上编写的软件生态，以及这些软件究竟能存活多久**。正是因为海量 PC 软件与 x86 牢牢绑定，形成的巨大软件惯性与向后兼容性需求，让 CISC 架构在 PC 时代反客为主，铸就了牢不可破的垄断地位。 ### [19:16](https://www.youtube.com/watch?v=Y8guMe665RE&t=1156s) 架构锁定的铁律与巨额迁移成本 IBM PC 及其兼容机生态（由 Compaq、Dell、Gateway 以及早已淹没在历史中的 AST Research 等厂商共同构建）的诞生，不仅奠定了个人电脑平台的格局，也让 Intel 的 x86 架构一跃成为行业事实上的标准。Rene Haas 指出，虽然学术界和工程界普遍认为 RISC 在理论上更加优秀，但这在商业现实面前无关紧要——一旦 IBM 敲定了 8086 芯片，DOS 以及后来的 Windows 针对其深度优化，整个 PC 时代的技术列车就此全速驶上 CISC 的轨道。苹果是少数几家经历过多次底层架构大震荡的公司：从早期的 Motorola 68000 转向苹果、IBM 与摩托罗拉联合组建的 PowerPC 联盟，打造出一种兼具 RISC 与 CISC 特征的混合产物。这种典型的 90 年代苹果式执念——既非纯粹 RISC 也非纯粹 CISC，而是另起炉灶搞私有技术——背后伴随着极高的架构迁移代价。 尽管 PowerPC 的软件迁移任务在当年已极其沉重，但与如今庞大臃肿的现代软件堆栈相比仍属小巫见大巫。Haas 提到自己近期在录制播客时与 NVIDIA 首席执行官黄仁勋（Jensen Huang）的交谈，黄仁勋当时精辟地概括了一句铁律："软件永不死"（software never dies）。正是由于海量既有代码的惯性，推动任何可编程架构的底层替换，都必须依赖功耗或成本上决定性的范式跃迁。人们常挂在嘴边的"十倍优势法则"未必需要精确达到 10 倍，但"绝对不可能只是 15%"，必须是颠覆性的飞跃，或者开辟全新场景。就像当年苹果研发掌上电脑 Newton 时，x86 的发热与功耗根本无法进入备选名单；而要想在已有软件生态的成熟场景下让用户"换马"，其性能代差的量化标准，或许就鲜明地体现在配备 Intel 芯片的最后一代 MacBook Pro 与搭载自研 M1 芯片的 Geekbench 跑分差距之中。 - **CISC 锁定的核心推手**：IBM 选型 8086、DOS/Windows 的排他性优化，以及 PC 兼容机阵营的规模效应。 - **架构迁移门槛**：常规 15% 的增量改进无法打破既有生态壁垒，必须依赖足以开辟新形态的功耗优势或跨代级的性能飞跃。 - **软件生态惯性**：黄仁勋所言的"软件永不死"，揭示了底层架构切换的最大阻力并非硬件设计本身，而是上层代码库的重构成本。 ### [22:35](https://www.youtube.com/watch?v=Y8guMe665RE&t=1355s) 赞助时刻：Vanta 自动化信任与合规管理 播客插入了老朋友 Vanta 的赞助介绍。作为领先的信任管理平台，Vanta 致力于实现安全审查与合规监控的自动化，涵盖 SOC 2、ISO 27001、GDPR 以及 HIPAA 等主流认证体系，帮助企业免除繁琐冗长、极度消耗资源的流程。 节目主持人援引 Jeff Bezos 广为人知的商业哲学作为类比——企业应该将所有精力与资源聚焦在"真正让自家啤酒变得更好喝"的事情上，即那些能切实推进产品体验与客户价值的环节，而将其余无关核心风味的杂务外包出去。合规审查虽是客户与合作伙伴的硬性准入要求，但对产品本身的功能体验没有任何加成。Vanta 通过单一口径的软件控制台，借助 API 无缝打通全套业务基础设施，消除了零散的手工电子表格和低效的人工比对；目前平台已集成超 300 种外部工具并引入 AI 能力，支持实时共享动态安全状态，服务全球超过 7,000 家企业客户。 ```mermaid flowchart TD accTitle: ARM Shared Success Licensing Model accDescr: Diagram illustrating ARM's two-part shared success licensing model consisting of upfront license fee and per-unit royalty. subgraph Client["芯片客户 / 终端厂商"] C_Eng["省去自研团队成本"] C_Prod["芯片量产上市销售"] end subgraph ARM["ARM 知识产权模型"] A_IP["ARM 核心架构设计 (R&D)"] A_Upfront["前期授权费 (License Fee)<br/>替代客户自身高额研发投入"] A_Royalty["量产版税分成 (Royalty)<br/>与客户出货规模深度绑定"] end A_IP --> A_Upfront A_Upfront -.替代研发成本.-> C_Eng C_Eng --> C_Prod C_Prod --> A_Royalty ``` ### [24:27](https://www.youtube.com/watch?v=Y8guMe665RE&t=1467s) ARM 的商业模式革命：授权与版税机制 在 1990 年代初期，主流半导体厂商如开发 x86、68000、AMD 29000 等微处理器的巨头，全部采用高度垂直整合的模式，众多企业斥巨资雇佣大批工程师从零设计专属微处理器。ARM 创始人 Robin Saxby 和 Tudor Brown 等人敏锐地意识到，通用微处理器之间的差异化极其有限，如果直接造芯片去硬碰已极度拥挤的市场无异于自杀。于是他们开创了一种在当时极具颠覆性的商业策略：不再自产实体芯片，而是将精巧的处理器内核设计授权给业界，采用"前期授权费（Upfront License Fee）+ 后续版税（Royalty）"的利益共享模式（Shared Success Model）。 这一模式的精妙之处在于，前期的授权费仅作为客户省下庞大自建团队研发支出的"替代支出"（Proxy for R&D），费率适中；而当客户产品大规模量产上市后，ARM 再按出货量抽取一定比例的版税提成。这种双赢机制在如今看来顺理成章，但在 90 年代初却面临着生态工具的严峻空白。当时 Synopsys 的高层次硬件描述语言才刚刚起步，Cadence 允许将第三方 IP 整合到后端设计流程的能力尚处萌芽期，配套的编译器和开发软件更是匮乏。因为是一张白纸，ARM 既没有现成的操作系统支持，也没有第三方应用程序生态，要在通用计算领域撕开一道口子几乎是不可能的任务。 - **垂直整合的打破**：芯片设计从"各家自建自研"转向标准 IP 模块授权。 - **共享成功模式**：前期较低门槛授权降低客户启动风险，后期依赖终端出货版税实现规模化收益。 - **早期 EDA 生态制约**：当时的芯片设计流程、语言工具链与软件生态均未成熟，IP 复用（IP reuse）概念在产业界仍属先锋尝试。 ### [27:36](https://www.youtube.com/watch?v=Y8guMe665RE&t=1656s) 诺基亚与德州仪器：基带中的关键战役 ARM 早期没有任何主流桌面软件生态的庇护，公司的生死存亡取决于一次充满戏剧性的历史契机——90 年代中期 2G/GSM 移动电话的全面爆发。当时，德州仪器（Texas Instruments, TI）是 GSM 移动基带芯片市场的最大供应商，而其核心大客户正是如日中天的诺基亚（Nokia）。在 GSM 手机中，基带芯片需要一颗极低功耗、体积小巧的微处理器来负责底层通信协议栈与状态机的调度运行。 在当时的移动通信设备中，根本不存在所谓的外部应用生态，正如主持人一针见血的调侃："当时的应用程序就是电话本身"（The application was the phone）。TI 在评估自研路线及市面上现存的所有处理器方案后，发现没有一款能比 ARM 架构更加简洁优雅。对 TI 而言，其核心竞争优势与利润来源是射频（Radio）与模拟基带算法，自研一颗控制类 CPU 毫无商业附加值。TI 决定引入 ARM 的 CPU 内核并将其集成进基带芯片组，驱动诺基亚的首批 GSM 手机走向全球市场。这次关键的"中标"（Design Win），成为了 ARM 发展史上无可替代的造市之战（market maker）。 ### [29:02](https://www.youtube.com/watch?v=Y8guMe665RE&t=1742s) 标准化的力量：从诺基亚基带到功能机生态 播客主播将 ARM 的发展历程与台积电（TSMC）当年的路径作类比：台积电从半导体制造栈的最底层切入，ARM 则在其上一层（IP 设计层）；两者起步时都没有去碰主流庞大的 PC 市场，而是专注于一些非先进制程或小众部件（例如德州仪器 TI 为手机打造的 CPU 组件）。这种早期不图暴利、依靠广泛授权成为行业底层标准的打法，极具 Windows 效应的影子——一旦全行业都基于你的标准进行开发，你在产业链中就会掌握巨大的话语权。 Rene Haas 证实了这一点：ARM 当年通过德州仪器的芯片打入诺基亚（Nokia）手机后，迅速确立了在 GSM 手机基带芯片领域的行业事实标准（de facto standard）。在那个功能机时代，手机上甚至没有真正的操作系统或 App，ARM 之所以能成为通用选择，完全是因为其出色的能效比与性能刚刚好契合需求。随后，GSM 手机逐渐变聪明，开始运行塞班（Symbian）操作系统。正是借由塞班，ARM 建立起了最早的一批开发者生态与软件粘性，开始有了“街头信誉”（street cred）。 ```mermaid flowchart LR accTitle: ARM早期在移动通讯市场的立足与扩散路径 accDescr: 展示ARM从德州仪器基带芯片切入诺基亚手机，进而成为GSM事实标准并孕育塞班软件生态的过程 A["ARM 处理器 IP"] --> B["TI 德州仪器基带芯片"] B --> C["诺基亚 GSM 手机"] C --> D["其他 GSM 芯片厂商跟进<br/>(确立 de facto 标准)"] D --> E["运行 Symbian 操作系统<br/>(形成初始软件生态粘性)"] ``` --- ### [31:00](https://www.youtube.com/watch?v=Y8guMe665RE&t=1860s) 关键踏板：初代 iPod 与移动操作系统的分水岭 如果说塞班让 ARM 初露锋芒，那么真正将其推向全新维度的转折点则是苹果的初代 iPod。2000 年代初，苹果推出了搭载微型东芝硬盘、机械滚轮和极简操作界面的第一代 iPod。在实质上，这部设备就是一个微型嵌入式计算机，而它的“心脏”正是 ARM 处理器。苹果在 iPod 上的成功运用，让 ARM 积累了在超便携设备上支撑操作系统与专用 UI 的实战经验。 时间来到 2000 年代中叶，苹果内部开启了秘密打造触屏设备（无论先立项的是 iPad 还是 iPhone）的探索，摆在团队面前的核心难题是：iPhone 的处理器究竟该选谁？当时业界都在追捧上网本（Netbook），整个 PC 行业甚至一度错判上网本就是未来的计算形态。英特尔（Intel）为了抓住这波浪潮，推出了低功耗 x86 芯片——Atom（凌动）。然而，面对真正的智能手机，英特尔陷入了经典的“创新者的窘境”（Innovator's Dilemma）。 - **初代 iPod 的硬件构成**：东芝特制微型硬盘、机械滚轮控制器、简易单色屏幕，核心运算全部由 ARM 架构芯片承载。 - **英特尔 Atom 的局限**：本质上是从 Core i7、i5、i3、奔腾（Pentium）、赛扬（Celeron）等桌面级架构强行“自上而下阉割”出来的产物，其功耗与成本虽勉强能应付上网本，但塞进体积严苛、依靠电池供电的手机中依然显得过热且费电。 - **创新者的窘境**：英特尔习惯了销售高性能、高毛利的桌面与服务器芯片，极难真正从零开始设计出极致低功耗的芯片；而 ARM 则始终坚守“自底向上”（Bottom-up）的超低功耗设计哲学。 --- ### [33:34](https://www.youtube.com/watch?v=Y8guMe665RE&t=2014s) 苹果内部的技术路线之争：Intel 方案 vs. iPod 派系 当时英特尔在苹果内部拥有极高的威望与地位。2005-2006 年间，苹果刚完成了从 PowerPC 到英特尔 x86 的世纪大迁移。为了让现有 Mac 软件能顺利重新编译并在英特尔平台上运行，苹果编译器团队付出了多年难以置信的心血。当时所有 Mac 笔记本以及 OS X 系统（包括后续的 Tiger、Leopard、Snow Leopard）全都针对 x86 做了深度优化。 因此在 2006 至 2007 年规划 iPhone 时，苹果内部爆发了一场路线大辩论： 1. **Intel 派系**：主张顺水推舟采用英特尔 Atom 芯片。其最大优势在于软件复用——直接把现有的 Mac OS X 进行裁剪降级，整体生态和编译器无缝继承。 2. **iPod 派系**：主张以 iPod 的 ARM 架构为基石打造全新的系统（即后来的 iOS）。他们坚信移动设备是完全不同的物种，必须从一张白纸重新出发，追求极限的功耗与能效，只有这样才能做出超越时代的产品。 最终，iPod 团队的理念赢得了胜利。虽然苹果并没有百分之百从零写起——他们继承了 Mac OS X 的 Unix/Mach 内核，但重构了运行时并重新编写了针对 ARM 的底层编译器——这不仅造就了初代 iPhone，也成为 ARM 商业史上最关键的一次 Design Win。 --- ### [36:17](https://www.youtube.com/watch?v=Y8guMe665RE&t=2177s) 移动革命的连锁反应与“炖汤里的水” 随着 2007 年 iPhone 惊艳问世，Andy Rubin 领导的 Android 生态（如 HTC、三星等）迅速跟进。由于开源社区早已在 Linux 上围绕 ARM 做了大量适配工作，ARM 顺理成章地成为了整个移动互联浪潮的绝对事实标准。到了 2007-2008 年，iPhone 与 Android 形成了无可阻挡的双重合力。 主播提问：在这场移动盛宴的“大杂烩”（Stew）里，ARM 到底扮演了什么角色？芯片具体又是谁造的？Rene Haas 指出，如果把 2007 年的 ARM 类比为 1981 年 PC 时代的英特尔，两者最大的区别在于商业模式： - **英特尔模式**：英特尔既拥有 x86 架构，又自己造芯片。当年仅因 IBM 出于供应链安全性强制要求“第二供应商”（Second Source），才不情愿地向 AMD 交叉授权了 x86。 - **ARM 模式**：ARM 从不垄断制造，而是将指令集架构与核心 IP 开放授权给全行业。在初代 iPhone 中，应用处理器是由三星（Samsung）基于 ARM 架构代工/定制制造的，之后苹果又逐步转向台积电（TSMC）。三星、高通（Qualcomm）、英伟达（NVIDIA，其 Tegra 系列芯片）等群雄并起，全都在 ARM 架构上制造自家的手机芯片。 这种“全行业多供应商”的平台模式释放出了极强的网络效应。对于下游整机厂商来说，再也不必受制于单一半导体巨头；而对于整个生态系统而言，ARM 架构就像是这锅大杂烩里最基础、不可或缺的“底汤之水”——虽然人们可能很少单挑出来尝它的味道，但没有它，整锅汤便根本无法成立。 ### [038:41] 移动浪潮中的“卖水人”与苹果的自研芯片之路 在智能手机行业爆发之初，ARM 处于一种极其独特且无可撼动的关键地位。正如 Rene Haas 所比喻的：“我们负责供水。任何人想要进入智能手机市场，除非通过 ARM，否则根本无路可走。”对于像苹果这样的设备厂商而言，这种生态带来了无与伦比的可移植性（Portability）。苹果在设计下一代 iPhone 时，市场上已经存在大量基于 ARM 的处理器供其挑选；这意味着只要选定 ARM 架构，就能拥有广阔的供应商海洋——不仅可以向采用 ARM 设计的三星购买芯片，甚至在收购 P.A. Semi 之后，苹果能够转变为自研芯片的厂商。 ARM 的合作机制提供了两条明确的路径：客户要么直接购买其他半导体厂商基于 ARM 设计制造的现成芯片，要么如果自身“足够勇敢、富有才华且聪明过人”，ARM 也会授予其权利，允许他们完全自行构建兼容 ARM 指令集的芯片。这正是苹果最终选择并开创的自研之路。 ```mermaid flowchart TD accTitle: ARM客户选择路径与商业模式 accDescr: 展示设备厂商获取ARM能力的两种主要方式以及对应的商业收费模式。 subgraph S1["厂商合作路径"] A["设备厂商需求"] --> B{"技术与实力选择"} B -->|"采纳通用方案"| C["购买第三方ARM芯片<br/>(如Samsung等)"] B -->|"具备顶尖芯片设计团队"| D["获取ARM架构授权<br/>(如Apple自研芯片)"] end subgraph S2["ARM商业模式"] E["前端授权费<br/>(Upfront License Fee)"] F["后端版税<br/>(Royalties)"] end C -.-> F D -.-> E D -.-> F ``` --- ### [039:29] ARM 的双重商业模式：授权费与版税的博弈 进入 2010 年代初期，ARM 经典的商业模式已经清晰确立：主要由前端的设计授权费（Licensing）和后期的芯片出货版税（Royalties）构成。在业务发展早期，由于很多客户尚未进入大规模量产阶段，绝大部分收入都来自授权费。在半导体行业中，这体现为“设计中标”（Design Wins）——虽然成功让客户对该架构做出承诺，但在没有大规模出货之前，根本看不到可观的版税收入。 - **乐观视角（杯子半满）**：如果这些设计未来能够顺利实现量产，前景将不可限量。 - **悲观视角（杯子半空）**：芯片迟迟未能出货甚至面临流片失败，前期押注的技术可能永远无法得见天日。 - **历史收入拐点**：由于芯片量产周期的漫长滞后性，在 ARM 发展历程中的很多年里，授权收入的规模一直大于版税收入，直到生态出货量爆发后两者的比例才发生结构性反转。 --- ### [040:35] 架构授权与绝不可打破的 ISA 铁律 在具体的授权方式上，ARM 区分了“实现设计授权”（Implementation License）与“架构授权”（Architectural License）。前者相当于 ARM 提供完整的房屋蓝图，内部拥有芯片设计工程师使用 Cadence 和 Synopsys 等 EDA 工具进行物理布局布线（Floor Planning），直接交付成品级核设计。而针对那些认为软硬件深度协同或自身工程实力能造出超越 ARM 官方性能水准的顶尖客户，ARM 则提供架构授权，允许其自行构建微架构实现。 然而，这种架构授权有着绝对不可逾越的底线：**客户绝不允许添加独占的自定义指令来破坏指令集架构（ISA）的兼容性**。Haas 强调，软件兼容性是 ARM 的核心超能力。如果允许客户随意增添特有指令（例如客户 A 引入“加速”指令，客户 B 引入“2倍加速”，客户 C 引入“3倍加速”），软件开发者就无法预测底层硬件支持情况，最终只能退化到“寻找最低公共分母”（Lowest Common Denominator）的保守编码方式，破坏整个开发者生态。如同当年 IBM PC 兼容机生态中 x86 的坚固护城河一样，Compaq、Dell 等克隆厂商之所以能够复制 PC 的繁荣，正是因为软件无需修改即可在不同机器上完美运行。 - **自研微架构的巨大成本**：如今采用纯架构授权自行设计 CPU 核心的客户越来越少。设计一颗现代高性能 CPU 核心需要投入三四百人甚至上千名工程师，工作极其艰巨复杂。 - **ROI 与研发资源分配**：如果耗费数亿美元团队资源最终仅能换取 5% 至 15% 的微弱性能优势，其投资回报率并不理想；多数厂商宁可将这批顶尖工程师投入到自身特有的专有 IP（如专用加速器）上，转而直接采用 ARM 现成的 CPU 设计。 --- ### [044:11] CPU 墓地、软件飞轮与硅谷芯片风投的退潮 回顾 20 世纪 80 年代末至 90 年代初的微处理器黄金期，市场上曾充斥着大量架构优异、性能顶尖的 CPU，如 Motorola 68000、PowerPC、AMD 29000、DEC Alpha 和 SPARC 等。然而，它们最终几乎全数进入了“CPU 墓地”（Graveyard of CPUs）。之所以被时代淘汰，根本原因在于软件生态飞轮效应的确立：一旦庞大的软件开发堆栈锁定在特定架构上，新硬件如果缺乏成熟的软件配套，就注定在市场上枯萎边缘化。 伴随着互联网泡沫时代的到来，科技行业的资本流向发生了剧烈倾斜： 1. **软件行业吸纳海量资本**：大量风投资金涌入 SaaS、订阅制以及企业级云端软件，推动了运行在云端、网络交换机和汽车等各层级软件复杂度的爆发式增长。 2. **半导体初创公司投资枯竭**：与软件热潮形成鲜明对比的是，硬件半导体领域的风险投资长期陷入低谷。曾经孕育底层计算架构创新的硅谷，出现了“名为硅谷（Silicon Valley），但半导体（Silicon）初创企业却根本融不到风投资金”的巨大讽刺。 作为这段历史的亲历者，Haas 回忆起自己在 1990 年代末加入的可配置处理器初创公司 Tensilica（由前 Synopsys 与 MIPS 团队于 1997 年创立，主打可定制指令扩展，2004 年 Haas 离职，公司最终于 2012 年被 Cadence 收购，累计出货超十亿颗内核）。在 Tensilica 和另一家名为 ARC 的公司之后，市场上几乎再也没有面向通用 CPU 创新的半导体初创企业诞生。 ```mermaid flowchart TD accTitle: 软件飞轮与半导体投资趋势对计算架构格局的影响 accDescr: 展示资本流向软件行业导致架构收敛，最终由x86与ARM主导市场的过程。 subgraph G1["资本流向分化"] A1["互联网与SaaS繁荣"] --> B1["软件生态飞轮加速<br/>(依赖既有架构兼容性)"] A2["硬件VC投资锐减"] --> B2["新兴CPU初创停滞<br/>(除Tensilica/ARC等极少数)"] end subgraph G2["竞争格局洗牌"] C1["历史架构凋零<br/>(Alpha, SPARC, PowerPC入土)"] B1 & B2 --> C1 C1 --> D["仅剩两大支柱架构"] end subgraph G3["双雄格局"] D --> E1["x86<br/>(40余年PC及服务器积累)"] D --> E2["ARM<br/>(极高能效比支撑边缘与数据中心)"] end ``` --- ### [047:14] 数据中心架构的历史洗牌与能效对决 这一深刻的行业变迁最终导致全球计算架构格局大幅收缩，仅存两大统治级赢家：拥有 40 余年历史积淀的 x86，以及席卷移动与边缘的 ARM。当话题转向“ARM 为何能够杀入数据中心”时，Haas 指出了两个决定性因素： - **客户的可选空间极其有限**：市场并不存在 17 种不同架构供数据中心挑选，除了传统的 x86，符合庞大通用软件生态要求的现代架构唯有 ARM。 - **算力负载下的功耗红线（Power Efficiency）**：无论是承载超大规模的通用计算负载，还是应对当下因 AI 模型爆发而激增的加速计算需求，数据中心对于极端能源效率的追求已成为决定成本与扩展规模的核心瓶颈，这恰恰是流淌在 ARM 基因里的最强优势。 ### [047:48] 胜者通吃与软件护城河：芯片架构的“晶圆厂级”壁垒 在探讨通用处理器市场的竞争格局时，ARM 首席执行官 Rene Haas 总结了 ARM 之所以能够登顶的三大决定性支柱：其一是架构天然具备卓越的低功耗基因；其二是业界在 ARM 架构上积累了惊人数量的软件生态创新；其三则是历史投资逐渐收窄导致其他替代选项彻底消失。Acquired 主持人将这种胜出戏称为“剩者为王”（Last man standing）的同义反复——随着竞争对手在沿途相继倒下，最终留在赛道上的自然就是赢家。 Rene 反驳了这一过程是因为“市场缺乏吸引力导致对手弃赛”的观点。恰恰相反，计算市场的吸引力空前高涨，但横亘在所有新入局者面前的，是一座由软件投资构筑的绝壁。Rene 打了个形象的比方：如果今天有人突然跑到市场上宣布“我们要基于摩托罗拉 68000（Motorola 68000）架构打造一款全新的 SoC 芯片”，根本不会有任何现成软件能在这块芯片上运转。主持人随即指出，这与晶圆代工业的垄断演进如出一辙——台积电（TSMC）、三星（Samsung）和格芯（GlobalFoundries）等少数巨头的形成，正是因为先进制程的资金门槛高不可攀。Rene 认同这一类比并进一步剖析：晶圆厂拼的是直接资本开支（CapEx），比如建设一座 2 纳米及更先进的晶圆厂需要耗资 300 亿至 350 亿美元；而处理器架构拼的则是软件生态的运营投入（OpEx），全球超过 2000 万名常年围绕 ARM 开发的软件工程师，共同汇聚成了难以撼动的庞大行业惯性。 - **三大胜出因素**：低功耗硬件设计、数十万亿美元级软件生态积累、竞争架构投资枯竭导致的生态排他性。 - **软件重力场**：即使硬件能复制，全球 2000 多万名专为 ARM 编写代码的开发者所形成的软件生态迁移成本几乎不可逾越。 - **重资产与重生态的对比**：先进制程晶圆厂依赖 300 亿至 350 亿美元的直接 CapEx 门槛，处理器架构则依赖数十万开发者生态数十年的累积 OpEx 门槛。 --- ### [050:14] 数据中心颠覆：ARM 如何击败 x86 的封闭垄断 主持人抛出了一个令人惊叹的技术反差：一款最初为了“不融化塑料外壳”、专为超低功耗掌上设备设计的架构，如今竟然在数据中心最严苛的重负载计算、AI 训练与推理中反客为主，成为主导架构。当被问及 x86 架构是否还有立足之地、全世界是否终将全面转向 ARM 时，Rene 坦承自己虽然身为 ARM CEO 很难做到绝对中立，但客观市场力量确实正在向 ARM 极度倾斜。 Rene 指出，ARM 与 x86 的核心对决首先体现在商业模式上：ARM 实行开放授权模式，客户设计的芯片可以在任何晶圆代工厂制造，也可以被任何芯片公司集成；反观 x86 阵营，客户完全受制于英特尔（Intel）和 AMD 这两家供应商。虽然 AMD 将制造交给台积电，英特尔近年也开始把部分产品外包给台积电，但客户不仅要押注这两家公司本身的研发节奏，还被迫捆绑接受它们自带的通信模块、网络存储和加速计算 IP。主持人插话指出，英特尔与 AMD 很难去模仿 ARM 的授权玩法，因为一旦从高毛利整片销售模式切换到低毛利的 IP 授权模式，现有商业逻辑就会崩溃。 这种商业模式差异在超大规模云厂商（Hyperscalers）自研芯片浪潮中爆发出了决定性优势。微软 Azure、谷歌 GCP 和亚马逊 AWS 均打造了基于 ARM 的自研服务器芯片，并在同等对比（Like-for-like）下普遍实现了 60% 的性能与能效提升。Rene 强调，这不仅仅是 ARM 指令集（ISA）本身的能效优势，更关键在于系统级定制的自由度——超大体量云厂商可以自由搭配定制内存接口、定制存储模块、定制服务器刀片、定制互连结构（Interconnect）与专用卸载引擎（Offload），从总体拥有成本（TCO）角度实现无与伦比的架构自由度。英特尔只能拿出一款固定规格的成品芯片对客户说“这就是给你准备的产品，自己想办法适配”，而 ARM 客户却能打造与自身云基础设施完美咬合的芯片。此外，市场上既有 Ampere 提供的现成标准品，也有英伟达（NVIDIA）Grace 这样的深度定制 CPU。在 Grace Blackwell 架构中，英伟达通过 NVLink 将 ARM CPU 与 GPU 高速互联，并打通 HBM 显存与 CPU 内存通道，这种高度协同的设计在封闭的 x86 世界中完全无法实现。更重要的是，Grace 能够原生运行所有主流操作系统和 ARM 通用计算软件栈，使客户现有的云端软件无需重构即可无缝部署在 AI 集群中。讲到这里，Rene 忍不住打趣道：“我原本是想帮 x86 说几句公道话的，结果讲着讲着又变成夸了一整天 ARM，这真的很难克制！” ```mermaid flowchart TD accTitle: ARM开放生态与x86封闭体系对比 accDescr: 展示ARM的开放多工厂授权及超大规模客户深度定制模式与x86固定采购模式的区别 subgraph ARM_Model["ARM 开放生态模式"] A["ARM 架构授权 & IP"] --> B["多样化客户群体"] B --> C["云巨头自研<br/>(AWS Graviton / Azure / GCP)"] B --> D["加速计算深度融合<br/>(NVIDIA Grace Blackwell)"] B --> E["通用芯片供应商<br/>(Ampere 等)"] C & D & E --> F["自由定制互连、内存与刀片<br/>代工选择灵活 (台积电/三星等)"] end subgraph x86_Model["x86 双寡头封闭模式"] G["x86 专利封闭双雄<br/>(Intel / AMD)"] --> H["标准化固定芯片成品"] H --> I["数据中心客户强制适配<br/>无法改变互连与内存底层逻辑"] end ``` - **60% 能效提升**：AWS、微软和谷歌的 ARM 自研芯片在同规格基准测试中展现出约 60% 的整体性能与能效红利。 - **架构融合壁垒**：NVIDIA Grace Blackwell 系统依靠 NVLink 与特有显存接口实现了 CPU 与 GPU 的极速耦合，彻底打破了传统 x86 系统总线带宽的物理瓶颈。 - **双重业务形态**：在数据中心领域，ARM 既支持 Ampere 等第三方提供货架式通用 CPU，又支持超大规模客户进行百万颗级别的完全自研芯片。 --- ### [054:08] 1500 亿美元估值与 AI 时代的算力饥渴：告别“足够好” 谈到公司的财务现状与资本市场表现，主持人列举了一组惊人的数字：在录制节目时的最近一个季度，ARM 实现了 9.39 亿美元的单季营收，年化营收规模（Run rate）逼近 40 亿美元，而公开市场给予的市值则高达约 1500 亿美元。面对“为什么投资人对 ARM 的未来如此狂热”的提问，Rene 从底层技术复利和终端市场变迁给出了深度剖析。 首先是出货量的大势所趋：全球每年制造的芯片总量在不断攀升，其中基于 ARM 架构的芯片占比持续增加。从家庭中的汽车、智能摄像头、智能灶台到各种物联网终端，无一例外都在由 ARM 驱动且处于快速增长赛道。其次，AI 技术引爆了对前所未有的算力密度的渴求。针对市场上对 AI 落地价值的种种短期质疑——例如主持人调侃 René 听起来很像质疑软件投资回报率的 Salesforce CEO 贝尼奥夫（Marc Benioff），吐槽“一个月花 40 美元买 Copilot 到底有没有 ROI”——Rene 直言，用近期的财务变现模型来评估 AI 是一种极其短视的误判。AI 的性质更接近汽车的发明、工业革命、智能手机革命或互联网革命，是一场全方位的底层生产力跃迁。 Rene 回忆起自己自 1984 年大学毕业进入德州仪器（Texas Instruments, TI）以来四十年的半导体从业生涯，以及此前在英伟达工作的经历。在算力和加速计算领域，所有从业者面临的最大梦魇莫过于市场陷入“足够好”（Good enough）的停滞状态。例如在 2000 年代末到 2010 年代初的“上网本”（Netbook）时代，当时的应用生态根本无法消耗更多的算力，消费者只需一台 199 美元的廉价简陋电脑就能完成日常全部任务，导致芯片算力创新步伐明显放缓。然而，AI 彻底终结了“足够好”的时代。无论是在云端训练类似于 OpenAI “草莓”（Strawberry）级别的下一代前沿大语言模型和前沿视频大模型，还是在边缘端的智能汽车、穿戴设备、头显以及各种家庭设备上跑本地推理（Inference），所有工作流都对算力提出了永无止境的极致需求，这也正是 ARM 未来巨大的增长引擎所在。 - **核心财务数据**：录制时季度营收 9.39 亿美元，折合年化营收近 40 亿美元，公开市场市值约 1500 亿美元。 - **产业革命视角**：将当前 AI 的爆发定性为媲美内燃机、工业革命、互联网和智能手机的底层范式转移，不可用单点软件付费的即期 ROI 一概而论。 - **“足够好”陷阱的终结**：AI 终结了 2000 年代末上网本（199 美元设备）所代表的算力过剩与创新减速期，训练端向巨型多模态与推理扩展，边缘端全场景推理爆发，重新打开了算力需求的指数级增长通道。 ### [57:43](https://www.youtube.com/watch?v=Y8guMe665RE&t=3463s) AI 计算机会与未被优化的边缘架构 在可预见的未来，AI 为教育、药物研发和投资等各个领域带来的变革潜力是极其惊人的，而 Arm 必然会处于这一浪潮的绝对核心。不论是在数据中心、汽车、智能手机，还是在可穿戴设备上，AI 的计算路径都会以某种形式穿过 Arm 架构。Rene Haas 引用了杰夫·贝索斯（Jeff Bezos）的名言作类比：贝索斯曾说他无法想象未来有客户会跑来说“真希望商品能再贵一点”，同样，你也很难想象未来有人会说“真希望 GPT-7 变得更笨一点”。面对现在有人抱怨“还没看到 AI 带来多大用处”的声音，Haas 反而感到兴奋，因为这意味着巨大的创新空间和尚未开发的市场潜力。 目前大众看到的许多硬件，尤其是边缘侧设备，大都是两三年前设计的。当时甚至没有人考虑过要在本地运行这些大语言模型，因此市场上现存的计算架构在处理即将爆发的 AI 工作负载时几乎完全没有经过优化。这种未被充分满足的计算需求，构成了巨大的市场空白（white space）。半导体行业预计在未来十年末将成长为万亿美元规模的市场，而 Arm 目前仅占约 40 亿美元的营收。凭借其技术不可或缺的底层地位，Arm 完全有能力在未来的万亿美元大蛋糕中攫取大得多的份额，这也是投资者对其前景极度看好的根本原因。 ```mermaid flowchart TD accTitle: AI 计算工作负载的软硬件演进 accDescr: 展示从边缘端未优化架构到云端协同、从训练转向海量推理的算力演进路径。 subgraph HardwareGap["现状：硬件架构滞后"] A["2-3 年前设计的边缘硬件"] --> B["完全未针对本地 LLM 优化"] B --> C["巨大的计算创新空白"] end subgraph ComputingShift["趋势：从训练走向推理"] D["数据中心集中训练<br/>(少数'老师')"] --> E["全终端海量推理<br/>(无数'学生')"] end subgraph ArmEcosystem["终端与云协同实现落地"] C & E --> F["端云混合计算与安全验证"] F --> G["Arm CPU 架构 + AI 指令集扩展 + 嵌入式 NPU"] end ``` ### [59:23](https://www.youtube.com/watch?v=Y8guMe665RE&t=3563s) “CPU 已死”的悖论与海量推理驱动的新格局 当被问及一个看似反直觉的问题——整个行业过去五到十年几乎完全沉迷于 GPU 与加速计算，甚至鼓吹将繁琐的串行计算踢出 CPU，而作为 CPU 巨头的 Arm 为什么依然胸有成竹时，Haas 给予了有力反击。部分投资人曾质疑“既然一切都在向 GPU 迁移，我们还需要 CPU 吗？”，Haas 直言这种观点极其荒谬且片面，就像在说“我把汽车从 V6 发动机升级到了 V8 发动机，难道我以后开车就不需要方向盘和轮胎了吗？”。在计算系统的基本架构中，没有 CPU 负责掌控系统的核心运行，任何加速器都无法独立运作。 加速计算和 GPU 的崛起对 Arm 而言其实是巨大的重大利好： - **数据中心基石**：以英伟达的 Grace Blackwell 架构为例，强大的 AI 集群首先需要性能顶尖的 CPU 来提供支撑并与 GPU 密切协同。 - **“老师”与“学生”的比喻**：如果把模型训练比作“老师”，那么模型推理就是“学生”。世界上学生的数量远远多于老师，因此推理工作负载的体量将数千倍于训练工作负载。 - **边缘设备的功耗红线**：推理最终必须跑在最贴近用户的微型终端上，例如 AR 头显或可穿戴设备。正如 Haas 指出的那样，“你不可能把一块 100 瓦的 GPU 顶在头上”，极低功耗限制必然依赖 CPU 的高效计算。 - **架构扩展与专用卸载**：Arm 既可以在 CPU 中融入面向 AI 加速的指令扩展（体现 RISC 的模块化优势），也可以配合像 Ethos 这样拥有 4 TOPS、8 TOPS 算力的自研小算力 NPU 进行工作卸载。 - **端云协同与隐私安全**：未来的主流交互模型是本地处理与云端计算相结合。边缘设备需要承担本地安全验证与隐私隔离，防止敏感数据泄露，这完全是 Arm 的主战场。 ### [01:03:14](https://www.youtube.com/watch?v=Y8guMe665RE&t=3794s) 英伟达 400 亿美元收购案与反垄断监管的深层意义 黄仁勋和英伟达多年前对 Arm 的巨额竞购，恰恰是 Arm 核心地位的最佳证明。2020 年英伟达宣布计划以 400 亿美元收购 Arm 时，Arm 的年营收仅约 20 亿美元，而英伟达当时的年营收约为 250 亿美元、市值在 3500 亿至 4000 亿美元之间。黄仁勋绝不是为了在财务报表上增加几十亿美元的营收才发起收购，而是深刻洞察到了 Arm 掌控着整座半导体行业的命脉。 在 2016 年软银以 320 亿美元收购 Arm 之后，由于核心业务转型尚未完全显效，2020 年宣布英伟达以约 400 亿美元接手时，外界甚至充斥着“英伟达当了冤大头、出价过高买了一家非增长型公司”的冷嘲热讽。但如今对照双方当前的体量，当时的估值显得无比便宜。更重要的是，这起收购案迅速遭到了全球反垄断监管机构、行业客户以及生态伙伴的强烈反对。这种普遍的抵制情绪从反面印证了一点：一家能够引发全行业同仇敌忾并让所有人担心“被断供”的公司，其重要性远超出外界最初给予的评级。 在反垄断法律实践中，这起交易原本被很多人视为纵向并购（垂直整合）而非横向整合——当时 Arm 的主要收入来自智能手机芯片架构，而英伟达并不涉足智能手机业务，理论上并不会直接消除某一终端市场的单一直接竞争对手。然而，作为整个科技行业和人类创新不可或缺的底层通用“基建”，若让单一硬件巨头独占 Arm 的所有权，将直接剥夺其他所有客户平等获取底层技术的权利，这也是该并购案最终被各方合力叫停的关键原因。 ### [01:06:25](https://www.youtube.com/watch?v=Y8guMe665RE&t=3985s) 监管长远考量与英伟达押注 Arm 的曲折历程 在谈到监管机构对英伟达（NVIDIA）此前拟收购 Arm 案的审查时，对话揭示了一个反常的监管逻辑：监管机构通常更关注眼前的短期市场影响，但在针对那笔世纪收购案时，他们却极为罕见地盯住了遥远的未来——即未来某一天“可能会发生什么”，而非眼下短期内“预计会发生什么”。话题随即转向英伟达自身与 Arm 之间看似不可思议的渊源。英伟达最初作为一家深植于 x86 个人电脑生态的独立显卡公司起家，后来向数据中心发起颠覆性进军时，数据中心本质上同样是纯粹的 x86 地盘。英伟达究竟是在何时开始意识到，Arm 架构绝非仅仅是那种“插在 2G 功能机里保证塑料壳不被融化”的低功耗玩具？ 作为前英伟达资深高管，Rene Haas 回顾了英伟达与 Arm 深度绑定的早期岁月。英伟达很早就通过 Tegra 处理器明确做出了加速移动业务并全力押注 Arm 的战略转向。在这场跨界探索中，英伟达曾收购了为初代苹果 iPod 打造音频解码芯片的 PortalPlayer，并在当年为微软抗衡 iPod 的音乐播放器 Zune 提供片上系统（SoC）。英伟达在 Windows CE 和嵌入式设备上与 Arm 频繁互动，但真正促使英伟达彻底“加倍下注”Arm 的催化剂有两个：第一是全球智能手机浪潮的爆发式狂飙；第二则是 2009 年前后微软公开承诺开发基于 Arm 的 Windows 系统（Windows on Arm）。 - **PC 游戏积累的优势**：当时 Rene Haas 正掌管英伟达的全部笔记本业务并亲自统筹 Windows on Arm 项目。英伟达坚信凭借在 Windows 生态、DirectX 技术以及 PC 游戏图形栈领域数十年的技术沉淀，自身在 Windows on Arm 市场具有得天独厚的先发优势。 - **跨平台软件移植的红利**：尽管让数十年来专门面向 x86 编译的代码“一键重新编译并流畅运行在 Arm 上”近乎编译器工程上的奇迹，但移动端的空前繁荣极大推进了这一进程，微软旗下包括 Office 在内的核心应用如今均已在 iPad 和各类移动 Arm 设备上实现原生运行。 ```mermaid flowchart TD accTitle: NVIDIA's Strategic Evolution with Arm accDescr: Flowchart tracing Nvidia's shift from early handheld gadgets to modern AI data centers. subgraph EarlyPhase["早期探索 (2000s)"] direction TB A["收购 PortalPlayer (iPod 音频芯片)"] --> B["自研 Tegra SoC (微软 Zune 与嵌入式)"] end subgraph ExpansionPhase["战略下注 (2010s)"] direction TB C["移动终端与智能手机浪潮爆发"] --> D["全面承接 Windows on Arm (2009)"] D --> E["拓展至汽车与智能硬件<br/>(NVIDIA Drive & 机器人)"] end subgraph DataCenterPhase["全栈融合 (当前)"] direction TB F["加速计算全栈原生运行于 Arm 平台"] --> G["深度定制 CPU+GPU 架构<br/>(Grace Hopper & Grace Blackwell)"] end EarlyPhase --> ExpansionPhase ExpansionPhase --> DataCenterPhase ``` --- ### [01:09:08](https://www.youtube.com/watch?v=Y8guMe665RE&t=4148s) 从车载计算到数据中心：Grace 芯片的定制化自由 在 Rene Haas 离开英伟达之后，英伟达与 Arm 的捆绑进一步加深，Arm 顺理成章地演变为英伟达在车载自动驾驶平台（NVIDIA Drive）以及机器人软硬件体系中的默认基石。在当下的汽车智能化领域，无论是由英伟达、瑞萨电子（Renesas）还是高通（Qualcomm）打造的计算方案，绝大部分软件栈均已完全原生构建在 Arm 架构之上。汽车芯片的这种普遍性，也反向强化了 Arm 在汽车供应链中不可撼动的统治地位。 当汽车领域的算力引擎全面走向成熟，数据中心的算力爆发接踵而至。英伟达敏锐地察觉到，必须跳出传统 x86 通用芯片的条框限制，利用 Arm 极高的架构定制自由度来实现前所未有的工程突破。英伟达随后推出的 Grace Hopper 超级芯片以及最新一代 Grace Blackwell 架构，正是通过将自研 Arm CPU 与英伟达顶级 GPU 进行紧密耦合与极速互连，打造出了在传统标准化芯片体系下“绝不可能以其他方式实现”的系统级性能飞跃与创新空间。 --- ### [01:10:14](https://www.youtube.com/watch?v=Y8guMe665RE&t=4214s) 商业模式进化：从售卖单体 IP 到“计算子系统”（CSS） 对谈进入最后的关键战略话题：Arm 正在发生的商业模式演进——计算子系统（Compute Subsystems，简称 CSS）。Rene Haas 解释道，外界常把 Arm 简化为“一家做 CPU 的公司”，但 Arm 内部的产品线远比外界想象的庞大，不仅涵盖 GPU、面向边缘人工智能的 NPU，还包括构建现代服务器级芯片必不可少的底层互连管网——相干网状网络（Coherent Mesh Network，简称 CMN）。构建一颗搭载 128 个 CPU 核心的现代 SoC，核心之间的通信互连以及它们与内存接口的高速吞吐，往往涉及极其复杂的底层布线工程。 为了向技术投资者清晰阐明这一概念，Rene Haas 引用了他在 Arm 首次公开募股（IPO）路演中屡试不爽的“乐高积木”比喻： > “想想看，这些组件就像一块块独立的乐高积木。对于那些技术实力极强的顶级客户，你可以直接把这些散装乐高卖给他们，他们自己就能动手拼出一座精美绝伦的自由女神像；但你也可以对客户说：‘看好了，只要按照我们预先设计好的特定形态将它们拼接在一起，你得到一座自由女神像的速度将比你自己摸索搭建快上成百上千倍。’这就是计算子系统。” 在计算子系统业务中，Arm 不再只是向客户投递分散的 CPU 和总线 IP，而是把 128 个 CPU 核心、相干网状网络、内存控制器与外围接口全部预先拼接整合，并在物理实现前完成严密的系统级全功能验证，确保“插上就能直接点亮运行”。 - **研发周期的质的飞跃**：这套软硬件验证方案能够为芯片设计公司直接省去 3 个月、6 个月甚至 9 个月的漫长工程验证周期，让终端芯片以极快的速度抢占市场窗口期。 - **与晶圆代工厂深度标定**：Arm 进一步加深与台积电（TSMC）、三星电子（Samsung）以及英特尔晶圆代工（Intel Foundry）的协同，保证在特定先进制程下以特定参数实现明确的性能指标，例如直接向客户提供保证可达成 4.4 GHz 主频的经过硬核标定的计算子系统方案，本质上提供了一套逼近完整芯片实现的“虚拟芯片组”（virtual chipset）。 --- ### [01:12:49](https://www.youtube.com/watch?v=Y8guMe665RE&t=4369s) 封装价值链与未来的无限悬念 针对主持人将这一举措形容为“芯片设计层面的打包捆绑（bundling）”，Rene Haas 更倾向于称其为一种端到端的高效“封装（packaging）”。他强调，对于开发智能手机芯片的客户而言，他们的核心商业价值在于手机拍照的图像信号处理器（ISP）或端侧 AI 算力；对于云端数据中心客户而言，其核心竞争力在于专有加速器架构或模拟 I/O。至于把上百颗通用 CPU 核心与内存接口严丝合缝地连在一起，这纯粹属于吃力且重复造轮子的底层基建，本身并不能为终端客户提供差异化的竞争价值。由对核心架构理解最透彻的 Arm 出面统筹计算子系统并定义芯片在先进制程下的极限性能，是全行业效率最高的分工方式。 面对这套已经无限接近完整成品芯片的参考设计，主持人最后抛出了全场最具冲击力的终极疑问：“既然你们现在已经掌握了制造非凡芯片的全部完整蓝图，未来有一天，我们是否会看到 Arm 自己拿起电话打给台积电说：‘嘿，直接替我们量产几百万颗实体芯片出来’？” 对此，Rene Haas 报以充满玄机与克制的微笑，给出了意味深长的回答：“关于这件事，我今天没有任何可以透露的（Nothing I can say about that today）。”双方在愉快的道别与悬念中结束了这场涵盖移动互联、PC 变革、汽车电子乃至 AI 数据中心霸权的技术商业长谈。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=Y8guMe665RE&t=0s) 每一台现代电子设备的数字大脑 — Rene Haas > [15:23](https://www.youtube.com/watch?v=Y8guMe665RE&t=923s) 一款 CPU 架构的真正价值，完全取决于在其之上编写的软件生态，以及这些软件究竟能存活多久 — Rene Haas > [19:16](https://www.youtube.com/watch?v=Y8guMe665RE&t=1156s) 软件永不死 — Jensen Huang > [27:36](https://www.youtube.com/watch?v=Y8guMe665RE&t=1656s) 当时的应用程序就是电话本身 — 主持人 > [38:41](https://www.youtube.com/watch?v=Y8guMe665RE&t=2321s) 我们负责供水。任何人想要进入智能手机市场，除非通过 ARM，否则根本无路可走 — Rene Haas > [40:35](https://www.youtube.com/watch?v=Y8guMe665RE&t=2435s) 客户绝不允许添加独占的自定义指令来破坏指令集架构（ISA）的兼容性 — Rene Haas > [44:11](https://www.youtube.com/watch?v=Y8guMe665RE&t=2651s) 名为硅谷（Silicon Valley），但半导体（Silicon）初创企业却根本融不到风投资金 — Rene Haas > [59:23](https://www.youtube.com/watch?v=Y8guMe665RE&t=3563s) 我把汽车从 V6 发动机升级到了 V8 发动机，难道我以后开车就不需要方向盘和轮胎了吗？ — Rene Haas > [59:23](https://www.youtube.com/watch?v=Y8guMe665RE&t=3563s) 你不可能把一块 100 瓦的 GPU 顶在头上 — Rene Haas > [01:10:14](https://www.youtube.com/watch?v=Y8guMe665RE&t=4214s) 想想看，这些组件就像一块块独立的乐高积木。对于那些技术实力极强的顶级客户，你可以直接把这些散装乐高卖给他们，他们自己就能动手拼出一座精美绝伦的自由女神像；但你也可以对客户说：‘看好了，只要按照我们预先设计好的特定形态将它们拼接在一起，你得到一座自由女神像的速度将比你自己摸索搭建快上成百上千倍。’这就是计算子系统 — Rene Haas > [01:12:49](https://www.youtube.com/watch?v=Y8guMe665RE&t=4369s) 关于这件事，我今天没有任何可以透露的 — Rene Haas ## Mentioned Arm Holdings, Rene Haas, Jensen Huang, NVIDIA, SoftBank, Apple, Apple Newton, Intel, x86, CISC, RISC, David Patterson, Motorola 68000, MIPS, IBM PC, Gary Kildall, Microsoft, DOS, PowerPC, MacBook Pro, M1, Vanta, Jeff Bezos, Robin Saxby, Tudor Brown, Texas Instruments, Nokia, TSMC, Symbian, iPod, Intel Atom, Andy Rubin, Android, Samsung, Qualcomm, MediaTek, Tesla, Amazon, AWS Graviton, P.A. Semi, Cadence, Synopsys, Tensilica, ARC, DEC Alpha, SPARC, AMD, GlobalFoundries, Microsoft Azure, Google Cloud Platform, Ampere, Grace Hopper, Grace Blackwell, OpenAI, Marc Benioff, PortalPlayer, Zune, Windows on Arm, Renesas, Compute Subsystems (CSS) ## Video description (Episode originally released Dec 02, 2024) ARM is an incredibly unlikely story. They were founded in Cambridge, England in 1990 to design a new chip architecture just for low-power devices (like the Apple Newton!), leaving the “serious computing” on desktop and servers to Intel’s x86. Now, nearly three decades later, ARM is the dominant architecture in all of computing today. ARM is in your phone, your car, data centers, the most advanced AI chips… there are hundreds (or thousands!) of ARM chips you encounter in your everyday life. In this episode, ARM Holdings CEO Rene Haas joins us to tell the story of how ARM become so dominant, weaving through the through the iPod, smartphone, and AI eras. Plus, their wild corporate story of going public, getting bought by SoftBank, going public again, and nearly being acquired by NVIDIA! *More Acquired:* - Get email updates https://www.acquired.fm/email and vote on future episodes! - Join the Slack http://acquired.fm/slack - Check out the latest swag in the ACQ Merch Store https://www.acquired.fm/store! _Note: Acquired hosts and guests may hold assets discussed in this episode. This podcast is not investment advice, and is intended for informational and entertainment purposes only. You should do your own research and make your own independent decisions when considering any financial transactions._ [Source](https://www.youtube.com/watch?v=Y8guMe665RE)

[0:00] Rene Haas: 每一台现代电子设备的数字大脑 [15:23] Rene Haas: 一款 CPU 架构的真正价值，完全取决于在其之上编写的软件生态，以及这些软件究竟能存活多久 [19:16] Jensen Huang: 软件永不死 [27:36] 主持人: 当时的应用程序就是电话本身 [38:41] Rene Haas: 我们负责供水。任何人想要进入智能手机市场，除非通过 ARM，否则根本无路可走 [40:35] Rene Haas: 客户绝不允许添加独占的自定义指令来破坏指令集架构（ISA）的兼容性 [44:11] Rene Haas: 名为硅谷（Silicon Valley），但半导体（Silicon）初创企业却根本融不到风投资金 [59:23] Rene Haas: 我把汽车从 V6 发动机升级到了 V8 发动机，难道我以后开车就不需要方向盘和轮胎了吗？ [59:23] Rene Haas: 你不可能把一块 100 瓦的 GPU 顶在头上 [70:14] Rene Haas: 想想看，这些组件就像一块块独立的乐高积木。对于那些技术实力极强的顶级客户，你可以直接把这些散装乐高卖给他们，他们自己就能动手拼出一座精美绝伦的自由女神像；但你也可以对客户说：‘看好了，只要按照我们预先设计好的特定形态将它们拼接在一起，你得到一座自由女神像的速度将比你自己摸索搭建快上成百上千倍。’这就是计算子系统 [72:49] Rene Haas: 关于这件事，我今天没有任何可以透露的

[2:33] Rene Haas: I can pretty much assure you that it's Arm. [5:20] Rene Haas: a crazy number [6:46] Rene Haas: before anything had a right to be a PDA. [6:46] Rene Haas: the ultimate productivity gain [1:19] Rene Haas: the digital brain of every modern electronic device. [9:24] Rene Haas: go three steps forward, two steps to your left, diagonally two steps, right three steps. [12:28] Rene Haas: move one step forward, one step backward [15:52] Rene Haas: a CPU is only as good as the software that's written on it. And how long that software survives. [20:00] Jensen Huang: software never dies. [22:34] Jeff Bezos: what actually makes your beer taste better [27:42] Ben or Dave: the application was the phone [27:42] neither really so low power nor really so low cost. [27:42] compiler miracle [27:42] Rene Haas: We're never going to break the ISA. We're not going to allow people to add custom instructions, because once you do that, you break software compatibility, which is one of the superpowers of ARM. [27:42] Rene Haas: when you have a high-margin business model, it's very hard to switch to a low-margin business model. [27:42] Rene Haas: if training is the teacher, inference is the student, and there are far more students than teachers in the universe. [27:42] Rene Haas: You're not going to run a 100-watt GPU on your head, I'm sorry, it's not going to happen. [70:14] Rene Haas: packaging instead of bundling [70:14] Rene Haas: Nothing I can say about that today.

[0:00] Hello, Acquired listeners. Today, we [0:02] have with us Rene Haas, the CEO of Arm [0:06] Holdings. Arm is the company that [0:08] develops the instruction set [0:10] architecture and many of the designs [0:12] underpinning CPUs all over your life [0:14] today, from our phones to our cars. And [0:17] Dave and I actually did an episode way [0:19] back in 2019 on the history of the [0:21] company, which had a fascinating start [0:23] out of Cambridge University. The company [0:25] was publicly traded, then taken private [0:27] in 2016 by SoftBank, then last year went [0:30] public again and is now valued at around [0:32] $150 billion. [0:34] Rene has quite the career himself in [0:36] semiconductors. He's been at Arm for the [0:38] last 11 years and before that was a VP [0:40] at Nvidia reporting to Jensen. Rene, [0:43] welcome to Acquired. Thank you very [0:45] much. Pleasure to be here with you both. [0:47] Delighted to have you. Pleasure is all [0:49] ours. [0:50] Well, I thought a fun way to start us [0:52] off, since there's a lot of people [0:53] listening to this that are going to see [0:56] Arm Holdings and say, "I know exactly [0:58] what that is. I know about the strategic [1:00] shift that they have going on. I've been [1:01] following every earnings call since they [1:04] went public." And there's people that [1:05] are going to say, "Arm, what is that?" [1:08] So, maybe to level set everyone on how [1:10] important Arm is in the world, what are [1:12] the types of devices where the Arm [1:15] instruction set architecture and Arm [1:17] designs are used? [1:19] So, Arm does CPUs. And what is a CPU? [1:22] The CPU is the digital brain of every [1:25] modern electronic device. That is your [1:28] television set, your thermostat, your [1:31] car. You know, we were having a chat [1:32] maybe about talking about a day in the [1:34] life inside Arm, and I can walk through [1:36] some of the all the Arm devices inside [1:37] my home. But simplest way to think about [1:39] it is we do CPUs, and that CPU is the [1:43] digital brain of every modern electronic [1:44] device. [1:46] And so, what is your relationship then [1:47] with, you know, in my head, Apple makes [1:50] the CPU in my phone. It's the A18 or in [1:53] my Mac the M4. What do you mean Arm does [1:56] CPUs? Yeah, so drilling down one level [1:59] deeper, we do the design, the ISA, which [2:03] is the instruction set architecture, [2:05] and we license that as either an [2:07] instruction set architecture to a [2:09] partner they can develop their own CPU [2:12] based on Arm. That's what Apple does. [2:14] Or we design and build our own CPUs and [2:17] license those CPUs to companies like [2:19] Samsung, MediaTek, Tesla, Qualcomm, [2:23] Amazon, etc. So, we deliver it in two [2:26] different ways, but those CPUs that you [2:28] mentioned inside your iPhone and inside [2:30] your MacBook, those are all Arm based. [2:33] And so, where we were going with this, I [2:35] mean, today, yeah, if you were to [2:36] imagine your house, my house, Ben's [2:38] house, any house, [2:40] how many devices have an Arm chip in [2:43] them? And is that a different question [2:45] than how many Arm chips are floating [2:47] around my house? It's a hard question to [2:49] answer in terms of just how many Arm [2:52] chips are in my house or how many Arm [2:54] chips get delivered in terms of a [2:56] typical application space because it [2:58] really varies. [3:00] But again, let's let's go back to first [3:01] principles, right? [3:02] Arm designs the CPU, and that is the [3:05] digital brain of every device, which [3:07] means it runs all the the complex [3:09] software that either runs the dashboard [3:12] or it runs the operating system or it [3:14] runs an application. So, [3:17] I was thinking about the question, and [3:18] I'm going to drop a bunch of brand names [3:19] here, but let's just kind of walk [3:21] through I pull my Audi into the garage. [3:23] That Audi has Arm processors. Those Arm [3:26] processors are what you see running the [3:29] display and that digital dashboard. [3:32] They're also helping with some of the [3:33] driver assist, and they're probably in [3:36] the power locks, power windows, etc., [3:38] etc. [3:39] I have a a a Nest doorbell camera. [3:42] That's Arm, and that's Arm that [3:44] basically runs the camera, interfaces [3:46] with the doorbell, etc., etc. [3:48] Walking by the LG refrigerator or Wolf [3:51] stove, I can assure you both of those [3:53] have Arm inside, too. They're probably [3:55] running the displays. They're probably [3:57] running the temperatures. In the stove, [3:58] they're definitely running the display. [4:00] They're running everything in terms of [4:01] the oven. [4:02] Turn on the television set, which is a a [4:04] Samsung. That Samsung digital TV, it's [4:08] actually running an operating system. [4:09] So, when you run all those apps and [4:11] everything that you see kind of shows up [4:12] on there, that's a version of Android. [4:15] That's all Arm. [4:16] Let's say I want to go downstairs and do [4:18] some gaming. My PS5 has Arm inside, [4:22] most likely running some of the display [4:23] controllers and and running some of the [4:25] stuff with it with a game controller. [4:27] And if I want to flip through on my [4:29] Pixel phone, that is Arm inside running [4:32] Android, and I've got my iPad next to [4:34] me. That's all Arm. So, you can imagine [4:36] just about everything that you [4:38] interacted with that does something, [4:40] that either runs an application, [4:42] recognizes your face, [4:45] gives you some display information, Arm [4:47] did all of that. I think it's probably [4:49] true there are hundreds of Arm chips or [4:53] devices with maybe Arm chips. How would [4:55] you describe There are hundreds of [4:56] instances of Arm around my house? [4:58] Probably hundreds. Yeah. Yeah, probably [5:00] hundreds. I mean, if you think about the [5:02] more your home is connected, all those [5:04] connected things have Arm inside. It's [5:08] hard to avoid it because you'd almost [5:09] have to go back to these old mechanical [5:12] type of controls on on machines that [5:15] actually don't have something that's [5:16] digital because if it's digital, [5:18] I can pretty much assure you that it's [5:19] Arm. [5:20] It's pretty wild. So, one stat that I [5:22] pulled just from your last quarter's [5:24] financial presentation is that in FY, [5:28] this is estimated 2024, there's almost [5:31] 29 billion Arm chips shipped. And that [5:35] is for every human on Earth, there is [5:38] four Arm based chips shipped in the last [5:40] 12 months. It's a crazy number, right? [5:43] When you think about the laptop market, [5:46] which is a big market, right? Everyone [5:47] wants to ship into laptops, big market, [5:50] etc., etc. [5:51] That's 200 million units plus or minus a [5:53] year, [5:55] which is a a fraction of that 29 [5:58] billion. [5:58] Right. A very small fraction. Very small [6:01] fraction, right? So, you look at this [6:02] and say, "Well, how is that [6:03] how is that possible?" Because laptop [6:05] computers they seem to be, you know, [6:06] pretty ubiquitous. But just walk through [6:08] that example I just gave you in terms of [6:09] those eight or nine examples in inside [6:11] the house, [6:13] and then you start to see, well, how do [6:14] you avoid it? Arm is in aircraft. You go [6:16] to an airport and you check in for your [6:18] flight and you look up at those displays [6:20] that are listening to gate information [6:22] and the flight information. [6:24] That's all Arm powered. That's running [6:26] that stuff in in the background. So, [6:27] it's everywhere. And at this point, [6:29] counterintuitively, [6:31] it's also in all the cloud architecture [6:33] that are running the web services that [6:35] all of these devices are communicating [6:37] with. And that is, as we'll get into [6:39] later in the episode, kind of a [6:40] narrative violation from the way that [6:42] the world thought about Arm a decade ago [6:44] versus what is true today. That's right. [6:46] I mean, the identity of the company, we [6:49] grew up, as you mentioned in the [6:50] opening, 30-plus years ago out of [6:52] Cambridge, and and the company's [6:54] original product that we were designed [6:56] into was the Apple Newton. And for those [7:00] who may or may not remember, [7:02] that was a PDA before anything had a [7:04] right to be a PDA, before before there [7:06] was the internet, before you had voice [7:07] recognition, before you had fingerprint [7:09] recognition. [7:10] But the chip that was designed that was [7:12] based upon Arm inside had two important [7:14] characteristics. It had to be running [7:16] off a battery, so as a result, it had to [7:18] be defined to be low power. [7:20] And secondly, performance and cost was [7:22] really important. And back in the day, [7:24] they used to build chips in two [7:26] different ways. They had plastic [7:27] packages, [7:28] which was pretty rare, and ceramic [7:30] packages, which were much better in [7:32] terms of heat dissipation, but were [7:34] costly and not that great in terms of [7:36] thermals. [7:37] So, one of the directives in terms of [7:39] the original design was let's get it in [7:41] a plastic package. So, as a result, from [7:43] the very early days, the early Arm [7:46] processor, the Arm 1 that was defined, [7:48] was basically to run off a off a [7:50] battery. [7:52] Yeah, which at the time didn't feel as [7:55] critical to the world since all [7:57] computers were basically plugged in all [7:59] the time or most of the computing that [8:01] people did was at computers that were [8:02] plugged in all the time. And now [8:03] obviously that's very different. [8:05] Absolutely. If I think back in time to [8:08] the first time that one could take one [8:10] of those large satellite phones and walk [8:13] around with them for 20 minutes without [8:14] having to plug them in, [8:16] it just seemed like magic, you know, [8:17] back in the day. If you could get 30 to [8:20] 40 minutes of battery life off of [8:21] anything that was doing something [8:23] sophisticated, it was considered to be [8:25] just a a complete game changer because [8:28] mobility was simply not something that [8:31] was very, very ubiquitous back in the [8:33] early days. If I think of stories around [8:34] this, one of the jobs I had in my career [8:36] was a field applications engineer. [8:38] And I'm going to date myself here, but [8:39] we used to [8:41] call in to the offices for messages. In [8:44] fact, we would be driving from account [8:46] to account. We'd find ourselves get to a [8:48] payphone. Once we got to that payphone, [8:50] we could then dial the office. The [8:52] office would list a whole bunch of [8:54] messages. The detail of those messages [8:56] were something like, "Call me back" or [8:58] "I'm not exactly sure what you asked" or [9:01] "I'm busy." [9:02] When we suddenly had a phone in our cars [9:05] that would allow us to do all these [9:07] things remotely, we thought, "Oh my [9:09] gosh, [9:10] this is the ultimate productivity gain [9:12] relative to uh [9:14] what seemed like Western Union looking [9:15] back in terms of making these phone [9:17] calls back and forth." But true story, [9:18] that's exactly how I my first field [9:20] applications job, that's how I used to [9:22] correspond with the home office. [9:24] And so, the arc that we're going to keep [9:27] calling back to over the course of this [9:28] episode, as you mentioned, the original [9:30] Arm processor was designed [9:33] with extreme low heat requirements in [9:36] mind, kind of low power requirements in [9:39] order to not quickly drain a battery in [9:40] a very inefficient world with way fewer [9:43] advances in lithium ion versus what we [9:44] have today. [9:46] And so, you sort of think, "Well, this [9:47] crappy processor architecture that's [9:49] extremely limited in its capabilities [9:52] will never be the dominant architecture [9:54] used in all of the most sophisticated [9:56] and advanced computing applications in [9:57] the world and yet it is. So, over the [10:00] next hour we're going to kind of wander [10:01] through how did we get here? But in [10:03] Acquired fashion I want to go way back [10:05] to the beginning and introduce this idea [10:08] of the reduced instruction set computer. [10:11] And I wanted to sort of turn it over to [10:12] you. You are wildly overqualified to do [10:15] this as the CEO of ARM, but maybe play [10:17] computer science professor or computer [10:18] science history professor with us for a [10:20] little bit. What was the development of [10:22] the RISC versus the CISC, the complex [10:25] instruction set computer like? [10:28] The concepts of of RISC and I think they [10:30] were originally, you know, kind of [10:31] conceived by professors at University of [10:34] California, Berkeley, David Patterson. [10:36] And the whole notion around RISC versus [10:38] CISC was [10:39] these original processors that were [10:41] invented and we're going to go way back [10:43] in time to [10:44] processor architectures, you know, such [10:46] as the x86 or the 68000s from Motorola [10:50] were deemed as [10:51] CISC processors which stand for you kind [10:53] of complex instruction set computers, [10:56] which basically meant that they had lots [10:58] and lots and lots of instructions that [11:00] they had to carry forward because the [11:03] software that was written for them prior [11:05] relied on them. [11:07] So, they were carrying a lot of baggage [11:09] to do these very very complicated [11:11] instructions which burned a lot of power [11:13] because the simplest way to think about [11:14] a [11:15] a CISC implementation is [11:18] an instruction by its definition because [11:19] it's complex means it has to run [11:21] multiple operations from a clock [11:23] standpoint to execute the instruction, [11:25] which means those transistors are [11:27] running more than they probably should [11:28] and you're burning up a bunch of power. [11:30] By example, at any given clock cycle, I [11:32] need to allow for the possibility of [11:34] doing something complicated like in this [11:37] instruction or in this operation I'm [11:39] going to go fetch something from memory [11:41] and load it into a register so that it [11:44] can be added and I can return the answer [11:46] all within the same clock cycle. So, you [11:48] sort of have like extra bandwidth [11:50] everywhere to accommodate doing [11:52] complicated things in one simple [11:54] assembly language, you know, line. Yeah, [11:57] that's a good way to describe it. [11:58] Another or another way to think of a [11:59] complex instruction set, a complex [12:01] instruction is [12:03] go three steps forward, two steps to [12:05] your left, diagonally two steps, right [12:08] three steps. [12:09] Now, if you can find an operation that [12:11] benefits from that specialized activity, [12:13] that's pretty good. [12:14] But not a lot of programs can. But once [12:17] a program has been written and relies on [12:19] that instruction [12:21] then by definition the architecture has [12:23] to carry that forward. So, you've got [12:25] all this kind of heavyweight stuff [12:27] that's involved. [12:28] So, the concepts were RISC were really [12:30] around simple movements. [12:32] Move one step forward, one step [12:34] backward, one step to your left, one [12:35] step to your right. I'm oversimplifying, [12:37] of course. But these are things like [12:40] add, subtract, et cetera, et cetera. And [12:41] the idea there being that if you have a [12:43] simpler set of instructions [12:46] that can be combined in such a way to be [12:48] much more efficient [12:50] then this is the concept of RISC versus [12:52] CISC, which gosh, I'm thinking now [12:53] probably back to the [12:56] to the 1980s where MIPS was invented and [12:59] and things of that nature that were the [13:00] original kind of RISC processors. And [13:02] this was all around reducing instruction [13:05] set complexity, but interestingly [13:07] enough, that was back in the day when a [13:08] lot of programs were written on [13:10] mainframes [13:11] or mini computers [13:13] with previous architectures. [13:15] And it was really interesting if I look [13:17] back in that time, you had a lot of [13:18] energy being spent on developing new [13:22] processor techniques when actually you [13:25] didn't have nearly the mountain of [13:27] software that you have today. [13:29] But yes, if you go back in time RISC was [13:32] seen as a much more efficient way to do [13:35] computing and one of the benefits you [13:37] had of that was just not only lower [13:40] power systems [13:41] but also going back in time, one of the [13:42] most more expensive things was actually [13:44] the memory associated to run all these [13:46] programs. If you could fit the program [13:48] in a smaller memory footprint which [13:50] again and it with a RISC machine you can [13:52] do that, there were some benefit to [13:54] that. So, yeah, that was [13:55] that was way way way back, I would say. [13:58] Probably '70s and '80s time frame. It's [14:00] so interesting and you can totally see [14:01] why CISC was [14:03] conceived of first or at least in the [14:04] early days believed to be better. It's [14:07] this really incredibly powerful system [14:10] where you can any given you know, [14:13] instruction can actually do a lot of [14:15] cool stuff behind the scenes. And when [14:17] you sort of juxtapose it to RISC which [14:19] in its early days had very few [14:21] instructions, a simple operation would [14:23] be load, you know, hey, go grab this [14:25] thing from memory, put it in a register. [14:28] Oh, I can't do anything else. That is [14:30] all we've allowed for in this [14:32] instruction. Just load. That's it. Oops. [14:34] Add. Oops. That's also it. Especially if [14:37] these larger programs were compiled and [14:39] then made use of those big instructions [14:41] from an assembly standpoint because [14:43] the other thing that was happening was [14:44] you were you were having a change from [14:46] everything being done in in what was [14:48] called low-level programming, assembly [14:49] language [14:50] to higher-level programming models such [14:52] as Fortran, Pascal, and then C and C++. [14:56] And when you're programming at the [14:57] higher-level languages you have these [15:00] compilers and then the what does the [15:01] compiler do? The compiler takes that [15:03] high-level language and tries to put it [15:06] into lower-level language, which are [15:08] what these instructions are. [15:09] So, the compilers [15:11] end up making use of these heavy [15:13] instructions. And as a result you've [15:16] just got heavier and more inefficient [15:18] code. And again, one of the things that [15:20] people were trying to do back in the day [15:21] was get to smaller memory footprints. [15:23] Everything you're describing [15:25] of the old CISC world sounds, like you [15:28] said, fits perfectly with the mainframe [15:30] and the mini computer era, you know, big [15:32] iron, big architecture. Nobody's worried [15:35] about power requirements. Complexity is [15:37] fine. IBM designs the whole thing. [15:40] You would think naively that the shift [15:42] to the PC era would have created the [15:45] right opening for RISC. [15:48] But actually [15:50] CISC continued through the PC era. So, [15:52] like what happened? Was RISC developed [15:53] just like slightly too late? Did ARM not [15:55] exist yet? So, let's continue down this [15:58] this history lesson here for for a [15:59] moment here. [16:00] One of the most amazing things that took [16:02] place with the IBM computer IBM PC back [16:05] in the day was [16:06] IBM, which was the world's leader in [16:09] computing, if you go back in time, [16:10] right? If you think about IBM 360 [16:13] and the IBM mainframes and the IBM minis [16:16] IBM was a one-stop shop, right? IBM did [16:19] the software, IBM did the service, IBM [16:21] did the hardware, IBM was was [16:22] everything. So, now in 1981 IBM decides, [16:26] I'm going to enter the PC market. [16:29] And they were behind. If you go back to [16:30] the late '70s, early 1980s [16:33] Apple has invented the first, you know, [16:36] end quote, home computer based on on the [16:38] Motorola architecture. Then back in the [16:40] day you had lots of [16:42] I wouldn't call them toy computers, but [16:44] things like TRS-80 and and Commodore and [16:47] and they all had these smaller weird [16:49] little processors into it. [16:51] So, the irony of the whole IBM story was [16:53] IBM, the behemoth of computing [16:56] decides that we're going to now enter [16:57] into building computers for the home. [17:00] And what does IBM decide to do? IBM [17:01] decides to not in-house the processor [17:05] nor do they decide to in-house the [17:07] operating system. [17:08] They decide that they're going to make [17:10] this platform, end quote, open. [17:13] They need an operating system, DOS. So, [17:15] something they can run off the disk and [17:17] they started talking to a company [17:20] that was actually not Microsoft. [17:21] Seattle Computer Products. The classic [17:23] CPM 80. They were talking to Gary [17:25] Kildall and and his company about doing [17:27] that, but they chose Microsoft. [17:29] And they were also looking at Motorola, [17:31] which was kind of considered that the [17:33] kingpin at the time to do the processor. [17:36] But for various different reasons they [17:37] decide on Intel and 8086. [17:40] The IBM PC is born. [17:42] And the irony of it is that there's [17:43] nothing about it that's very IBM-like [17:46] because it uses external memory. It uses [17:48] external hard drives. It uses an Intel [17:51] processor. [17:52] And it uses an operating system from [17:54] Microsoft. [17:55] So a little crazy if you kind of look [17:57] back in time that you'd look at it and [17:59] say, why would IBM actually, and this is [18:01] what took off with the birth of all the [18:03] the clones because you could build a [18:05] clone [18:06] of that system because if you bought a [18:07] processor from Intel and you bought a [18:09] hard drive from [18:11] Conner or Maxtor or Seagate and you [18:13] bought a monitor from [18:15] one of the third parties in Taiwan and [18:16] you got a license to DOS [18:19] you're in business and off you go. So [18:22] to your question though in terms of, [18:23] okay, so why why didn't somebody do [18:25] something on RISC? [18:27] Therein lies the magic of software [18:29] compatibility [18:30] and software legacy because all these [18:32] early programs, it was stuff like Lotus [18:35] 1-2-3 [18:36] were now written [18:38] to run on the x86 processor and [18:41] optimized on that. So, what happened was [18:44] over the 1980s as the IBM PC compatible [18:47] market started to take off you had all [18:50] this software that was written for that [18:52] platform. [18:53] And the dirty little secret about CPU [18:56] architectures and and there's been lots [18:58] of them over the days, whether it's [19:00] again, back to SPARC or MIPS or [19:04] ARC or Tensilica, Tensilica where I used [19:06] to work, 29000, 68000, DEC Alpha [19:10] a CPU is only as good as the software [19:12] that's written on it. And how long that [19:15] software survives. [19:16] So, the IBM PC and its clones ultimately [19:20] built in by companies by like Compaq and [19:22] by Dell and by Gateway and all these [19:24] other companies that have long gone, AST [19:26] Research, if you remember those guys. [19:29] This is what created the birth of not [19:33] only the IBM PC platform, obviously, but [19:35] the Intel x86 architecture. And that's [19:39] why as a default CISC, end quote, [19:43] because that's what x86 was [19:45] is the de facto. It really wasn't a, oh, [19:47] RISC was better cuz it probably was, but [19:50] kind of didn't matter. Once IBM selected [19:52] 8086 [19:54] and DOS was optimized for that and then [19:56] subsequently Windows, [19:58] you know, off it went. Now, the one [20:00] company was quite interesting that has [20:02] probably made the most pivots in this [20:03] area was was Apple, right? Because Apple [20:05] was, you know, originally 68,000 based [20:08] It's a Motorola chip. Yep. And then they [20:10] created a consortium [20:12] for the PowerPC. PowerPC, right? With [20:14] with IBM and Motorola. [20:16] with IBM. With IBM, yeah, exactly. Which [20:19] was kind of a RISC CISC hybridy thing. [20:21] That's very 90s Apple to have something [20:23] that is like neither RISC nor CISC, but [20:25] entirely reinvented and proprietary. [20:28] Yeah, and that's what Power was. You [20:29] know, that was a big switching cost. But [20:32] I think the other thing that's [20:32] interesting about that is it was a large [20:34] switching cost because the amount of [20:35] software work that was there, [20:37] but not nearly the amount of software [20:40] that exists today. I was having a [20:41] discussion on the on a podcast that I [20:43] did with Jensen. Yeah, you guys just [20:46] launched your own podcast, right? [20:47] We did. We did. And Jensen made a [20:48] comment on the podcast that software [20:51] never dies. [20:52] That continues to be a very true theme [20:55] relative to [20:56] just the amount of heavy lifting [20:59] required to switch an architecture. But [21:01] yeah, a long-winded answer to your story [21:03] if you kind of go back in time. Why did [21:04] CISC kind of make it? It was the IBM PC. [21:07] And once that took off, that's been a [21:08] very sticky platform. It's so funny. [21:10] Yeah, cuz RISC was there and arguably [21:12] would have been better for PCs of like, [21:14] "Hey, new paradigm, a lot of new [21:15] software going to get written." But it [21:17] was that decision to go with x86 that [21:20] locked CISC in for the PC era. [21:22] And in CPUs, and I would argue for any [21:25] any programmable architecture, to get to [21:28] something that drives a major switching [21:30] cost, you need a a fairly large paradigm [21:32] shift in terms of benefits on power or [21:35] benefits on on cost. You know, people [21:37] will talk about you need the 10x [21:39] advantage to make the switch. [21:41] I'm not sure it's 10x, but it's not 15%. [21:45] It's got to be something that's quite [21:46] material [21:47] that's going to change in terms of lift [21:50] and/or it has to drive us a level of [21:53] innovation that could not be done when [21:55] you're starting out, which kind of goes [21:56] all the way back to the Newton. [21:58] There was no way that an x86 could have [22:01] been an option. You simply could not [22:02] build the product. You either have to [22:04] start in a space where something is very [22:06] very new and you need some very very [22:09] unique computing paradigm and/or you've [22:11] got to drive some different level of [22:12] innovation. And to quantify the if it's [22:15] not 10x, what is it? I bet if you just [22:17] go look at the Geekbench scores from [22:20] whatever Apple's latest and greatest [22:21] Intel-based MacBook Pros were before [22:23] switching to the M1, that's probably the [22:25] exact quantification of how much better [22:27] does something need to be to stay in an [22:29] existing paradigm and switch from one [22:31] horse to another. [22:32] That's about right. Yep. Huh. We want to [22:35] thank our long-time friend of the show, [22:38] Vanta, the leading trust management [22:40] platform. Vanta, of course, automates [22:42] your security reviews and compliance [22:45] efforts. So, frameworks like SOC 2, ISO [22:47] 27001, GDPR, and HIPAA compliance and [22:50] monitoring. Vanta takes care of these [22:52] otherwise incredibly time and resource [22:54] draining efforts for your organization [22:55] and makes them fast and simple. Yep, [22:58] Vanta is the perfect example of the [23:00] quote that we talk about all the time [23:01] here on Acquired, Jeff Bezos, his idea [23:03] that a company should only focus on what [23:06] actually makes your beer taste better, [23:08] i.e., spend your time and resources only [23:10] on what's actually going to move the [23:12] needle for your product and your [23:13] customers and outsource everything else [23:15] that doesn't. Every company needs [23:17] compliance and trust with their vendors [23:18] and customers. It plays a major role in [23:20] enabling because customers and partners [23:22] demand it, but yet it adds zero flavor [23:25] to your actual product. Vanta takes care [23:27] of all of it for you. No more [23:28] spreadsheets, no fragmented tools, no [23:30] manual reviews to cobble together your [23:32] security and compliance requirements. It [23:34] is one single software pane of glass [23:37] that connects to all of your services [23:38] via APIs and eliminates countless hours [23:41] of work for your organization. There are [23:43] now AI capabilities to make this even [23:45] more powerful and they even integrate [23:47] with over 300 external tools. Plus, they [23:50] let customers build private integrations [23:51] with their internal systems. And perhaps [23:53] most importantly, your security reviews [23:55] are now real-time instead of static, so [23:58] you can monitor and share with your [23:59] customers and partners to give them [24:00] added confidence. So, whether you're a [24:02] startup or a large enterprise and your [24:04] company is ready to automate compliance [24:05] and streamline security reviews like [24:07] Vanta's 7,000 customers around the globe [24:10] and go back to making your beer taste [24:11] better, head on over to [24:13] vanta.com/acquired and just tell them [24:15] that Ben and David sent you. And thanks [24:18] to friend of the show, Christina Vanta [24:20] CEO, all Acquired listeners get a $1,000 [24:23] of free credit. vanta.com/acquired. [24:27] Okay, so we've perfectly set the table [24:29] for CISC in the PC era, pretty locked [24:33] in, not going anywhere. ARM is founded. [24:36] It's using a RISC-based approach. What [24:38] is ARM doing for its first couple [24:40] decades in existence? What markets does [24:43] it serve? So, [24:45] go back to kind of the the invention of [24:47] ARM. And one of the unique things that [24:49] ARM drove also back in the day that I [24:51] think couldn't be done today, but [24:54] perfect time, perfect place, perfect [24:55] strategy, all of this is also luck and [24:57] timing. [24:59] All of those processors that I just [25:01] described to you, x86, [25:03] 68,000, AMD 29,000, the list and list [25:07] goes on. [25:08] They were all vertically integrated. And [25:10] believe it or not, a lot of people used [25:12] to spend a lot of time designing their [25:13] own microprocessors. [25:15] And ARM had an idea that [25:18] that's a lot of work, that's a lot of [25:20] effort, [25:22] there's not a lot of differentiation [25:24] that one microprocessor can have versus [25:26] another microprocessor. So, why don't we [25:28] come up with a business model that [25:30] rather than building our own [25:33] and trying to enter the market against [25:36] what is very very crowded, [25:38] I'm going to license it and I'm going to [25:41] make it available to companies rather [25:43] than developing their own, [25:46] just run on ARM. And I'm going to [25:48] license it. I'm not going to charge, no [25:50] pun intended, an arm and a leg for it. [25:53] I'm going to have a business model [25:54] that's going to require [25:56] an upfront licensing fee, which is [25:57] modest, and I'll take a royalty when you [26:00] ship in production. And the idea back [26:02] then was on a shared success model, [26:04] which I think again, back to the [26:06] founders, you know, back to people like [26:07] Robin Saxby and and Tudor Brown, [26:10] that was really a rather brilliant idea. [26:12] Because the notion was pay me an upfront [26:14] license fee, which is kind of a proxy [26:16] for R&D. In other words, you're not [26:18] going to spend the money on the [26:19] engineers anymore to [26:21] to do the development. The licensing fee [26:23] will be sort of a proxy for R&D, so it's [26:25] not it's not an exorbitant fee and more [26:28] importantly, it's not money you wouldn't [26:29] be spending anyway. So, by licensing the [26:31] technology, you're not going to need to [26:33] hire the engineers to develop the [26:35] product cuz I've already done that for [26:36] you. [26:37] And then on the back end, [26:39] if you ship a whole bunch of products, [26:42] which is good for you, [26:44] then pay me a percentage of it because [26:45] it's good for me, too. So, it's a shared [26:47] success model. You look back and say, [26:49] "Wow, brilliant. Of course, why wouldn't [26:51] everybody do that?" But back when ARM [26:53] started in the early 1990s, [26:55] one of the things that was really not [26:56] there yet [26:58] was all of the tools and methodologies [27:01] and flows needed in the ecosystem to [27:03] make it work. Synopsys and high-level [27:06] design language, pretty new. [27:08] Cadence doing back-end design where you [27:10] could just take someone else's design [27:12] and integrate it into an overall flow, [27:15] pretty new. A set of software tools that [27:17] were as involved, all pretty new. So, [27:19] ARM was really driving a lot of [27:22] innovation. And because we were so new, [27:24] you know, again going back to the [27:26] superpower of a CPU is really the [27:27] software, we had no software. There were [27:30] no [27:31] application ecosystem that ran on ARM. [27:33] There weren't operating systems that ran [27:35] on ARM. So, [27:36] it was very difficult in the early days [27:39] to get some stickiness from a software [27:40] standpoint. So, our very first design [27:43] win that kind of made the company, [27:45] and it's again, it's a classic story of [27:47] of accidental empires where right time, [27:49] right place, [27:51] but now mobile phones are taking off in [27:52] the mid-1990s [27:54] and Texas Instruments is one of the [27:56] largest suppliers of baseband chips for [27:59] 2G phones and GSM phones. [28:02] And what they needed inside the phone [28:06] was a small microprocessor that could [28:08] help the baseband machine run. So, [28:12] the idea of the processor was not to [28:16] run any kind of applications cuz back in [28:18] the mid-1990s, there were no [28:19] applications that ran on a GSM phone. [28:22] The application was the phone. The [28:24] application was the phone, right? So, [28:25] the customer was TI, [28:28] but the big customer was Nokia. They was [28:30] the the first Nokia GSM phones that used [28:33] the TI chip that had an ARM CPU inside. [28:37] And TI chose ARM because [28:40] they looked around everything they had [28:41] and they didn't really have anything [28:42] that was as elegant as ARM. And they [28:45] thought, "Well, why would I design my [28:47] own CPU?" Cuz the value back then with [28:49] TI's product was in the radio. Wasn't [28:51] really in the processor. Every company, [28:54] if you look back in the chip world, [28:56] has a [28:57] design that was the market maker for [28:59] them. That was it for us. It really [29:02] reminds me [29:03] of the TSMC story and journey just, you [29:06] know, a couple years later, right? Of [29:08] like starting with like, "Okay, we're [29:09] going to take a layer of the stack here, [29:11] you know, them at the most lowest level [29:13] layer of production." And you guys one [29:15] layer up from that. And we're going to [29:18] make it available to all these people [29:19] who want chips, but like, "We're not [29:21] going after the PC market. We're not [29:22] going after, you know, anything big [29:25] that's going to be what it is today. Oh, [29:27] we'll start with this small stuff and [29:30] applications like these TI CPUs in a [29:32] component, not leading edge in the [29:35] fab terms. Great, we'll take that." And [29:38] it's just amazing over the next 20, 30 [29:40] years [29:42] how far it's come. And it's the same [29:44] echo of the Windows story, which is it's [29:47] fine to not make that much money early [29:49] on, but once everyone standardizes on [29:51] you, you have a lot of power in a [29:54] market. [29:55] Well, exactly right. Once we found our [29:57] way into the TI handset chipset that [30:01] went into the Nokia phone, now we have [30:03] traction. And now other folks who are [30:05] trying to build baseband chips for GSM [30:09] phones, [30:10] ARM becomes the de facto standard. [30:13] Not so much, quite frankly, because we [30:16] ran any operating system or we ran any [30:18] apps cuz there were none. It was just [30:20] simply, "Hey, it works pretty well. It's [30:23] got the right power. It's got the right [30:25] performance." And and off you go, which [30:26] is a lot of ways, you know, designs [30:28] ultimately sort of take off. [30:30] So, then you get into a what was the [30:31] lift, if you will, underneath the wings [30:33] of the architecture? [30:35] Well, fast forward, these GSM phones got [30:37] a little bit smarter. [30:38] And they began to run an operating [30:40] system called Symbian. [30:42] So, we actually began to have some level [30:45] of stickiness in terms of there was a [30:47] software community and development [30:49] ecosystem that that started to learn and [30:50] run on ARM. [30:52] But, I would say if I was to look back [30:53] and say, "Well, what what was the design [30:56] that took ARM completely into the next [30:58] level?" It was it was the iPhone. [31:00] If you look back at the iPhone, [31:02] because ARM now had some street cred, if [31:04] you will, in terms of low power, and we [31:07] had street cred in terms of we could run [31:09] small operating systems and small [31:11] applications, [31:13] we were chosen as the engine inside the [31:16] first iPod. Oh, I didn't realize that. [31:19] Yeah, so if you go back to early 2000s, [31:22] when the first iPod came out, [31:24] Yeah, those early little Toshiba hard [31:25] drives that kind of had no other use [31:27] case except for That's right. So, if you [31:29] remember that iPod, right? That iPod had [31:31] a kind of a crude display and it would [31:32] have uh you It had an operating system. [31:34] Had a little operating system. It had a [31:36] thumb wheel, so you had a UI. It had it [31:38] had all the things of a tiny little [31:40] computer. [31:41] The iPod was based on ARM. [31:44] So, fast forward, now this is early [31:46] 2000s. [31:47] As the 2000s are are moving forward and [31:49] Apple starts to futz around with are we [31:51] going to build a phone? Are we going to [31:53] build an iPad? And [31:54] revisionist history, you know, there's [31:56] all kinds of stories which one they were [31:57] going to build first, but it's probably [31:58] less important in that [32:00] they had a decision to make in terms of [32:02] what was the processor going to be [32:03] inside the iPhone. [32:05] Industry, you know, legend is that they [32:07] did talk to Intel about using Intel. And [32:09] Intel's processor of choice back then [32:11] was something called the Atom. Which was [32:13] their low power [32:14] or attempted low power device? [32:16] Respectfully, it was not really so low [32:18] power and it was not really so low cost. [32:19] It was kind of a very, very [32:21] stripped-down x86. And they were [32:22] building [32:24] Gosh, all this history, I'm kind of [32:25] going back in time here. If you guys [32:26] probably remember product called a [32:27] netbook? [32:28] Oh, yes. Of course. [32:29] The PC industry was lined up that [32:31] netbooks were the future, and that was [32:32] just flat out wrong. Exactly. [32:34] Well, it was right until the iPhone. [32:37] It was right until the iPhone. And Atom [32:39] was the chip inside the netbook. [32:42] Intel was coming from a very lofty place [32:45] of selling, you know, very, very high [32:46] performance and and very, very good Core [32:48] i7s, Core i5s. [32:50] The Atom, [32:52] you know, this is the classic [32:53] innovator's dilemma, right? You know, [32:54] innovate from the bottom versus the top. [32:55] Intel was having to come all the way [32:57] down from i7, i5, i3, Pentium, Celeron, [33:02] down to a little itty-bitty Atom, which [33:04] was designed for the netbook and was [33:07] probably okay for a stripped-down low [33:09] power netbook, a laptop, excuse me. [33:13] But, for a phone that needs to run at [33:16] even more low power not great, right? [33:18] But, Intel has got all of the street [33:21] cred inside of Apple at the time because [33:23] they've made that transition [33:25] by now away from power [33:27] to x86. So, all the the laptops inside [33:31] of Apple are all all running on x86. [33:34] Which is on its own a miracle. Like, [33:36] they changed a compiler to make it so [33:38] that applications written targeting a [33:41] power platform, the power architecture, [33:44] could suddenly now, with some changes, [33:46] compile to Intel. Oh my god, that is a [33:48] compiler miracle. Massive amount of [33:51] work, years and years of work by Apple. [33:53] So, you can imagine the debates inside [33:55] of Apple in 2006, 2007. [33:57] goal for the operating system, whether [34:00] you you know, phone, tablet, whatever it [34:02] was supposed to be initially, but was [34:04] basically run OS 10 [34:06] or a version of it on a mobile device. [34:09] And OS 10 ran on Intel at that point in [34:11] time. [34:11] Yeah, you you got you guys are are [34:13] bringing back all kinds of stuff that I [34:14] completely thought I had forgotten in my [34:16] memory. This is is a whole different [34:17] exercise on how neural nets work cuz you [34:19] know, you guys are uncovering all this [34:20] other stuff, but you had operating [34:22] systems like Leopard and Snow Leopard [34:24] and all all these things that were [34:26] pretty powerful, hefty operating [34:27] systems, right? They're running all on [34:29] on x86. So, [34:32] Intel and Apple have made the shift now [34:34] in the mid-2000s away from power into [34:36] Intel. You have all this investment [34:38] that's been made on these Mac operating [34:40] systems, as I mentioned, all of these [34:42] Tigers and Leopards that are all [34:43] optimized to Intel. [34:45] You have a big franchise inside of Apple [34:48] that is all based on Intel and the Mac [34:51] operating system. [34:53] And then you've got this little fussy [34:54] little iPod that runs on ARM with a kind [34:57] of a crude display. Which is basically [34:59] an embedded system. [35:00] Which is basically an embedded system. [35:01] So, you can imagine that [35:03] an easy choice would be we're going to [35:06] build this on Atom [35:08] and we're going to have the operating [35:09] system of Mac OS and and this new thing [35:12] look the same because software will be [35:14] easier, we'll strip it down, and we'll [35:16] just basically take our laptop and our [35:19] desktop operating system, [35:21] strip it down to the phone and run it on [35:23] Intel. [35:24] Or we can build up from this iPod, use [35:29] ARM, and build something called iOS, [35:32] which is the operating system for the [35:33] phone, [35:34] and it's going to be different than the [35:36] Mac OS, but you know what? This market [35:39] is very different. It's going to require [35:41] a different level of efficiency, [35:42] different level of power. If we clean [35:44] sheet it and do it right this way, or [35:46] the bias was at the time from the iPod [35:48] team was this is the right way to do it, [35:51] we'll end up with a better product at [35:52] the end of the day. So, that was the [35:53] debate inside. [35:55] And ultimately, the iPod team won. Sort [35:58] of, right? Didn't they kind of split the [35:59] baby where it was an ARM processor, but [36:01] it was a version of Mac OS's kernel that [36:04] had a new compiler written to target ARM [36:07] instead of [36:08] Oh, yeah, yeah, yeah, for sure. But, you [36:09] know, they didn't start from scratch. [36:10] But, yes, they started cutting things [36:12] down into simplify it and build it up. [36:15] But, yes, that was the key design win [36:16] for us. [36:17] Once that happened, [36:19] then very quickly you had followers from [36:22] the Android ecosystem, [36:24] the Samsungs of the world. And if you go [36:26] back in time, companies like HTC. [36:29] As Andy Rubin and Android started to [36:30] take off, [36:32] now ARM was seen as as the [36:34] de facto standard, and you had a lot of [36:37] work that was already now being done [36:38] around Linux and such. So, [36:40] we had the one-two punch of having the [36:43] iPhone and ultimately the Android [36:44] ecosystem designing around ARM, and this [36:46] is [36:47] 2007, 2008 time frame. [36:50] And at this point in time, just so [36:51] listeners can kind of anchor on what [36:53] ingredient to the stew does ARM provide? [36:55] They were standardizing Apple and these [36:57] Android vendors on ARM as the [37:00] instruction set architecture. Who was [37:02] actually making the processor in the [37:04] first iPhone? [37:06] Or in these other Android phones? If [37:08] 1981 is 2007, ARM is Intel. Except, the [37:13] benefit that ARM has is that [37:15] instead of Intel being Intel, in other [37:17] words, Intel builds the x86 and owns the [37:19] architecture, [37:20] ARM is licensing the architecture to [37:23] companies like Samsung. [37:26] To your question, if you go all the way [37:27] back in time, believe it or not, that [37:28] first iPhone chip, I think, was built by [37:30] Samsung for Apple. [37:33] And then ultimately, I think Apple went [37:35] to TSMC. [37:36] The chip vendors back in the day are [37:38] companies like Samsung, Qualcomm, [37:41] believe it or not, Nvidia. The Tegra [37:42] stuff was all ARM-based. [37:44] So, it was crowded. And why not? You [37:46] have this the smartphone market that's [37:49] now starting to take off, and chip [37:51] vendors now have an opportunity to build [37:54] chips for these phones based on ARM. And [37:56] again, if I do my IBM PC parallel, [37:59] it would have been as if Intel would [38:01] have licensed x86, and they did, to one [38:04] guy, AMD, because they were forced to. [38:06] And this is kind of interesting if you [38:08] just do the parallels. [38:10] Because IBM was so worried about [38:12] multiple sourcing, cuz the x86 was such [38:14] a a critical part, is that they [38:16] exercised, and I think I have this [38:18] right, to work on a second source for [38:20] x86. So, what you had with ARM was [38:24] multiple sources. So, you can see why [38:27] the business model suddenly became very [38:28] powerful. [38:29] Because now, to your point, what did we [38:31] provide in the stew? [38:33] Kind of the whatever the most basic [38:34] ingredient is in stew. I don't like stew [38:36] personally, so I don't know what the [38:37] best ingredient is, but let's assume [38:39] it's water. That without water, you have [38:41] nothing. We supplied the water. There [38:43] was no way anybody could do anything to [38:46] enter the smartphone market unless you [38:47] went through ARM. And there's an element [38:49] of portability. It's beautiful. If [38:51] you're Apple and you want to design the [38:52] next version of your phone, you're [38:53] thinking, "Well, there are a bunch of [38:55] ARM-based processors out there. So, as [38:57] long as we pick ARM, we have this whole [38:58] different sea of vendors, including [39:00] eventually ourselves after we acquire PA [39:02] Semi, [39:03] that we can sort of pick as our chip [39:05] vendor." That's right. So, they can [39:07] either pick companies that build ARM [39:10] chips, [39:11] or if they're brave enough and talented [39:13] enough and and smart enough, ARM will [39:16] give you the rights to build an [39:18] ARM-compatible chip yourself. So, rather [39:20] than licensing it from [39:22] or buying the chip from Samsung, who [39:24] used one of our designs, [39:26] you can just go build your own, which is [39:28] what Apple did. All right. So, now that [39:29] we're in these early 2010s period, this [39:31] is probably a good place to explain the [39:34] dual ARM business models. So, at least [39:37] at that point in history, how does ARM [39:39] make money? So, [39:41] back to the simple concept of licensing [39:43] and royalties, our business model way [39:46] back in the day, and it still pretty [39:48] much holds, is that we have an upfront [39:50] fee for licensing and royalties. As you [39:52] can imagine, when you're starting out [39:54] and a lot of companies aren't actually [39:56] shipping any volume, [39:58] the vast majority of your revenues come [40:00] from licensing. And the proxy for that [40:03] in the chip world is design wins. So, [40:05] you get a lot of design wins, you get [40:06] people committed to the architecture, [40:08] but they don't actually ship any volume, [40:10] so you don't really get to a a mix of of [40:13] royalties until you're in volume. So, [40:16] it took a long time, but licensing was [40:18] bigger than royalties for many years. [40:19] And [40:20] you could look at that glass half Yeah, [40:22] you could look at that glass half full [40:23] and say, "Wow, the future is going to be [40:25] bright if you ever get there." Glass [40:27] half empty would be [40:29] Hey, this stuff doesn't work. Yeah. [40:31] Yeah, but I'm I'm betting on the front [40:32] end and are these things ever going to [40:34] see the light of day? [40:35] Now, one thing that we we changed I [40:37] wouldn't say changed, but another [40:39] version of the business model is the [40:40] license. You can either [40:42] license an core that we built, we call [40:44] that an implementation, that is we [40:46] basically do the blueprint and says the [40:49] house looks like this. [40:50] And you let This means in-house you have [40:52] your own chip designers, they're using [40:54] Cadence and Synopsys, and they're floor [40:55] planning, and they're actually you know, [40:57] they're doing what everyone sort of [40:58] imagines Nvidia is doing over there. [41:01] That's right. There were a set of [41:02] customers that believe that either due [41:05] to the link between hardware and [41:07] software or the ability of their [41:09] engineers to develop something that [41:11] would be higher performance than what we [41:12] could build, [41:13] we had these architectural licenses. And [41:15] it allowed customers to build their own [41:17] implementation. Now, one of the things [41:19] that sometimes gets confused about these [41:21] licenses is that [41:24] are they able to run software that's not [41:26] ARM compliant? In other words, can they [41:28] add some special instructions that [41:30] nobody else has, which gives them a [41:32] unique advantage? And they're not [41:33] allowed to do that. And the reason is [41:35] very very simple. [41:37] Once instructions look different across [41:40] a number of different architectures that [41:42] a customer has, software can't [41:44] understand it. And let me drill on that [41:46] a little bit further. If customer A has [41:49] an instruction that says accelerate, and [41:51] customer B has an instruction that says [41:53] accelerate 2X, and customer 3 has an [41:56] instruction that says accelerate 3X, [41:59] if I'm a software developer and I'm [42:00] writing software for ARM, [42:02] I really don't want my program taking [42:04] advantage of the 3X instruction, cuz I [42:07] don't know that everybody has it. So, I [42:09] end up going to something we kind of [42:11] call inside as a lowest common [42:12] denominator approach [42:14] that [42:15] the software developer would not make [42:17] use of those instructions. So, it's one [42:18] of the great things a company has done [42:20] in its early days, and and we've [42:21] maintained it certainly since I've been [42:22] been running it. [42:24] We're never going to break the ISA. [42:26] We're not going to allow people to add [42:27] custom instructions, because once you do [42:29] that, you break software compatibility, [42:32] which is one of the superpowers of ARM. [42:34] And if you think about, you know, why [42:35] did X86 stay so sticky on the IBM PC? [42:38] It's because [42:39] Intel was the only game in town, so of [42:41] course they were going to run. And [42:42] that's why Compaq and Dell and [42:45] all these other clone guys were able to [42:47] copy the PC, because the software just [42:49] ran. And if they were not able to do it [42:51] in such a way that IBM did it, they [42:53] could never be successful. So, [42:55] we offer these licenses, they're [42:56] architectural licenses, but all they [42:58] really do is allow people to build their [43:00] own implementations. And I will say, [43:01] just adding on to that, I know we're [43:02] going back and forth between the future [43:04] and the past, we used to do a lot of [43:05] them [43:06] because customers used to believe that [43:08] A, they could build a better design than [43:10] ARM, and or B, there was something [43:12] specific in the software they wanted to [43:13] take advantage of. Not many people do [43:15] them anymore. They're really hard. And [43:17] back to the 10 to 15% advantage or even [43:19] 5% advantage, [43:21] the ROI isn't all that high. And if [43:23] you're going to have three or four [43:25] engineers designing an ARM CPU that you [43:27] can buy from ARM anyway, [43:29] why not take those three or four hundred [43:31] engineers and put them on IP that you do [43:33] as a customer that only you do? Nobody's [43:37] building a CPU with three or four [43:38] engineers. Sorry, it's it's three or [43:39] four hundred or thousand. [43:41] Yeah, three or four hundred. Yeah, no, [43:42] not three or four. Uh, if I said three [43:43] or four, that's a that's a big way. [43:45] Yeah, three or four hundred at least. [43:46] It's a lot of work. It's hard. Well, I [43:48] imagine for probably almost every [43:50] customer out there now, [43:53] the ecosystem [43:54] and compatibility of software across all [43:58] vendors, you know, all applications out [44:00] there is [44:01] worth so much that it you know, they [44:03] wouldn't even consider going out and [44:05] altering the instruction set, right? [44:07] Because then they would lose [44:08] compatibility with the rest of the [44:09] ecosystem. And I know we're hopping [44:11] around, you know, in terms of history [44:12] dates, but that's one of the things that [44:14] I think gets lost in terms of what's [44:16] gone on with [44:17] with CPUs and software compatibility [44:19] over the last 15 20 years, because [44:22] as we were talking about in 1980s, early [44:24] 1990s, I mentioned a lot of [44:26] microprocessors, right? The 68000, [44:28] PowerPC, 29000, DEC Alpha, SPARC. [44:33] There's a pretty large graveyard of [44:35] CPUs. And they were very good products, [44:38] very very good in terms of performance, [44:39] very very good in terms of their design. [44:42] And they've just entered the graveyard [44:44] of CPUs. And you say to yourself, "Well, [44:46] why why do they all die off?" Well, once [44:49] the flywheel of software gets built onto [44:52] a certain architecture, [44:54] it's very very difficult for A, if [44:56] you're developing a new piece of [44:57] hardware to say, "Well, I'll choose one [44:59] of the ones I just mentioned, because [45:00] there really isn't a software [45:02] story around it." So, they all began to [45:04] to wither away. [45:06] Once the internet took off, and [45:08] particularly as you got into the dot-com [45:10] era and a little bit after it, [45:12] huge amounts of investment started to go [45:15] into software companies. [45:17] And software as a service, [45:18] subscriptions, SaaS models, recurring [45:20] revenue, everything around the software [45:22] industry, which was wonderful. Two [45:23] things happened with that. Number one, [45:26] it drove an increased innovation and [45:29] investment into software, and all levels [45:32] of software, the complexity of software, [45:33] software stacks that run in the cloud, [45:35] that run in a network switch, that run [45:37] in an automobile. [45:39] And at the same time, semiconductor [45:41] investments, which is changing a little [45:42] bit now, began to wane. And very little [45:45] venture money [45:47] started to go into startups, and [45:49] semiconductor startups in particular. [45:52] Well, that's the fertile ground where [45:54] new innovation happens, whether it's [45:55] around new compute architectures, [45:57] including CPUs. [45:58] So, you had [46:00] very very little innovation [46:03] taking place with companies building [46:05] CPUs and startups. In fact, I was with [46:07] one of the very last ones [46:08] founded in the late 1990s, company [46:10] called Tensilica. And we were a bunch of [46:13] ex-Synopsys and ex-MIPS guys [46:16] building configurable processors. And [46:18] the idea there being that [46:20] you could build a custom piece of [46:22] processor with your own custom [46:23] extensions, et cetera, et cetera. And [46:26] we started in 1997, I think, and I left [46:29] in 2004. Company was ultimately bought [46:31] by Cadence, I think, in 2012. Shipped a [46:33] lot of cores, I think hundreds and maybe [46:35] over a billion cores. But the point was [46:37] is after Tensilica and another company [46:39] called ARC that was doing the same [46:40] thing, there was very very little [46:42] innovation taking place or investment in [46:46] semiconductor CPU startups. The great [46:48] irony of the namesake of Silicon Valley [46:51] is that if you were a silicon startup, [46:52] you could no longer raise venture [46:54] capital dollars there. [46:55] Yes, exactly. Exactly. So, what you have [46:58] is [46:59] as all these architectures start to wane [47:02] away, [47:03] and the amazing amount of investment [47:05] that's now going into the software [47:07] industry in general, and all of the [47:10] investment going into stuff going into [47:12] the cloud, [47:14] two architectures really ultimately [47:16] remain. X86, which has been around for [47:19] 40 plus years, [47:21] and ARM. You know, we were talking [47:22] earlier about the data center. [47:24] Why ARM in the data center? Well, you [47:26] know, two things. First off, [47:28] the choices aren't massive. It's not [47:30] like there's 17 different choices, as we [47:32] just talked about. And number two, one [47:34] of the things that's becoming extremely [47:35] important in the data center is its [47:36] power efficiency, because when you're [47:38] running these extremely large loads, [47:40] whether it's general purpose compute, [47:42] and now with the advent of running [47:44] accelerated compute with AI models, you [47:47] need incredible efficiency in the [47:48] processor space. So, I think we we've [47:50] arrived at this place both as a [47:52] combination of having A, a really really [47:55] good low power architecture, B, an [47:58] incredible amount of software innovation [48:00] that's been done on ARM, and C, just [48:03] optionality has kind of gone away, [48:05] because investment has has waned. [48:07] That last one is just like a last man [48:09] standing. You know, why is the winner [48:11] the winner? Well, there was going to be [48:13] a winner as all the sort of competitors [48:15] fell by the wayside. It's almost [48:16] tautological that, you know, whoever [48:18] becomes the winner, there was going to [48:20] be one who was left standing, or two in [48:22] this case. I would argue it's not one of [48:24] these industries where last man standing [48:27] has occurred because the market is [48:29] uninteresting. It's actually the [48:31] reverse. The market's never been more [48:33] interesting, but because of the massive [48:35] amount of investment required from a [48:37] software standpoint, [48:39] optionality is limited, because if you [48:41] were to rock up today and say, "I want [48:42] to go build a system on chip based upon [48:44] the Motorola 68000 architecture," [48:47] what software exists is going to run on [48:48] it? It's so funny. It really is just [48:50] like the fab industry, right? Of like [48:52] the capital investment required and the [48:55] software investment required is so [48:57] massive that like you get to where we [48:59] are now, where you've got TSMC, you've [49:01] got Samsung, You've got GlobalFoundries. [49:03] Yeah, but at the leading edge, like, you [49:05] know, [49:06] it's kind of all that's left, right? [49:08] There are definite parallels. You know, [49:09] the fab industry [49:10] is direct CapEx, and you'd look at it [49:12] and say, "If I'm going to build a 2 [49:14] nanometer fab and beyond, I'm going to [49:16] have to have 30 to 35 billion dollars of [49:18] CapEx." [49:19] Our industry is not that, but on the [49:21] flip side, it's not unlike that when you [49:23] think about the [49:25] OpEx of all of the 20 million developers [49:29] and plus that have developed on ARM. [49:31] You're actually having to tilt that. [49:32] Incredible momentum there. [49:34] I still am floored [49:37] by this architecture that was originally [49:41] built not to melt plastic, to be super [49:43] low power, ended up becoming, I'm sure [49:47] you have better stats than I do, but a [49:49] dominant architecture running in data [49:51] centers doing this heavy compute load, [49:54] AI training, inference. Maybe Rene, I [49:57] could ask you with your most honest [49:58] assessment on [50:00] where is there still a place for x86 [50:04] architectures versus where is there a [50:06] place? Like should the whole world be [50:08] ARM? Is it just actually better? Or are [50:11] there different use cases for each? [50:14] You know, I'm I'm going to try hard to [50:16] be unbiased even though my even though [50:18] my my job is the CEO of ARM. [50:22] There's a lot of things that that are in [50:24] our favor. One of them is quite frankly [50:27] the fact that we have an open model [50:30] where our products can be built at any [50:32] fab by any chip company. [50:35] So, if you're looking at x86, you're [50:37] looking at two people who build it. [50:40] Now, one of them builds at TSMC, AMD, [50:43] and the other one builds in-house at [50:46] Intel, although they build a whole bunch [50:47] of stuff at TSMC too these days. [50:49] But it's just two people. And not only [50:51] are you betting on those two people, [50:54] but the IP around the chip that they [50:58] build, whether it's around [50:59] communications, whether it's around [51:01] accelerated computing, whether it's [51:02] around [51:04] network storage, you're you're banking [51:06] on on that to to bring a lot to the [51:08] party. And then one might look at it and [51:10] say, "Well, why isn't Intel and AMD just [51:11] license x86 and just flatten out the [51:14] playing field and and [51:17] maybe that playbook probably could have [51:19] been run a while ago. Well, also, when [51:21] you have a high-margin business model, [51:22] it's very hard to switch to a low-margin [51:24] business model. Bingo. ARM came from a [51:26] very different place. So, [51:28] as a result, we we have a huge advantage [51:31] just just with our model. Now, in the [51:33] data center, we have another fairly [51:35] significant advantage in that [51:38] if you look at customers like Microsoft [51:40] or Google or AWS, all who have custom [51:44] chip efforts on ARM, [51:46] all who have talked about getting 60% [51:48] benefit in terms of performance on a [51:50] like-for-like basis, [51:52] that's not just the ARM ISA. That's not [51:55] just the fact that we are more efficient [51:57] than x86. That's in that they can build [52:00] a custom SOC [52:02] with a custom piece of memory, let's [52:03] say, or custom storage, or custom blade, [52:08] or custom interconnect, or custom [52:10] offload, where from a TCO standpoint, [52:13] their optionality is incredible. [52:16] And as a result, their flexibility in [52:18] terms of building something [52:20] that is absolutely right for an Azure [52:23] estate or a GCP estate or an AWS estate, [52:27] because they they have the kind of [52:29] volume and spend that can drive that. [52:32] So, again, you know, one of the one of [52:33] the benefits we get with the [52:34] hyperscalers is because, no pun [52:36] intended, the scale is so large [52:38] and doing custom chips, they can get an [52:41] ROI on it. [52:42] You can't do that with x86. Right. You [52:45] go to Intel and they say, "Here's my [52:46] product for you." Here's my product, [52:48] right? And then you've got to kind of [52:50] put the pieces together and see how it [52:51] all fits. So, [52:53] that in itself gives us a big advantage, [52:55] and we have optionality with with people [52:58] like Ampere, for example, who do [52:59] standard products, [53:01] but that kind of optionality of there's [53:02] a standard marketplace and/or a custom [53:04] play, or Grace, for example, you know, [53:07] the CPU from from Nvidia, you could buy [53:09] Grace and/or [53:10] the way they ship it today increasingly [53:12] with Grace Blackwell, where it's highly [53:14] integrated. [53:15] And again, why Grace Blackwell versus [53:19] Intel plus Blackwell or AMD plus [53:20] Blackwell? [53:22] Well, if you look at the architecture [53:23] and some of the things that they do with [53:25] NVLink and how they couple the CPU to [53:28] the GPU and how the interface between [53:31] HBM memory and CPU memory, [53:34] just they can't do that in in x86 world. [53:37] And then by the way, in a Grace [53:38] Blackwell system, the other benefit you [53:40] have is that Grace can run all major [53:42] pieces of the operating system. [53:44] You can run an AI cluster, AI cloud, [53:47] and the software stacks that are sort of [53:49] native that run for an ARM general [53:51] purpose compute can run in your AI [53:53] cluster. So, that in itself gives, you [53:56] know, huge huge optionality. So, [53:59] I don't know how we if we started on [54:00] this. I was trying to I was I've [54:01] advocating what to do about x86 and I [54:03] started talking about ARM all day, but [54:04] you know, hard. Just just hard. [54:07] Yeah. [54:08] Yeah, it makes total sense. Okay, so, [54:10] we'll call a spade a spade. We're at the [54:11] present. We have come forward today, and [54:13] I want to talk to you about a couple [54:14] things. One, how the business model has [54:17] evolved and how you deal with your [54:19] customers differently and your your sort [54:20] of products that you sell to customers [54:23] now and the way in which you work with [54:24] customers. And the other of which is [54:27] last quarter as of recording, you did [54:29] 939 million in revenue, so right around [54:32] a run rate of $4 billion. [54:34] The market cap is about $150 billion. [54:37] Investors think the future is very, very [54:39] bright for this company as we move into [54:41] this world of AI and connected devices [54:43] everywhere. [54:44] Why are people so insanely bullish on [54:46] ARM? What is the incredible future hold [54:49] and why is that valuation the valuation? [54:51] We've been talking for 40 minutes or so, [54:53] but hopefully these last 40 minutes have [54:54] been helping sort of build that Yes, [54:56] very much so. build that case study. [54:58] I think it kind of goes back to the [55:01] fundamental advantages both from a [55:03] technology standpoint and probably more [55:05] importantly, as it tends to be with with [55:08] this world, the market forces that are [55:10] are in our favor. [55:11] If you just start with the fact that [55:13] more and more chips are shipped every [55:15] year and more and more of those chips [55:17] are based on ARM, [55:18] and you look at the end markets, whether [55:21] the examples I gave you in in my house, [55:23] from my car to my camera to my stove to [55:26] my [55:27] they are all ARM-based and they are all [55:30] in a growth mode. You look at it and [55:31] say, "Gosh, there's a there's a ton of [55:33] tailwind associated with with this [55:35] company." [55:36] And maybe people are a bit more excited [55:38] since the IPO, I don't know, is around [55:40] the fact that [55:42] AI has created this next level [55:46] of compute need. Now, one can argue [55:50] incessantly around, "Well, gosh, you [55:52] know, $40 for Copilot, am I really [55:53] getting the ROI on that? And what are [55:55] the near-term economic models?" And [55:57] You sound like Marc Benioff. I just [55:59] think the near-term economic models on [56:01] on AI is kind of the wrong way to think [56:03] about it. I look at it much more in the [56:05] parallels of the automobile, the [56:07] industrial revolution, the smartphone [56:09] revolution, the internet revolution. [56:12] And for a company like ARM, because AI [56:16] requires a next level of compute [56:19] capacity and capability. In other words, [56:22] and it's not just, you know, running [56:24] strawberry training models in the [56:26] massive amount that's required to to [56:28] train all these next generation LLMs or [56:30] even beyond large language models, [56:32] video-related models, [56:34] but it's actually then running those [56:35] applications, [56:37] the inference in your car, on your [56:39] stove, [56:41] in your headset, on your wearable. [56:43] Inference is going to run across all [56:44] those workspaces. [56:46] That all requires a lot of compute. [56:49] And one of the things that we used to [56:51] talk about, you know, when I was at [56:52] Nvidia, was [56:54] what is the death for anybody who's [56:56] either in the computing category or [56:58] accelerated computing category, and [56:59] that's to eat when you get to end quote [57:01] good enough. [57:02] I remember being in good enough. I've [57:03] been [57:05] in the semiconductor industry since I [57:06] got out of school in 1984 and started [57:08] TI, and there's and there's definitely [57:09] been periods of of good enough. I think [57:11] the late 2000s, early 2010s felt like [57:14] good enough. Like netbooks were a good [57:15] definition of good enough, where [57:18] at that time it didn't seem like you had [57:20] the application space and area to drive [57:23] the need for more compute. So, what did [57:24] you end up building? A little crummy [57:26] $199 computer because [57:28] it could do everything your big computer [57:30] did. So, we've definitely had periods in [57:32] the ever in our industry where good [57:35] enough has existed and the need for [57:37] compute innovation has slowed. It's [57:38] never stopped, but it's slowed. [57:41] With AI, [57:43] in the foreseeable future, [57:46] you look at it and say, [57:48] "This appears to be almost the benefits [57:50] that AI could bring, whether it's around [57:52] education, [57:53] drug research, [57:55] investment, it's mind-boggling. [57:58] So, ARM's going to be in the center of [57:59] that. [58:00] Whether it's in the data center, whether [58:01] it's in your automobile, whether it's on [58:03] your smartphone, whether it's in your [58:05] wearable, the AI compute path is going [58:07] to run through ARM on some way, shape, [58:09] or form. So, It's kind of like the Bezos [58:11] comment of, you know, I can't imagine a [58:13] future where my customers ever say, "You [58:15] know, gosh, I wish this were a little [58:16] more expensive." You can't imagine a [58:18] future where, "Gosh, I wish GPT-7 were [58:21] just a little dumber, you know." I [58:23] actually like the fact that people look [58:25] at it and say, "I'm not really seeing [58:26] much benefit from this yet." [58:28] Because that actually says, "Oh my gosh, [58:30] what a fantastic opportunity to innovate [58:32] and do more." And a big part of it is [58:35] the hardware [58:36] that you're seeing today, particularly [58:38] the edge-based hardware. [58:40] Those were designed [58:42] a couple years ago when these large [58:44] language models weren't even needing to [58:45] run locally. [58:47] So, you have completely unoptimized [58:49] architectures everywhere [58:51] to take advantage of [58:53] the AI capability that we're going to [58:55] unharness. So, to me, [58:57] I look at this and it's like white space [58:59] in terms of the the compute opportunity. [59:02] Which back to the question that Ben [59:04] asked in terms of why people so bullish [59:05] on the company, [59:07] I'd like to think that's why it is. We [59:08] play in a super large market. [59:11] Semiconductors are a trillion-dollar [59:12] market by the end of the decade. [59:15] You said we're $4 billion. [59:17] We probably could take a bigger chunk of [59:18] that $1 trillion market at some point in [59:20] time because of the importance of the [59:22] company. [59:23] And this is a good lead-in to this [59:25] question I have for you. I've heard you [59:27] espouse this idea. I'm sure there's a [59:29] way to rationalize these two things, but [59:31] it almost feels heretical. [59:33] You open the episode by saying, we do [59:35] CPUs. [59:37] The whole industry over the last 5 or 10 [59:39] years, including David and I on our [59:41] Nvidia episodes, had this obsession with [59:44] GPUs, with accelerated computing, with [59:46] get those stupid serial workloads off [59:50] the CPU, get them onto the GPU where you [59:52] can do pure magic with it. This enabled [59:54] the whole AI revolution. [59:56] You're the CPU company. And I've heard [59:58] you talk about, okay, now that we know [60:01] some of the use cases that are happening [60:04] on GPUs, [60:05] history has shown us that those kind of [60:07] tend to migrate back to the CPU over [60:09] time, and CPU the definition of CPU kind [60:11] of changes. How do you view the state of [60:14] things right now with everyone being so [60:16] excited about GPUs and incredibly [60:19] parallel GPUs of the future, and CPUs, [60:21] yeah, CPUs are fine, but they're a known [60:22] quantity. [60:24] I think accelerated computing and the [60:26] advent of GPUs is fantastic for Arm. [60:29] Because what it indicates is that [60:31] there's lots of compute out there, and [60:33] more compute needs to run in such a way [60:35] that you're have not only base compute, [60:37] but accelerated compute. The reason I I [60:39] think it's kind of [60:41] over simplified because it's almost the [60:42] notion of, oh, and I've met with [60:44] investors who have have had this [60:45] questions to us and say, well, [60:46] everything's moving to the GPU, do you [60:48] need a CPU anymore? [60:51] It's almost like I saying, well, I've [60:52] got this V6 engine going to a V8, I [60:55] don't need tires and a steering wheel [60:56] anymore, do I? [60:57] It's nonsensical. Just think about the [60:59] architecture of it. [61:00] So, with the advent of of all of these [61:03] accelerated computing models that are [61:05] doing again, it's primarily the data [61:07] center. Let's just be very real about [61:09] this, right? It's [61:10] it's all happening in the data center. [61:13] It's a fantastic outcome for for CPUs. [61:16] Why is that? Well, number one, all these [61:18] data centers need CPUs, obviously, and I [61:20] just gave the example in Grace Blackwell [61:22] why that's a great positioning piece for [61:23] for Arm. [61:25] But more importantly, all of that [61:27] training converts into inference. If [61:29] training is the teacher, inference is [61:31] the student, and there are far more [61:33] students than teachers in the universe, [61:35] and that's why there'll be far more [61:36] inference workloads than than training. [61:39] And that's going to run everywhere [61:41] relative to the smallest devices, [61:44] whether it's wearables, whether it's a a [61:46] headset, augmented reality. [61:49] You're not going to run a 100 watt GPU [61:51] on your head. I'm sorry, it's not going [61:53] to happen, right? You're going to [61:54] You're You're going to have to get into [61:55] very, very different form factors. Now, [61:58] naturally, a CPU's going to be there. [62:00] You can't have an accelerator [62:03] out there without something that's [62:04] running the main of the system. [62:06] That's a fantastic opportunity for Arm, [62:08] because it means a couple of things for [62:09] us. We can solve that in a few ways. We [62:11] can add more and more capability to our [62:13] CPUs, which we are today, around [62:16] extensions that help with AI [62:18] acceleration, and this goes back to [62:20] RISC versus CISC and things that we can [62:22] add in terms of just extensions that [62:24] will help with AI. [62:26] But also back to the customization, you [62:28] could add small AI acceleration, which [62:30] we do today with our Ethos NPUs that are [62:32] four tops, eight tops, etc. [62:34] That will do some level of offload. And [62:36] I think the model the model will be for [62:39] these edge devices to run in conjunction [62:41] with cloud, [62:43] where you're going to have some [62:44] processing happening locally, some [62:45] processing going to be happening in the [62:46] cloud. [62:48] You're going to need to have some level [62:49] of security and authentication and [62:51] attestation locally so that the models [62:53] know that it's you and it's not somebody [62:55] else, and [62:56] the information's kept private to you. [62:57] So, [62:58] game on. I mean, all this all this GPU [63:00] accelerated compute is wonderful for us, [63:03] because it's it's just going to drive [63:04] incredible incredible demand. And the [63:06] idea that the only way you'll ever run a [63:08] computer is through a a large GPU in the [63:11] data center, it's just [63:13] it's not the way the world works. And [63:14] the last thing I'll say on this, and I [63:15] love Jensen, he's done a brilliant job [63:17] with the company, [63:18] but remember he tried to buy Arm. Well, [63:20] I was going to say, there's no better [63:21] data point than this than Nvidia tried [63:23] to buy Arm. [63:25] And when he tried to buy Arm, Arm was a [63:27] $2 billion company, and he was a $25 [63:29] billion company. [63:30] He certainly didn't do it because he [63:32] wanted to be revenue accretive. [63:33] He knew the importance of what Arm meant [63:36] to the industry. Was that really the the [63:38] valuations of both companies at the time [63:40] of the No, that was their revenue rates. [63:42] I think [63:43] But Nvidia was [63:44] Nvidia tried to buy us for $40 billion [63:46] back in 2020, and I think their market [63:49] cap was [63:51] 350 to 400. Wasn't anything close to [63:52] what it is now. [63:53] If you looked at from the outside in, [63:56] back then, 2020, [63:58] we had not yet gone public, and we [64:00] hadn't really started the turnaround in [64:01] our in our core businesses yet. So, [64:04] there were a lot of people at the time [64:05] looking at the deal, Masa bought Arm in [64:07] 2016 for $32 billion, [64:10] and basically sold it 4 years later for [64:13] 32 plus some change, you know, $40 [64:15] billion. [64:16] There were a lot of critics of the deal [64:17] that said, Nvidia overpaid for this [64:19] thing, because it's not really a growth [64:21] company. [64:21] Do you mean in SoftBank overpaid? No, [64:23] I'm sorry. That Nvidia overpaid for Arm. [64:26] Hmm. Yeah. That had the acquisition gone [64:29] through, Nvidia would have been [64:30] overpaying for it. Yeah, I'm sorry. The [64:31] price that they put down, $40 billion. [64:33] There was a lot of criticism that they [64:34] had overpaid back in the day. [64:37] Now, you looking back at now, and it [64:38] seems Well, laughable. Yeah. Laughable. [64:41] Yeah, laughable in terms of their market [64:42] cap's probably 10x, their revenue's 4x. [64:45] And by the way, the last thing I would [64:46] say about that about that acquisition, [64:48] you know, first off, a lot of people [64:50] thought Nvidia overpaid. [64:52] Secondly, a lot of people hated it. And [64:54] then there was a lot a lot of opposition [64:57] that we got from regulators, customers, [65:00] ecosystem partners, [65:02] which I think belied kind of the [65:04] importance of the company in a in a kind [65:05] of a roundabout way that said, gosh, [65:08] this is [65:09] a company being bought for this amount [65:11] of money at this valuation and so people [65:13] against it, [65:14] maybe the company is more important than [65:16] than folks had originally gave us credit [65:18] for. [65:19] This kind of seems like an area where [65:21] regulation did exactly what it's [65:23] supposed to. You were a broad horizontal [65:25] provider that served a whole bunch of [65:27] customers that was integral to an [65:28] industry and is kind of essential for [65:31] the further advancement of humankind. I [65:33] mean, truly in our sort of most [65:35] important innovation area. [65:37] And one of your customers wanted to own [65:40] all of it, which over time presumably [65:42] means all the other customers wouldn't [65:44] quite have the same access to it. It was [65:46] a fascinating case study because I [65:47] learned a lot about M&A and and [65:49] regulatory. [65:51] And one of the things that had surprised [65:52] our teams that were advocating on the [65:55] deal was that generally most of the [65:57] blocking takes place and let me back I [66:00] would say this way. It was a vertical [66:01] merger, right? So, it wasn't a [66:02] horizontal merger. It was a vertical [66:03] merger. [66:04] And typically in a vertical merger, [66:06] people will object to the merger if it [66:08] forecloses a market or the stifles [66:11] competition in a in a given market. [66:13] But at the time, we were predominantly [66:16] smartphone revenue, and Nvidia's not a [66:18] smartphone company. [66:19] And the folks looked at it and said, [66:20] well, because it doesn't really violate [66:23] a vertical integration mantra, [66:25] and regulators tend to care more about [66:27] the near term than the long term, this [66:28] should be okay. [66:29] But what they actually did in that case [66:31] was cared much more about the long term [66:33] of the what may happen someday versus [66:35] what we think will happen in the near [66:37] term. [66:38] I'm curious actually if you know, since [66:39] you were at Nvidia for a long time, [66:42] the Arm journey for Nvidia also seems [66:45] like an improbable one, right? Because [66:48] Nvidia started as, you know, obviously a [66:50] graphics card company for PCs which ran [66:53] on x86, and then [66:56] did this incredible shift into the data [66:58] center. But at the time as they were [67:00] making that shift, data center was also [67:02] an x86 environment. When did the company [67:05] start really realizing, [67:08] hey, this Arm platform is going to be a [67:10] lot more than just not melting plastic [67:12] in, you know, [67:13] 2G phones? Nvidia's been an amazing [67:15] partner for Arm. When I was working [67:17] there, [67:18] we [67:19] made a very distinct pivot to try to [67:22] accelerate our mobile business with [67:24] Tegra, and really accelerate everything [67:26] we were doing with Arm. [67:28] Nvidia bought a company by name of [67:29] PortalPlayer. I may remember those guys, [67:31] and they they were actually doing the [67:33] audio chip for iPod back in the day. [67:35] And we at Nvidia were, you know, we were [67:37] doing [67:38] we were actually doing the SOC for Zune. [67:40] I don't know if you guys remember that. [67:41] That's right. Oh, yeah. Yeah, that was [67:43] the the Microsoft equivalent iPod. We [67:46] had been flirting with all kinds of [67:48] stuff that were Arm based, whether it [67:50] was with Microsoft with Windows [67:52] CE and Zune. [67:54] But the real thing that had Nvidia [67:55] double down on it was A, when when the [67:58] smartphone thing, you know, really took [68:00] off. That was number one. [68:02] And then number two, and this was the [68:03] business I was managing at the time, [68:05] this is 2009-ish [68:07] time frame, is when Microsoft made the [68:09] commitment to do Windows on Arm. [68:12] And we felt at Nvidia at the time that [68:14] we were very well positioned to do very [68:16] well in that market because [68:18] of all the history that Nvidia had with [68:21] the Windows ecosystem, all the work that [68:23] they had done with PC gaming. DirectX, [68:26] yeah. Yeah, I was running the business [68:28] for all laptops back then for Nvidia. [68:30] So, I I took over all the Windows on Arm [68:32] stuff. So, I was doing that first hand [68:34] myself. Which Windows on Arm is like [68:36] another miracle if you can make it [68:38] happen. All that translation layer, all [68:41] those compilers, everything that's, you [68:43] know, been written for decades [68:45] specifically for x86 chips, [68:48] theoretically, you're going to be able [68:49] to press one button, compile your code [68:51] differently, and now it runs on Arm. I [68:52] mean, that is quite the promise. [68:54] Yeah, and and a lot of the native stuff [68:55] now has all been ported to Arm, and that [68:57] really benefited from stuff on mobile, [68:59] right? If you think about all the all [69:00] the apps, [69:01] all the Microsoft apps that run on iPads [69:03] today, you know, whether it's Office and [69:04] and whatnot. So, we we got a huge [69:06] benefit of that. But [69:08] going back to your question, David, in [69:09] terms of of Nvidia, they stuck with Arm [69:11] for quite some time. We stuck with Arm [69:14] with Windows on Arm, and then after I [69:15] left, Arm became the default platform [69:18] for everything they're doing on [69:19] automotive. So, if you look at the [69:21] Nvidia Drive platform, [69:23] everything Nvidia does around robotics, [69:25] that's all Arm based. So, everything [69:28] that they do that uses [69:30] accelerated computing, [69:32] that whole software stack is all all [69:33] runs on arm. [69:35] Which is, you know, why arm is so [69:37] ubiquitous in automotive. If you look at [69:39] work done by Renesas, or work done by [69:41] Nvidia, or work done by Qualcomm, a lot [69:45] of those software stacks are are now [69:47] native and and all run on onto arm. So, [69:50] it's a why we're so strong in the [69:51] automotive space. And then back to your [69:52] Nvidia question, they were very [69:54] committed to arm for a long long time. [69:56] And a combination of A Tegra, B Windows, [69:59] then all the stuff on auto, And then the [70:01] data center probably really starting to [70:03] come online. [70:03] center really started to take it off. [70:04] And I think they saw, you know, back to [70:05] the customization, the way they [70:07] architected Grace Hopper, and now Grace [70:10] Blackwell, gives them a a degree of [70:12] innovation that they can't get any other [70:14] way. [70:14] All right. On a closing topic, I want to [70:16] ask about what seems to me to be a [70:18] little bit of a strategic evolution. Can [70:21] you tell us what you're doing with [70:22] subsystems, and how that came to be? [70:24] Yeah. So, subsystems are kind of a [70:26] natural extension of a of an IP business [70:29] model. So, the core model of of doing [70:31] CPUs, and I say we do CPUs, I [70:33] oversimplify. There There's a lot of [70:35] other products we do inside the company. [70:36] We do GPUs, we do NPUs for AI. [70:40] We do all of the the complex [70:42] interconnect that's required to build a [70:44] server chip, CMNs, which are coherent [70:47] mesh networks. And these are [70:49] essentially the the plumbing. If you're [70:51] building an SOC that has 128 CPUs, you [70:55] need this mesh network that helps [70:57] connect the CPUs together and then [70:59] interfaces them in the memory. It's just [71:00] a lot of a lot of plumbing. [71:02] And this is the analogy, I know your [71:04] audience is pretty technical, but I used [71:05] during our our road show. Think of all [71:07] those things as disparate Lego blocks. [71:10] To very sophisticated customers, you can [71:13] basically sell them or provide to them [71:16] these Lego blocks, [71:17] and they will provide a beautiful copy [71:19] of the Statue of Liberty. Or, you can [71:22] basically say, "Look, connect everything [71:24] exactly this way, in this particular [71:26] form, and you will get the Statue of [71:28] Liberty a heck of a lot faster than if [71:30] you built it yourself." [71:32] It really is like Lego. [71:35] That's what compute subsystems are. We [71:36] basically take the 128 CPUs, we take the [71:40] coherent mesh network, other controllers [71:42] and memory interfaces, [71:44] and not only do we stitch them together, [71:46] but we also verify that this is all [71:49] going to functionally work and be [71:50] correct. That when you put it into your [71:52] design, it's just going to work. And [71:55] that can save 3 months, 6 months, 9 [71:58] months of engineering time. They can get [72:00] a product out to market a heck of a lot [72:01] sooner. We can take that a step deeper, [72:04] which we do, in terms of we may work [72:06] with a a TSMC, or a Samsung, or an [72:08] Intel, [72:09] and say, "We're going to actually now [72:11] say, if you build it this way, [72:14] with these type of characteristics, we [72:16] will guarantee that you will get 4.4 GHz [72:19] of frequency output. We know that you [72:21] can get this kind of performance." So, [72:23] we are taking it much further than we [72:25] have. It's It's almost, I would say, a a [72:28] virtual [72:29] chipset, but not quite to the final [72:31] building a chip, but it's it's pretty [72:32] darn close. And then you say, "Well, [72:34] well, why why would you do that?" Yeah, [72:36] this is a lot of integration. I mean, [72:37] it's a lot of bundling from just the [72:39] instruction set architecture to [72:41] uh [72:42] design to now this kind of complete [72:45] solution. Hey, connect it all this way, [72:47] and, you know, 4.4 GHz are yours. Yep. [72:49] I'll call it packaging instead of [72:51] bundling, but it is a way of [72:53] It's a way of providing a full solution [72:56] that will simply allow customers to get [72:58] to market a heck of a lot faster. It [73:00] provides us a lot of benefit, because we [73:01] can do early prototyping from a software [73:03] standpoint earlier. But for customers, [73:05] the big benefit is they get to market [73:07] much faster than they would. And back to [73:09] the IP standpoint, [73:11] connecting up all the CPUs, taking the [73:12] IP that we deliver, that's not really [73:15] value add from [73:18] an end customer, right? An end customer [73:19] that's building a phone chip wants to [73:22] focus on the ISP in the camera. If [73:24] you're a cloud customer, you may want to [73:26] focus on the accelerator, or something [73:28] on analog IO. [73:30] So, for us, our position is, if it's [73:32] around the computer, [73:34] and essentially what's running the main [73:36] software of the system, and how that [73:37] performs in a certain fab, we're [73:39] probably in the best position to be able [73:42] to define what the best performance [73:43] output will look like. [73:44] All right. So, you now have this [73:46] essentially reference design for how to [73:48] make an amazing chip. Are we ever going [73:50] to see arm call up TSMC and say, "Hey, [73:54] go make a few million of these?" [73:56] Nothing I can say about that today. [73:59] Fair enough. [73:59] Great. Great. Well, Rene, this has been [74:01] awesome. Thank you so much. That was [74:03] great. Thank you. Awesome. Listeners, [74:05] we'll see you next time. [74:06] We'll see you next time.