{
  "text": "一颗采用12纳米工艺的芯片\n竟然在人工智能推理\n测试里压过了4纳米GPU\n它没有最先进的制程\n没有昂贵的HBM\n也不能训练大模型\n却在某些任务上跑得更快\n耗电更少\n听起来像IBM突然找到了\n击败英伟达的捷径\n但真正的问题是\n如果north pole这么强\n为什么它到现在仍然\n只是一款研究原型\n答案藏在它最强的优势里\n也藏在它最难绕开的限制里\n这颗芯片叫north pole\nIBM在2023年正式公开它时\n给出的规格并\n不符合当时的主流叙事\n12纳米工艺\n220亿个晶体管\n256个核心\n芯片面积约795平方毫米\n放在一个人人都盯着\n4纳米3纳米的行业里\n12纳米看起来甚至有些落后\n可IBM偏偏用这颗旧制程芯片做出\n了一组让人无法忽略的数据\n在Resnet50图像分类和YOLO\nV4目标检测测试中\nnorth pole相对采用相近12纳米工艺的GPU\n实现约25倍的每瓦增速\n约5倍的每晶体管处理效率\n以及约22倍更低的延迟\nIBM还表示\n它在这些限定任务上\n连采用更先进4纳米\n工艺的GPU也能压过去\n先纠正一个流传很广的说法\n22倍并不是每一单位\n能量多完成22倍工作\nIBM公开论文里的22倍指的是\n相对可比12纳米GPU的延迟优势\n能效指标是约25倍\n把延迟和能效混成\n一句话会让数字更刺激\n却也会误导观众\nnorth pole确实很快很省电\n但正确理解这些成绩比\n重复一个夸张数字更重要\n对north pole更重要的是\n制程从来不是芯片表现的唯一答案\n先进制程可以让晶体管更小\n通常也有利于密度性能和功耗\n但如果一套架构把大量\n电力耗在搬运数据上\n再先进的晶体管也只能缓解问题\n不能让问题消失\nnorth pole真正供给的\n正是现代AI计算最昂贵\n也最不容易被普通\n用户看见的一部分\n数据移动\n传统计算机长期遵循冯诺依曼架构\n处理器负责计算\n内存负责存数据\n两者分开\n处理器需要一组权重\n就去内存取\n算完产生结果\n再把结果送回去\n这个循环每秒发生无数次\n早期软件的计算量有限\n这种分工足够通用\n也足够灵活\n但神经网络把问题放大了\nAI推理里的单次运算并不神秘\n大量工作只是乘法和加法\n真正可怕的是数量\n以及每一次运算都要\n及时拿到正确数据\nGPU的计算能力增长很快\n内存带宽却没有以相同速度增长\n于是计算单元越来越像\n一座扩建过渡的工厂\n机器很多\n原料却还在路上\nGPU加入HBM\n就是给工厂修更宽的高速公路\n但高速公路昂贵\n封装复杂\n也会持续耗电\nnorth pole提出的不是把路再拓宽\n而是把仓库直接搬进工厂\nnorth pole把分布式SRAM放进芯片内部\n让每个核心旁边都有本地内存\n整颗芯片共有192兆分布式SRAM\n片上内存带宽达到约13太字节每秒\n模型权重和中间数据\n尽量留在芯片内部\n只有输入和最终结果\n需要跨过芯片边界\nIBM负责人darmendramoda把它形容为\n一个完整网路装在一颗芯片上\n这句话听起来像宣传\n但很准确的概括了设计目标\n让计算靠近数据\n把搬运距离压到最低\n芯片内部的256个核心并不是一群\n等待中央调度器发号施令的通用\n工人\nNorthpoole会提前由\n编译器安排数据怎么走\n计算何时发生\n各个核心做哪一段任务\n运行时不需要不断判断下一步\n也没有复杂的数据依赖分支\n它牺牲动态调度\n换来更高利用率和更可预测的执行\n这里出现了第一个关键转折\nNorthpo并不是一块更好的通用GPU\n它是一块把自己限制\n的很死的推理芯片\n它不支持大模型训练\n训练需要前向计算反向传播\n梯度更新和频繁变化的数据流\n需要高度灵活的内存与调度\nNorthpole从设计之初就放弃这些能力\n只专注神经网络的前向推理\n这条路线并不是IBM突然想到的\nnorth pole之前还有true north\nIBM团队从2,000年代\n开始研究类脑计算得到过\n美国国防高级研究计划局Synaps项目\n支持\nTruenorth在2014年公开\n强调用大量低功耗\n神经元核心处理感知任务\n它展示了类脑芯片可以极省电\n却也受限于编程方式和适用模型\nNorthpoole继承了让数据待\n在计算附近的思想\n但没有照搬生物神经元\n它转向更适合现代深度\n网路的数字计算和张量操作\n从true north到Northpoole\n真正变化的是目标\nTruenorth更像在探索机器能否用\n接近大脑的稀疏方式处理感知\nNorthpole则直接瞄准\n部署中的效率问题\n它不要求开发者接受\n一种完全陌生的神经网络\n而是尝试运行常见卷积\n网络和经过适配的Transformer\n类脑灵感保留下来\n工程路径却更接近现代AI加速器\n这段历史解释了为什么Northpoole\n很难用普通显卡思路评价\n他不是把更多单元塞进芯片\n也不是单纯提高频率\n他先设定10条架构原则\n再围绕推理重新组织芯片\n没有运行时调度\n没有外部内存依赖\n控制分散到各个核心\n每条原则都在减少通用性\n同时压低数据移动和控制开销\n这个取舍并不愚蠢\n一个模型可能只训练几轮\n却要被调用几百万几亿次\n训练决定模型怎么形成\n推理决定它一生消耗多少电\n如果某种架构能把\n高频推理做到极致\n即使训练仍然交给GPU\n它也可能在AI基础设施\n里占据非常大的位置\nnorth pole押注的正是这件事\n不去争夺所有工作\n只吃下重复次数最多的那一段\nIBM后来把测试从图像\n模型推进到语言模型\n研究团队从Granit Bobby Codebase\n改造出一个约30亿参数\n14层Transformer的模型\n再把14个Transformer层分别\n映射到14张north pole卡\n输出层放到另外两张卡上\n16张卡通过Pcie组成一台普通2U服务器\n以流水线方式执行推理\n测试结果很亮眼\nnorth pole把单TOKEN延迟压到1毫秒以内\n相对对照组中能源\n效率最高的GPU快46.9倍\n同时相对延迟最低的GPU节能72.7倍\n16卡系统总吞吐量\n达到每秒28,356个TOKEN\n这组数字同样需要仔细读\n46.9倍和72.7倍不是对同一张GPU\n同一个指标进行的全面碾压\nIBM分别选择了最节能的\n对照GPU和延迟最低的对照GPU\n用来展示north pole\n同时避开了速度与能效的传统取舍\n这依然是非常强的结果\n但不能简化成比\n所有英伟达GPU快70倍\n测试模型精度批次上下文\n长度和系统配置都会影响比较\n随后IBM把规模进一步扩大\n2025年公开的一套端到端研究\n系统使用288张north pole加速卡\n装进18台2U服务器\n整个42U机架提供115 PETAOPS的四位整数\n计算能力和3.7 PBS聚合内存带宽\n功耗约30千瓦\n它能同时运行三套Granite3.38B模型\n在2048上下文长度下\n服务28名并发用户\n单用户TOKEN间延迟约2.8毫秒\n30千瓦对普通家庭当然是惊人数字\n但放进AI数据中心\n它的意义完全不同\n今天的高密度GPU机架常常逼迫\n数据中心升级供电和液冷\nnorth pole系统能够在空气冷却\n和常规基础设施下运行\n说明它追求的不只是\n实验室里的单芯片纪录\n而是让推理系统在现有机房里落地\n对企业而言\n省下的不只是电费\n还有机房改造散热和供电扩容\n可是看到这里\n最大的疑问反而更明显\n既然数据这么漂亮\n为什么IBM没有直接\n把north pole推向市场\n与英伟达争夺AI加速卡订单\n第一个答案是内存容量\nnorth pole每颗芯片只有192兆分布式SRAM\n片上内存快得惊人\n也贵得惊人\n容量无法像外部DRAM\n或HBM那样轻易堆高\n任何映射到单张卡上的权重和\n中间激活都必须挤进这192兆\n模型一大\n就只能切层\n增加卡片数量\n再用流水线串起来\n这也是为什么30亿\n参数模型需要16张卡\n80亿参数模型扩展到数百张卡\nnorth pole用更多芯片换取\n模型完整驻留片上的速度\n对于固定模型\n稳定流量和极低延迟任务\n这种交易可能划算\n对于模型经常变化\n上下文不断增长\n或者负载忽高忽低的环境\n它未必比GPU更经济\n第二个答案是上下文\n大语言模型推理不只要存权重\n还要保存不断增长的KV缓存\n对话越长\n缓存越大\nnorth pole的片上内存架构天生\n不喜欢无限扩张的上下文\n288卡系统展示的Granit 8B\n运行在2048上下文长度下\n而公开版本Granit模型\n本身能够支持更长上下文\n这说明实验更接近低\n延迟可控长度的企业任务\n而不是让用户一次塞进整套\n代码库或几十万TOKEN文档\n第三个答案是模型适配成本\n开发者不能从网上下载任意模型\n直接丢进north pole\n模型需要量化到较低精度\n重新训练或微调\n随后由专用工具完成分层和映射\nNorthpoole没有消灭GPU\n恰恰相反\nGPU仍然负责训练调整和准备模型\nNorthpoole只在最后的\n部署阶段发挥优势\n它适合模型已经稳定\n任务会被重复执行很多次的场景\n例如视觉识别工业检测时事决策\n固定企业助手和部分代理流程\n它不适合每天更换模型\n频繁实验\n需要训练和推理混跑的研发环境\nGPU最大的护城河\n从来不只是单向性能\n而是通用性和CUDA生态\n开发者可以用同一套\n硬件训练微调推理\n也可以运行大量非AI计算\nnorth pole则要求用户\n围绕它重新设计流程\n低精度同样是成绩的一部分\nNorthpole在8位4位甚至更低\n精度下追求高吞吐\n语言模型实验也需要量化\n把权重从16位压到4位\n可以显著减少存储和计算\n但需要重新校准\n避免准确率明显下降\n硬件成绩不能脱离模型质量看\n若一套系统快10倍\n却在关键任务上经常答错\n速度没有商业意义\nIBM的研究说明\n在选定模型和精度下\n可以维持目标质量\n却不能自动代表所有\n模型都能无损迁移\n流水线还带来另一种代价\n模型每一层固定落在某张卡上\nTOKEN像产品一样依次通过生产线\n系统总吞吐量可以很高\n可单个请求的表现\n会受最慢阶段限制\n若某层计算更重\n或者请求长度差异很大\n负载平衡就会变复杂\nGPU集群也有通信开销\n但软件生态已经积累了成熟的张量\n并行流水线并行和动态批处理工具\nNorthpoole要商业化\n必须证明自己的\n编译器和运行时同样可靠\n企业部署还会追问容错\n一套288卡系统里\n只要一张卡或一条链路出现故障\n模型映射如何重新安排\n升级模型时是否要\n重新编译整条流水线\n不同客户的模型能不能共享硬件\n这些问题不会出现\n在单项芯片跑分里\n却决定数据中心是否\n敢把生产流量交给他\n研究论文证明了可行\n产品团队还要把故障\n监控和更新变成日常操作\n第四个答案是基准本身\nNorthpo在Resnet50UOV4和经过专门\n改造的Granite模型上表现突出\n不等于它能在所有AI\n工作负载中领先\n现代前沿模型大量使用混合\n专家长上下文稀疏注意力多\n模态输入和动态工具调用\n每种结构对内存通信\n和控制的需求都不同\n预先安排好的数据流\n能够消除调度开销\n也会降低应对动态变化的能力\n所以标题里12纳米让\n4纳米GPU相形见绌\n只在一个严格范围内成立\n它证明的是\n在特定神经网络推理任务中\n架构优势可以大到抵消制程差距\n它没有证明12纳米普遍胜过4纳米\n更没有证明IBM已经做出一张\n能够替代英伟达数据中心GPU的\n产品north pole目前仍是研究原型\nIBM真正推向商业\n市场的AI加速器叫Spire\nSpire采用5纳米工艺\n拥有32个加速核心和256亿个晶体管\n安装在75瓦Pcie卡上\n已经进入IBM z Linux one和power系统\n它面向企业本地推理\n生成式AI\n和代理任务\n强调数据留在企业系统内部\nnorth Pole与Spire来自\nIBM长期AI硬件研究\n但不是同一款芯片\n也不能把Spire的商业化\n算成north Pole已经上市\n这个区别反而揭示了实验室\n记录与商业产品之间的距离\n研究芯片可以把一个想法推到极致\n商业芯片还要考虑良率成本软件\n兼容性维护客户迁移和长期供应\nnorth pole最重要的任务\n也许不是成为下一张热门加速卡\n而是证明一条路线\n值得产业继续投入\n这条路线已经集中了\n英伟达最敏感的问题\n英伟达现在的优势建立在GPU\nCUDA网路HBM和整机系统上\n他卖的不是一颗芯片\n而是一整座AI工厂\n但随着推理占总算\n力需求的比例上升\n客户会越来越认真地计算\n每个TOKEN的电力延迟和成本\n只要工作负载足够稳定\n专用芯片就有机会从GPU\n手里切走一部分任务Google TPU\nAmazon Trion和Inferentia\n微软Maya以及各种边缘NPU\n本质上都在做类似的选择\n不是所有AI计算都必须\n交给最通用最昂贵的GPU\nNorthpointe把这个选择推得更远\n它不是简单增加矩阵乘法单元\n而是重新安排内存与计算的关系\n把少搬数据放在多做计算之前\nAI行业过去几年的扩张方式很直接\n模型变大就买更多GPU\n上下文变长就上更多HBM\n机架功率增加就修更大的数据中心\n这个办法有效\n却越来越昂贵\n电网接入时间液冷系统HBM供应\n和先进封装都开始成为限制\nnorth pole提醒行业\n有些增长不必来自更大的蛮力\n也可以来自不做无用功\n这里还藏着一个经常\n被忽略的经济问题\n数据中心买GPU时\n采购价格只是一部分成本\n芯片几年生命周期里\n电力冷却机房空间和\n利用率会持续产生支出\n一块便宜但经常等待\n内存的芯片未必真的便宜\n一套前期适配复杂\n长期每次推理都更省电的专用系统\n在高调用量场景里可能更有优势\nnorth pole的商业价值要\n用总拥有成本计算\n而不是只看晶圆制程\n可总拥有成本不能只算电\n专用芯片的开发工具模型转换\n工程师培训备件和供应规模都要\n花钱\n英伟达GPU贵却很容易找到开发者\n框架和现成方案企业遇到问题也能\n从庞大社区和合作伙伴中找到答案\nnorth pole若想把能效优势变成订单\n节省的电费必须大到\n足以覆盖迁移成本\n还要保证芯片能持续供货\n这也是先进制程比较\n最容易误导的地方\n4纳米GPU承担的是更广泛任务\n包含缓存调度图形和通用计算能力\n12纳米north pole删除了\n大量不需要的能力\n两者像一辆能在各种道路行驶的\n卡车和一列只跑固定路线的电车\n电车在自己的轨道上效率极高\n不能据此断言卡车技术落后\n正确结论是\n当路线足够固定\n专用设计可以用成熟制程\n赢过先进制程的通用芯片\n成熟制程还有现实优势\n晶圆成本更低\n产能更充足\n制造经验更稳定\nAI行业把注意力集中在\n台积电最先进节点\n但并非所有推理芯片都\n需要增强最稀缺产能\n如果架构能在12纳米实现目标\n同样预算可以购买更多芯片\n也能降低供应链对\n单一先进节点的依赖\nnorth pole给出的不只是技术路线\n也是一种供应链选择\n但反过来\n专用系统也有被\n模型变化淘汰的风险\n如果模型架构每半年就大改一次\n今天为某种数据流设计的硬件\n明天可能难以适配\nGPU虽然没有在每项任务上做到极致\n却能承受行业快速变化\n客户为通用性支付的溢价\n实际上是在购买\n未来的不确定性保险\nNorthpo与英伟达\n因此不是简单的谁把谁打败\n它们代表两种不同答案\nGPU的答案是用强大的通用并行计算\n外部高带宽内存和成熟软件生态\n覆盖尽可能多的任务\nnorth pole的答案是主动\n放弃训练和部分灵活性\n把固定推理路径压缩到片上\n用结构换效率\n未来最可能出现的不是\n专用芯片一夜取代GPU\n而是AI基础设施开始分层\n训练和快速实验继续留在GPU上\n稳定的大规模推理转向\nTPU ASIC或north pole式架构\n手机汽车和工业设备\n使用更小的边缘NPU\n工作负载被拆开以后\n英伟达仍然可能是最大赢家\n却不再天然拥有每一个环节\nnorth pole最值得记住的数字\n也许不是22倍26倍46.9倍或者72.7倍\n而是192兆\n就是这块看起来不大的片上内存\n让它绕过数据搬运\n赢得惊人的速度和能效\n也正是这192招\n限制了模型大小上下文和灵活性\n它的优势和代价\n来自同一个设计决定\n这也回答了开头的问题\nNorthpole没有成为商品\n不是因为实验失败\n而是因为它证明的太专一\n它证明了12纳米可以在特定\n推理任务上击败4纳米GPU\n证明了内存墙并非不可改变\n也证明了AI硬件还有一条\n不靠无限增加功耗的路\n可要把这条路变成普遍适用的产品\nIBM还得解决容量软件和模型适配\n对英伟达而言\n真正的威胁不是IBM明天\n推出一张north pole显卡\n而是越来越多客户\n开始问同一个问题\n既然大部分电力花在搬数据上\n为什么还要为每次推理\n都使用最昂贵的通用GPU\n只要这个问题被持续追问\nAI芯片战争就会从制程和算力转向\n内存位置数据流和每个TOKEN的成本\n所以\n12纳米并没有推翻4纳米\nnorth pole推翻的是\n另一种更顽固的迷信\n只要制程更新\nGPU更大\n功耗更高\nAI就一定会更有效率\nIBM用一颗研究芯片说明\n架构仍然可以改变规则\n英伟达今天不会因此失去王座\n但它必须面对一个更难防守的战场\n在这个战场上\n竞争对手不需要做出更大的GPU\n只需要让大量任务不再需要GPU\n试试核验\nIBM Northpole采用12纳米工艺\n拥有220亿个晶体管与256个核心属实\n分布式DRAM容量为192兆\n片上带宽约13太字节每秒\n属实在限定视觉推理任务\n中超过一款4纳米GPU\n属实相对可比12纳米GPU约25倍\n能效和约22倍低延迟属实\n每焦耳工作量提高\n22倍的说法不准确\n16卡系统达到每秒28,356TOKEN属实\n288卡研究系统与30千瓦\n功耗数据有论文支持\nNorthpole全面击败现代英伟达GPU\n不属实Northpole已经商业化不属实\nIBM已经商业化\n另一款Spire加速器属实",
  "language": "zh",
  "duration": 1392.0,
  "segments": [
    {
      "id": 0,
      "start": 0.133,
      "end": 2.533,
      "text": "一颗采用12纳米工艺的芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 1,
      "start": 2.566,
      "end": 4.2,
      "text": "竟然在人工智能推理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 2,
      "start": 4.2,
      "end": 6.5,
      "text": "测试里压过了4纳米GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 3,
      "start": 6.666,
      "end": 8.366,
      "text": "它没有最先进的制程",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 4,
      "start": 8.6,
      "end": 10.366,
      "text": "没有昂贵的HBM",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 5,
      "start": 10.4,
      "end": 12.166,
      "text": "也不能训练大模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 6,
      "start": 12.266,
      "end": 14.533,
      "text": "却在某些任务上跑得更快",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 7,
      "start": 14.666,
      "end": 15.7,
      "text": "耗电更少",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 8,
      "start": 15.766,
      "end": 17.866,
      "text": "听起来像IBM突然找到了",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 9,
      "start": 17.866,
      "end": 19.4,
      "text": "击败英伟达的捷径",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 10,
      "start": 19.866,
      "end": 21.366,
      "text": "但真正的问题是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 11,
      "start": 21.366,
      "end": 23.1,
      "text": "如果north pole这么强",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 12,
      "start": 23.2,
      "end": 24.933,
      "text": "为什么它到现在仍然",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 13,
      "start": 24.933,
      "end": 26.7,
      "text": "只是一款研究原型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 14,
      "start": 26.866,
      "end": 29.0,
      "text": "答案藏在它最强的优势里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 15,
      "start": 29.0,
      "end": 31.533,
      "text": "也藏在它最难绕开的限制里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 16,
      "start": 31.533,
      "end": 33.4,
      "text": "这颗芯片叫north pole",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 17,
      "start": 33.966,
      "end": 37.3,
      "text": "IBM在2023年正式公开它时",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 18,
      "start": 37.333,
      "end": 38.533,
      "text": "给出的规格并",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 19,
      "start": 38.533,
      "end": 40.5,
      "text": "不符合当时的主流叙事",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 20,
      "start": 40.8,
      "end": 42.1,
      "text": "12纳米工艺",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 21,
      "start": 42.3,
      "end": 44.133,
      "text": "220亿个晶体管",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 22,
      "start": 44.3,
      "end": 45.9,
      "text": "256个核心",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 23,
      "start": 46.2,
      "end": 49.266,
      "text": "芯片面积约795平方毫米",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 24,
      "start": 49.4,
      "end": 50.866,
      "text": "放在一个人人都盯着",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 25,
      "start": 50.866,
      "end": 53.266,
      "text": "4纳米3纳米的行业里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 26,
      "start": 53.566,
      "end": 55.966,
      "text": "12纳米看起来甚至有些落后",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 27,
      "start": 56.066,
      "end": 59.0,
      "text": "可IBM偏偏用这颗旧制程芯片做出",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 28,
      "start": 59.0,
      "end": 61.333,
      "text": "了一组让人无法忽略的数据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 29,
      "start": 61.4,
      "end": 64.333,
      "text": "在Resnet50图像分类和YOLO",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 30,
      "start": 64.333,
      "end": 66.2,
      "text": "V4目标检测测试中",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 31,
      "start": 66.533,
      "end": 70.133,
      "text": "north pole相对采用相近12纳米工艺的GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 32,
      "start": 70.533,
      "end": 72.9,
      "text": "实现约25倍的每瓦增速",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 33,
      "start": 73.133,
      "end": 75.566,
      "text": "约5倍的每晶体管处理效率",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 34,
      "start": 75.733,
      "end": 78.1,
      "text": "以及约22倍更低的延迟",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 35,
      "start": 78.333,
      "end": 79.666,
      "text": "IBM还表示",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 36,
      "start": 79.7,
      "end": 81.533,
      "text": "它在这些限定任务上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 37,
      "start": 81.533,
      "end": 83.466,
      "text": "连采用更先进4纳米",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 38,
      "start": 83.466,
      "end": 85.666,
      "text": "工艺的GPU也能压过去",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 39,
      "start": 85.7,
      "end": 87.8,
      "text": "先纠正一个流传很广的说法",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 40,
      "start": 88.166,
      "end": 90.266,
      "text": "22倍并不是每一单位",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 41,
      "start": 90.266,
      "end": 92.533,
      "text": "能量多完成22倍工作",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 42,
      "start": 92.6,
      "end": 95.966,
      "text": "IBM公开论文里的22倍指的是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 43,
      "start": 95.966,
      "end": 99.1,
      "text": "相对可比12纳米GPU的延迟优势",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 44,
      "start": 99.333,
      "end": 101.533,
      "text": "能效指标是约25倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 45,
      "start": 101.6,
      "end": 103.366,
      "text": "把延迟和能效混成",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 46,
      "start": 103.366,
      "end": 105.533,
      "text": "一句话会让数字更刺激",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 47,
      "start": 105.766,
      "end": 107.466,
      "text": "却也会误导观众",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 48,
      "start": 107.933,
      "end": 110.066,
      "text": "north pole确实很快很省电",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 49,
      "start": 110.133,
      "end": 112.4,
      "text": "但正确理解这些成绩比",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 50,
      "start": 112.4,
      "end": 114.8,
      "text": "重复一个夸张数字更重要",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 51,
      "start": 114.933,
      "end": 116.866,
      "text": "对north pole更重要的是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 52,
      "start": 117.266,
      "end": 120.1,
      "text": "制程从来不是芯片表现的唯一答案",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 53,
      "start": 120.5,
      "end": 123.066,
      "text": "先进制程可以让晶体管更小",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 54,
      "start": 123.166,
      "end": 125.9,
      "text": "通常也有利于密度性能和功耗",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 55,
      "start": 126.1,
      "end": 128.0,
      "text": "但如果一套架构把大量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 56,
      "start": 128.0,
      "end": 130.1,
      "text": "电力耗在搬运数据上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 57,
      "start": 130.266,
      "end": 133.133,
      "text": "再先进的晶体管也只能缓解问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 58,
      "start": 133.266,
      "end": 134.8,
      "text": "不能让问题消失",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 59,
      "start": 135.2,
      "end": 137.166,
      "text": "north pole真正供给的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 60,
      "start": 137.366,
      "end": 140.066,
      "text": "正是现代AI计算最昂贵",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 61,
      "start": 140.366,
      "end": 141.9,
      "text": "也最不容易被普通",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 62,
      "start": 141.9,
      "end": 143.4,
      "text": "用户看见的一部分",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 63,
      "start": 143.666,
      "end": 144.733,
      "text": "数据移动",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 64,
      "start": 145.0,
      "end": 148.133,
      "text": "传统计算机长期遵循冯诺依曼架构",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 65,
      "start": 148.466,
      "end": 149.933,
      "text": "处理器负责计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 66,
      "start": 150.066,
      "end": 151.733,
      "text": "内存负责存数据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 67,
      "start": 151.866,
      "end": 153.066,
      "text": "两者分开",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 68,
      "start": 153.2,
      "end": 154.966,
      "text": "处理器需要一组权重",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 69,
      "start": 155.166,
      "end": 156.4,
      "text": "就去内存取",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 70,
      "start": 156.566,
      "end": 157.9,
      "text": "算完产生结果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 71,
      "start": 157.9,
      "end": 159.333,
      "text": "再把结果送回去",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 72,
      "start": 159.533,
      "end": 161.8,
      "text": "这个循环每秒发生无数次",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 73,
      "start": 161.9,
      "end": 164.133,
      "text": "早期软件的计算量有限",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 74,
      "start": 164.366,
      "end": 166.266,
      "text": "这种分工足够通用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 75,
      "start": 166.266,
      "end": 167.5,
      "text": "也足够灵活",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 76,
      "start": 167.933,
      "end": 170.366,
      "text": "但神经网络把问题放大了",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 77,
      "start": 170.7,
      "end": 173.566,
      "text": "AI推理里的单次运算并不神秘",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 78,
      "start": 174.1,
      "end": 176.566,
      "text": "大量工作只是乘法和加法",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 79,
      "start": 176.733,
      "end": 178.4,
      "text": "真正可怕的是数量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 80,
      "start": 178.966,
      "end": 180.533,
      "text": "以及每一次运算都要",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 81,
      "start": 180.533,
      "end": 182.133,
      "text": "及时拿到正确数据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 82,
      "start": 182.333,
      "end": 184.533,
      "text": "GPU的计算能力增长很快",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 83,
      "start": 184.933,
      "end": 187.6,
      "text": "内存带宽却没有以相同速度增长",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 84,
      "start": 187.766,
      "end": 189.7,
      "text": "于是计算单元越来越像",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 85,
      "start": 189.7,
      "end": 191.566,
      "text": "一座扩建过渡的工厂",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 86,
      "start": 191.766,
      "end": 192.8,
      "text": "机器很多",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 87,
      "start": 192.866,
      "end": 194.333,
      "text": "原料却还在路上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 88,
      "start": 194.666,
      "end": 196.266,
      "text": "GPU加入HBM",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 89,
      "start": 196.266,
      "end": 198.766,
      "text": "就是给工厂修更宽的高速公路",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 90,
      "start": 198.8,
      "end": 200.466,
      "text": "但高速公路昂贵",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 91,
      "start": 200.566,
      "end": 201.566,
      "text": "封装复杂",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 92,
      "start": 201.566,
      "end": 203.066,
      "text": "也会持续耗电",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 93,
      "start": 203.333,
      "end": 205.9,
      "text": "north pole提出的不是把路再拓宽",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 94,
      "start": 205.9,
      "end": 208.266,
      "text": "而是把仓库直接搬进工厂",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 95,
      "start": 208.566,
      "end": 211.666,
      "text": "north pole把分布式SRAM放进芯片内部",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 96,
      "start": 211.733,
      "end": 214.266,
      "text": "让每个核心旁边都有本地内存",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 97,
      "start": 214.666,
      "end": 218.066,
      "text": "整颗芯片共有192兆分布式SRAM",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 98,
      "start": 218.266,
      "end": 221.533,
      "text": "片上内存带宽达到约13太字节每秒",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 99,
      "start": 221.666,
      "end": 223.4,
      "text": "模型权重和中间数据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 100,
      "start": 223.4,
      "end": 225.166,
      "text": "尽量留在芯片内部",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 101,
      "start": 225.2,
      "end": 226.933,
      "text": "只有输入和最终结果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 102,
      "start": 226.933,
      "end": 228.6,
      "text": "需要跨过芯片边界",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 103,
      "start": 229.133,
      "end": 231.966,
      "text": "IBM负责人darmendramoda把它形容为",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 104,
      "start": 231.966,
      "end": 234.5,
      "text": "一个完整网路装在一颗芯片上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 105,
      "start": 234.566,
      "end": 236.466,
      "text": "这句话听起来像宣传",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 106,
      "start": 236.866,
      "end": 239.133,
      "text": "但很准确的概括了设计目标",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 107,
      "start": 239.266,
      "end": 240.933,
      "text": "让计算靠近数据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 108,
      "start": 241.066,
      "end": 243.066,
      "text": "把搬运距离压到最低",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 109,
      "start": 243.266,
      "end": 246.733,
      "text": "芯片内部的256个核心并不是一群",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 110,
      "start": 246.733,
      "end": 249.366,
      "text": "等待中央调度器发号施令的通用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 111,
      "start": 249.366,
      "end": 250.166,
      "text": "工人",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 112,
      "start": 250.6,
      "end": 251.966,
      "text": "Northpoole会提前由",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 113,
      "start": 251.966,
      "end": 253.966,
      "text": "编译器安排数据怎么走",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 114,
      "start": 254.133,
      "end": 255.533,
      "text": "计算何时发生",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 115,
      "start": 255.566,
      "end": 257.566,
      "text": "各个核心做哪一段任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 116,
      "start": 257.766,
      "end": 260.7,
      "text": "运行时不需要不断判断下一步",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 117,
      "start": 260.733,
      "end": 263.066,
      "text": "也没有复杂的数据依赖分支",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 118,
      "start": 263.3,
      "end": 264.8,
      "text": "它牺牲动态调度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 119,
      "start": 264.966,
      "end": 268.1,
      "text": "换来更高利用率和更可预测的执行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 120,
      "start": 268.166,
      "end": 270.166,
      "text": "这里出现了第一个关键转折",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 121,
      "start": 271.066,
      "end": 274.0,
      "text": "Northpo并不是一块更好的通用GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 122,
      "start": 274.2,
      "end": 275.666,
      "text": "它是一块把自己限制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 123,
      "start": 275.666,
      "end": 277.266,
      "text": "的很死的推理芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 124,
      "start": 277.6,
      "end": 279.466,
      "text": "它不支持大模型训练",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 125,
      "start": 279.7,
      "end": 282.4,
      "text": "训练需要前向计算反向传播",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 126,
      "start": 282.5,
      "end": 285.1,
      "text": "梯度更新和频繁变化的数据流",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 127,
      "start": 285.266,
      "end": 287.666,
      "text": "需要高度灵活的内存与调度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 128,
      "start": 288.333,
      "end": 291.366,
      "text": "Northpole从设计之初就放弃这些能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 129,
      "start": 291.5,
      "end": 293.966,
      "text": "只专注神经网络的前向推理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 130,
      "start": 294.0,
      "end": 296.566,
      "text": "这条路线并不是IBM突然想到的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 131,
      "start": 297.3,
      "end": 299.533,
      "text": "north pole之前还有true north",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 132,
      "start": 299.766,
      "end": 301.733,
      "text": "IBM团队从2,000年代",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 133,
      "start": 301.733,
      "end": 304.3,
      "text": "开始研究类脑计算得到过",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 134,
      "start": 304.3,
      "end": 307.866,
      "text": "美国国防高级研究计划局Synaps项目",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 135,
      "start": 307.866,
      "end": 308.666,
      "text": "支持",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 136,
      "start": 309.0,
      "end": 311.266,
      "text": "Truenorth在2014年公开",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 137,
      "start": 311.366,
      "end": 313.0,
      "text": "强调用大量低功耗",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 138,
      "start": 313.0,
      "end": 315.166,
      "text": "神经元核心处理感知任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 139,
      "start": 315.3,
      "end": 318.0,
      "text": "它展示了类脑芯片可以极省电",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 140,
      "start": 318.366,
      "end": 321.5,
      "text": "却也受限于编程方式和适用模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 141,
      "start": 322.4,
      "end": 324.366,
      "text": "Northpoole继承了让数据待",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 142,
      "start": 324.366,
      "end": 326.066,
      "text": "在计算附近的思想",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 143,
      "start": 326.366,
      "end": 328.466,
      "text": "但没有照搬生物神经元",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 144,
      "start": 328.733,
      "end": 330.666,
      "text": "它转向更适合现代深度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 145,
      "start": 330.666,
      "end": 333.333,
      "text": "网路的数字计算和张量操作",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 146,
      "start": 333.466,
      "end": 335.533,
      "text": "从true north到Northpoole",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 147,
      "start": 335.6,
      "end": 337.4,
      "text": "真正变化的是目标",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 148,
      "start": 338.066,
      "end": 340.7,
      "text": "Truenorth更像在探索机器能否用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 149,
      "start": 340.7,
      "end": 343.266,
      "text": "接近大脑的稀疏方式处理感知",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 150,
      "start": 343.7,
      "end": 345.366,
      "text": "Northpole则直接瞄准",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 151,
      "start": 345.366,
      "end": 346.8,
      "text": "部署中的效率问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 152,
      "start": 346.9,
      "end": 348.366,
      "text": "它不要求开发者接受",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 153,
      "start": 348.366,
      "end": 350.4,
      "text": "一种完全陌生的神经网络",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 154,
      "start": 350.5,
      "end": 352.7,
      "text": "而是尝试运行常见卷积",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 155,
      "start": 352.7,
      "end": 354.966,
      "text": "网络和经过适配的Transformer",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 156,
      "start": 356.066,
      "end": 357.666,
      "text": "类脑灵感保留下来",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 157,
      "start": 357.7,
      "end": 361.1,
      "text": "工程路径却更接近现代AI加速器",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 158,
      "start": 361.133,
      "end": 363.366,
      "text": "这段历史解释了为什么Northpoole",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 159,
      "start": 363.366,
      "end": 365.566,
      "text": "很难用普通显卡思路评价",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 160,
      "start": 365.9,
      "end": 368.1,
      "text": "他不是把更多单元塞进芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 161,
      "start": 368.166,
      "end": 369.966,
      "text": "也不是单纯提高频率",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 162,
      "start": 370.133,
      "end": 372.266,
      "text": "他先设定10条架构原则",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 163,
      "start": 372.333,
      "end": 374.8,
      "text": "再围绕推理重新组织芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 164,
      "start": 374.9,
      "end": 376.366,
      "text": "没有运行时调度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 165,
      "start": 376.6,
      "end": 378.133,
      "text": "没有外部内存依赖",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 166,
      "start": 378.333,
      "end": 380.2,
      "text": "控制分散到各个核心",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 167,
      "start": 380.533,
      "end": 382.8,
      "text": "每条原则都在减少通用性",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 168,
      "start": 382.933,
      "end": 385.533,
      "text": "同时压低数据移动和控制开销",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 169,
      "start": 385.7,
      "end": 387.366,
      "text": "这个取舍并不愚蠢",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 170,
      "start": 387.566,
      "end": 389.666,
      "text": "一个模型可能只训练几轮",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 171,
      "start": 390.0,
      "end": 392.866,
      "text": "却要被调用几百万几亿次",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 172,
      "start": 393.066,
      "end": 395.0,
      "text": "训练决定模型怎么形成",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 173,
      "start": 395.266,
      "end": 397.733,
      "text": "推理决定它一生消耗多少电",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 174,
      "start": 397.766,
      "end": 399.133,
      "text": "如果某种架构能把",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 175,
      "start": 399.133,
      "end": 400.666,
      "text": "高频推理做到极致",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 176,
      "start": 401.0,
      "end": 403.166,
      "text": "即使训练仍然交给GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 177,
      "start": 403.2,
      "end": 405.1,
      "text": "它也可能在AI基础设施",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 178,
      "start": 405.1,
      "end": 407.0,
      "text": "里占据非常大的位置",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 179,
      "start": 407.266,
      "end": 409.533,
      "text": "north pole押注的正是这件事",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 180,
      "start": 410.066,
      "end": 411.8,
      "text": "不去争夺所有工作",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 181,
      "start": 411.966,
      "end": 414.933,
      "text": "只吃下重复次数最多的那一段",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 182,
      "start": 415.066,
      "end": 417.866,
      "text": "IBM后来把测试从图像",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 183,
      "start": 417.866,
      "end": 419.733,
      "text": "模型推进到语言模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 184,
      "start": 420.166,
      "end": 422.4,
      "text": "研究团队从Granit Bobby Codebase",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 185,
      "start": 422.4,
      "end": 424.5,
      "text": "改造出一个约30亿参数",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 186,
      "start": 424.6,
      "end": 426.4,
      "text": "14层Transformer的模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 187,
      "start": 426.466,
      "end": 428.566,
      "text": "再把14个Transformer层分别",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 188,
      "start": 428.566,
      "end": 430.733,
      "text": "映射到14张north pole卡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 189,
      "start": 430.933,
      "end": 433.3,
      "text": "输出层放到另外两张卡上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 190,
      "start": 433.566,
      "end": 437.6,
      "text": "16张卡通过Pcie组成一台普通2U服务器",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 191,
      "start": 437.733,
      "end": 440.0,
      "text": "以流水线方式执行推理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 192,
      "start": 440.3,
      "end": 441.9,
      "text": "测试结果很亮眼",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 193,
      "start": 442.0,
      "end": 445.666,
      "text": "north pole把单TOKEN延迟压到1毫秒以内",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 194,
      "start": 445.766,
      "end": 447.566,
      "text": "相对对照组中能源",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 195,
      "start": 447.566,
      "end": 450.5,
      "text": "效率最高的GPU快46.9倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 196,
      "start": 450.933,
      "end": 455.533,
      "text": "同时相对延迟最低的GPU节能72.7倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 197,
      "start": 455.866,
      "end": 457.533,
      "text": "16卡系统总吞吐量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 198,
      "start": 457.533,
      "end": 460.3,
      "text": "达到每秒28,356个TOKEN",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 199,
      "start": 460.4,
      "end": 462.8,
      "text": "这组数字同样需要仔细读",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 200,
      "start": 463.366,
      "end": 467.9,
      "text": "46.9倍和72.7倍不是对同一张GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 201,
      "start": 468.066,
      "end": 470.466,
      "text": "同一个指标进行的全面碾压",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 202,
      "start": 470.766,
      "end": 473.133,
      "text": "IBM分别选择了最节能的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 203,
      "start": 473.133,
      "end": 476.6,
      "text": "对照GPU和延迟最低的对照GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 204,
      "start": 476.766,
      "end": 478.366,
      "text": "用来展示north pole",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 205,
      "start": 478.666,
      "end": 481.666,
      "text": "同时避开了速度与能效的传统取舍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 206,
      "start": 481.766,
      "end": 483.7,
      "text": "这依然是非常强的结果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 207,
      "start": 483.7,
      "end": 485.133,
      "text": "但不能简化成比",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 208,
      "start": 485.133,
      "end": 487.566,
      "text": "所有英伟达GPU快70倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 209,
      "start": 487.733,
      "end": 490.866,
      "text": "测试模型精度批次上下文",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 210,
      "start": 490.866,
      "end": 493.733,
      "text": "长度和系统配置都会影响比较",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 211,
      "start": 493.9,
      "end": 496.866,
      "text": "随后IBM把规模进一步扩大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 212,
      "start": 497.1,
      "end": 499.566,
      "text": "2025年公开的一套端到端研究",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 213,
      "start": 499.566,
      "end": 503.0,
      "text": "系统使用288张north pole加速卡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 214,
      "start": 503.0,
      "end": 505.1,
      "text": "装进18台2U服务器",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 215,
      "start": 505.4,
      "end": 509.566,
      "text": "整个42U机架提供115 PETAOPS的四位整数",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 216,
      "start": 509.566,
      "end": 513.0,
      "text": "计算能力和3.7 PBS聚合内存带宽",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 217,
      "start": 513.3,
      "end": 514.766,
      "text": "功耗约30千瓦",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 218,
      "start": 514.933,
      "end": 518.666,
      "text": "它能同时运行三套Granite3.38B模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 219,
      "start": 518.866,
      "end": 521.166,
      "text": "在2048上下文长度下",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 220,
      "start": 521.166,
      "end": 523.166,
      "text": "服务28名并发用户",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 221,
      "start": 523.6,
      "end": 526.6,
      "text": "单用户TOKEN间延迟约2.8毫秒",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 222,
      "start": 526.766,
      "end": 530.1,
      "text": "30千瓦对普通家庭当然是惊人数字",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 223,
      "start": 530.566,
      "end": 532.333,
      "text": "但放进AI数据中心",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 224,
      "start": 532.366,
      "end": 533.933,
      "text": "它的意义完全不同",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 225,
      "start": 534.066,
      "end": 536.9,
      "text": "今天的高密度GPU机架常常逼迫",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 226,
      "start": 536.9,
      "end": 539.133,
      "text": "数据中心升级供电和液冷",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 227,
      "start": 539.3,
      "end": 541.9,
      "text": "north pole系统能够在空气冷却",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 228,
      "start": 541.9,
      "end": 544.366,
      "text": "和常规基础设施下运行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 229,
      "start": 544.566,
      "end": 546.166,
      "text": "说明它追求的不只是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 230,
      "start": 546.166,
      "end": 548.266,
      "text": "实验室里的单芯片纪录",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 231,
      "start": 548.6,
      "end": 551.733,
      "text": "而是让推理系统在现有机房里落地",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 232,
      "start": 551.866,
      "end": 553.066,
      "text": "对企业而言",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 233,
      "start": 553.2,
      "end": 554.933,
      "text": "省下的不只是电费",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 234,
      "start": 554.933,
      "end": 557.9,
      "text": "还有机房改造散热和供电扩容",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 235,
      "start": 557.966,
      "end": 559.1,
      "text": "可是看到这里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 236,
      "start": 559.2,
      "end": 561.133,
      "text": "最大的疑问反而更明显",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 237,
      "start": 561.466,
      "end": 562.766,
      "text": "既然数据这么漂亮",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 238,
      "start": 562.9,
      "end": 564.466,
      "text": "为什么IBM没有直接",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 239,
      "start": 564.466,
      "end": 566.266,
      "text": "把north pole推向市场",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 240,
      "start": 566.266,
      "end": 568.866,
      "text": "与英伟达争夺AI加速卡订单",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 241,
      "start": 568.966,
      "end": 570.9,
      "text": "第一个答案是内存容量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 242,
      "start": 571.2,
      "end": 575.366,
      "text": "north pole每颗芯片只有192兆分布式SRAM",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 243,
      "start": 575.8,
      "end": 577.366,
      "text": "片上内存快得惊人",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 244,
      "start": 577.366,
      "end": 578.666,
      "text": "也贵得惊人",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 245,
      "start": 578.766,
      "end": 581.133,
      "text": "容量无法像外部DRAM",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 246,
      "start": 581.133,
      "end": 583.4,
      "text": "或HBM那样轻易堆高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 247,
      "start": 583.533,
      "end": 586.0,
      "text": "任何映射到单张卡上的权重和",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 248,
      "start": 586.0,
      "end": 589.533,
      "text": "中间激活都必须挤进这192兆",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 249,
      "start": 590.0,
      "end": 590.933,
      "text": "模型一大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 250,
      "start": 590.933,
      "end": 592.166,
      "text": "就只能切层",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 251,
      "start": 592.333,
      "end": 593.766,
      "text": "增加卡片数量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 252,
      "start": 593.8,
      "end": 595.6,
      "text": "再用流水线串起来",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 253,
      "start": 595.666,
      "end": 597.2,
      "text": "这也是为什么30亿",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 254,
      "start": 597.2,
      "end": 599.2,
      "text": "参数模型需要16张卡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 255,
      "start": 599.366,
      "end": 602.266,
      "text": "80亿参数模型扩展到数百张卡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 256,
      "start": 602.5,
      "end": 604.9,
      "text": "north pole用更多芯片换取",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 257,
      "start": 604.9,
      "end": 607.066,
      "text": "模型完整驻留片上的速度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 258,
      "start": 607.4,
      "end": 608.7,
      "text": "对于固定模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 259,
      "start": 608.8,
      "end": 611.266,
      "text": "稳定流量和极低延迟任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 260,
      "start": 611.333,
      "end": 613.133,
      "text": "这种交易可能划算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 261,
      "start": 613.266,
      "end": 615.0,
      "text": "对于模型经常变化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 262,
      "start": 615.1,
      "end": 616.566,
      "text": "上下文不断增长",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 263,
      "start": 616.566,
      "end": 618.766,
      "text": "或者负载忽高忽低的环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 264,
      "start": 618.9,
      "end": 620.966,
      "text": "它未必比GPU更经济",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 265,
      "start": 621.166,
      "end": 622.966,
      "text": "第二个答案是上下文",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 266,
      "start": 623.366,
      "end": 625.7,
      "text": "大语言模型推理不只要存权重",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 267,
      "start": 625.9,
      "end": 628.733,
      "text": "还要保存不断增长的KV缓存",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 268,
      "start": 629.0,
      "end": 629.966,
      "text": "对话越长",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 269,
      "start": 630.066,
      "end": 630.966,
      "text": "缓存越大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 270,
      "start": 631.333,
      "end": 633.666,
      "text": "north pole的片上内存架构天生",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 271,
      "start": 633.666,
      "end": 635.9,
      "text": "不喜欢无限扩张的上下文",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 272,
      "start": 636.1,
      "end": 639.2,
      "text": "288卡系统展示的Granit 8B",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 273,
      "start": 639.2,
      "end": 642.066,
      "text": "运行在2048上下文长度下",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 274,
      "start": 642.133,
      "end": 644.266,
      "text": "而公开版本Granit模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 275,
      "start": 644.266,
      "end": 646.8,
      "text": "本身能够支持更长上下文",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 276,
      "start": 646.9,
      "end": 648.766,
      "text": "这说明实验更接近低",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 277,
      "start": 648.766,
      "end": 651.166,
      "text": "延迟可控长度的企业任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 278,
      "start": 651.4,
      "end": 653.566,
      "text": "而不是让用户一次塞进整套",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 279,
      "start": 653.566,
      "end": 655.766,
      "text": "代码库或几十万TOKEN文档",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 280,
      "start": 655.966,
      "end": 658.466,
      "text": "第三个答案是模型适配成本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 281,
      "start": 658.966,
      "end": 661.466,
      "text": "开发者不能从网上下载任意模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 282,
      "start": 661.733,
      "end": 663.333,
      "text": "直接丢进north pole",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 283,
      "start": 663.566,
      "end": 665.8,
      "text": "模型需要量化到较低精度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 284,
      "start": 666.066,
      "end": 667.8,
      "text": "重新训练或微调",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 285,
      "start": 668.2,
      "end": 671.1,
      "text": "随后由专用工具完成分层和映射",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 286,
      "start": 671.5,
      "end": 673.333,
      "text": "Northpoole没有消灭GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 287,
      "start": 673.533,
      "end": 674.4,
      "text": "恰恰相反",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 288,
      "start": 674.666,
      "end": 677.933,
      "text": "GPU仍然负责训练调整和准备模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 289,
      "start": 678.366,
      "end": 679.7,
      "text": "Northpoole只在最后的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 290,
      "start": 679.7,
      "end": 681.466,
      "text": "部署阶段发挥优势",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 291,
      "start": 681.6,
      "end": 683.5,
      "text": "它适合模型已经稳定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 292,
      "start": 683.9,
      "end": 686.533,
      "text": "任务会被重复执行很多次的场景",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 293,
      "start": 687.266,
      "end": 691.2,
      "text": "例如视觉识别工业检测时事决策",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 294,
      "start": 691.5,
      "end": 694.133,
      "text": "固定企业助手和部分代理流程",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 295,
      "start": 694.333,
      "end": 696.2,
      "text": "它不适合每天更换模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 296,
      "start": 696.366,
      "end": 697.333,
      "text": "频繁实验",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 297,
      "start": 697.933,
      "end": 700.7,
      "text": "需要训练和推理混跑的研发环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 298,
      "start": 700.9,
      "end": 702.533,
      "text": "GPU最大的护城河",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 299,
      "start": 702.6,
      "end": 704.4,
      "text": "从来不只是单向性能",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 300,
      "start": 704.466,
      "end": 706.6,
      "text": "而是通用性和CUDA生态",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 301,
      "start": 706.9,
      "end": 708.333,
      "text": "开发者可以用同一套",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 302,
      "start": 708.333,
      "end": 710.5,
      "text": "硬件训练微调推理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 303,
      "start": 710.6,
      "end": 713.066,
      "text": "也可以运行大量非AI计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 304,
      "start": 713.3,
      "end": 715.333,
      "text": "north pole则要求用户",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 305,
      "start": 715.333,
      "end": 717.333,
      "text": "围绕它重新设计流程",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 306,
      "start": 717.5,
      "end": 719.566,
      "text": "低精度同样是成绩的一部分",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 307,
      "start": 720.3,
      "end": 722.9,
      "text": "Northpole在8位4位甚至更低",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 308,
      "start": 722.9,
      "end": 724.533,
      "text": "精度下追求高吞吐",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 309,
      "start": 725.1,
      "end": 727.333,
      "text": "语言模型实验也需要量化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 310,
      "start": 727.533,
      "end": 729.933,
      "text": "把权重从16位压到4位",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 311,
      "start": 730.2,
      "end": 732.7,
      "text": "可以显著减少存储和计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 312,
      "start": 732.7,
      "end": 734.333,
      "text": "但需要重新校准",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 313,
      "start": 734.5,
      "end": 736.4,
      "text": "避免准确率明显下降",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 314,
      "start": 736.666,
      "end": 739.166,
      "text": "硬件成绩不能脱离模型质量看",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 315,
      "start": 739.566,
      "end": 741.333,
      "text": "若一套系统快10倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 316,
      "start": 741.333,
      "end": 743.8,
      "text": "却在关键任务上经常答错",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 317,
      "start": 743.866,
      "end": 745.4,
      "text": "速度没有商业意义",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 318,
      "start": 745.6,
      "end": 747.066,
      "text": "IBM的研究说明",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 319,
      "start": 747.066,
      "end": 749.066,
      "text": "在选定模型和精度下",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 320,
      "start": 749.066,
      "end": 750.733,
      "text": "可以维持目标质量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 321,
      "start": 750.733,
      "end": 752.5,
      "text": "却不能自动代表所有",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 322,
      "start": 752.5,
      "end": 754.3,
      "text": "模型都能无损迁移",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 323,
      "start": 754.733,
      "end": 756.866,
      "text": "流水线还带来另一种代价",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 324,
      "start": 757.266,
      "end": 759.733,
      "text": "模型每一层固定落在某张卡上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 325,
      "start": 759.933,
      "end": 762.733,
      "text": "TOKEN像产品一样依次通过生产线",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 326,
      "start": 763.366,
      "end": 765.5,
      "text": "系统总吞吐量可以很高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 327,
      "start": 765.566,
      "end": 766.966,
      "text": "可单个请求的表现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 328,
      "start": 766.966,
      "end": 768.7,
      "text": "会受最慢阶段限制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 329,
      "start": 768.8,
      "end": 770.5,
      "text": "若某层计算更重",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 330,
      "start": 770.5,
      "end": 772.7,
      "text": "或者请求长度差异很大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 331,
      "start": 772.866,
      "end": 774.7,
      "text": "负载平衡就会变复杂",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 332,
      "start": 775.166,
      "end": 777.333,
      "text": "GPU集群也有通信开销",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 333,
      "start": 777.5,
      "end": 780.466,
      "text": "但软件生态已经积累了成熟的张量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 334,
      "start": 780.466,
      "end": 783.9,
      "text": "并行流水线并行和动态批处理工具",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 335,
      "start": 784.133,
      "end": 785.766,
      "text": "Northpoole要商业化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 336,
      "start": 785.866,
      "end": 787.0,
      "text": "必须证明自己的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 337,
      "start": 787.0,
      "end": 789.5,
      "text": "编译器和运行时同样可靠",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 338,
      "start": 789.7,
      "end": 791.7,
      "text": "企业部署还会追问容错",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 339,
      "start": 791.866,
      "end": 793.966,
      "text": "一套288卡系统里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 340,
      "start": 794.3,
      "end": 796.9,
      "text": "只要一张卡或一条链路出现故障",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 341,
      "start": 797.266,
      "end": 799.6,
      "text": "模型映射如何重新安排",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 342,
      "start": 799.866,
      "end": 801.533,
      "text": "升级模型时是否要",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 343,
      "start": 801.533,
      "end": 803.5,
      "text": "重新编译整条流水线",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 344,
      "start": 803.6,
      "end": 806.333,
      "text": "不同客户的模型能不能共享硬件",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 345,
      "start": 806.5,
      "end": 807.9,
      "text": "这些问题不会出现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 346,
      "start": 807.9,
      "end": 809.766,
      "text": "在单项芯片跑分里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 347,
      "start": 809.766,
      "end": 811.566,
      "text": "却决定数据中心是否",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 348,
      "start": 811.566,
      "end": 813.533,
      "text": "敢把生产流量交给他",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 349,
      "start": 813.933,
      "end": 815.6,
      "text": "研究论文证明了可行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 350,
      "start": 815.8,
      "end": 817.6,
      "text": "产品团队还要把故障",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 351,
      "start": 817.866,
      "end": 820.533,
      "text": "监控和更新变成日常操作",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 352,
      "start": 820.666,
      "end": 822.666,
      "text": "第四个答案是基准本身",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 353,
      "start": 823.4,
      "end": 827.2,
      "text": "Northpo在Resnet50UOV4和经过专门",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 354,
      "start": 827.2,
      "end": 829.7,
      "text": "改造的Granite模型上表现突出",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 355,
      "start": 829.733,
      "end": 831.466,
      "text": "不等于它能在所有AI",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 356,
      "start": 831.466,
      "end": 833.266,
      "text": "工作负载中领先",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 357,
      "start": 833.366,
      "end": 835.966,
      "text": "现代前沿模型大量使用混合",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 358,
      "start": 835.966,
      "end": 839.733,
      "text": "专家长上下文稀疏注意力多",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 359,
      "start": 839.733,
      "end": 842.166,
      "text": "模态输入和动态工具调用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 360,
      "start": 842.733,
      "end": 844.9,
      "text": "每种结构对内存通信",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 361,
      "start": 844.9,
      "end": 846.9,
      "text": "和控制的需求都不同",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 362,
      "start": 847.3,
      "end": 848.9,
      "text": "预先安排好的数据流",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 363,
      "start": 848.9,
      "end": 850.533,
      "text": "能够消除调度开销",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 364,
      "start": 850.733,
      "end": 853.366,
      "text": "也会降低应对动态变化的能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 365,
      "start": 853.666,
      "end": 855.733,
      "text": "所以标题里12纳米让",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 366,
      "start": 855.733,
      "end": 857.9,
      "text": "4纳米GPU相形见绌",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 367,
      "start": 858.166,
      "end": 860.3,
      "text": "只在一个严格范围内成立",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 368,
      "start": 860.566,
      "end": 861.5,
      "text": "它证明的是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 369,
      "start": 861.5,
      "end": 863.9,
      "text": "在特定神经网络推理任务中",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 370,
      "start": 864.0,
      "end": 866.933,
      "text": "架构优势可以大到抵消制程差距",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 371,
      "start": 867.066,
      "end": 870.133,
      "text": "它没有证明12纳米普遍胜过4纳米",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 372,
      "start": 870.466,
      "end": 872.9,
      "text": "更没有证明IBM已经做出一张",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 373,
      "start": 872.966,
      "end": 875.933,
      "text": "能够替代英伟达数据中心GPU的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 374,
      "start": 875.933,
      "end": 878.9,
      "text": "产品north pole目前仍是研究原型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 375,
      "start": 879.366,
      "end": 880.766,
      "text": "IBM真正推向商业",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 376,
      "start": 880.766,
      "end": 882.866,
      "text": "市场的AI加速器叫Spire",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 377,
      "start": 883.4,
      "end": 885.1,
      "text": "Spire采用5纳米工艺",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 378,
      "start": 885.333,
      "end": 889.466,
      "text": "拥有32个加速核心和256亿个晶体管",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 379,
      "start": 889.6,
      "end": 892.066,
      "text": "安装在75瓦Pcie卡上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 380,
      "start": 892.3,
      "end": 896.133,
      "text": "已经进入IBM z Linux one和power系统",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 381,
      "start": 896.2,
      "end": 897.933,
      "text": "它面向企业本地推理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 382,
      "start": 898.133,
      "end": 899.2,
      "text": "生成式AI",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 383,
      "start": 899.5,
      "end": 900.533,
      "text": "和代理任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 384,
      "start": 900.6,
      "end": 902.966,
      "text": "强调数据留在企业系统内部",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 385,
      "start": 903.2,
      "end": 904.966,
      "text": "north Pole与Spire来自",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 386,
      "start": 904.966,
      "end": 907.733,
      "text": "IBM长期AI硬件研究",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 387,
      "start": 907.933,
      "end": 909.666,
      "text": "但不是同一款芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 388,
      "start": 909.7,
      "end": 911.666,
      "text": "也不能把Spire的商业化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 389,
      "start": 911.666,
      "end": 913.666,
      "text": "算成north Pole已经上市",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 390,
      "start": 913.8,
      "end": 916.266,
      "text": "这个区别反而揭示了实验室",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 391,
      "start": 916.266,
      "end": 918.766,
      "text": "记录与商业产品之间的距离",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 392,
      "start": 919.066,
      "end": 922.133,
      "text": "研究芯片可以把一个想法推到极致",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 393,
      "start": 922.533,
      "end": 926.2,
      "text": "商业芯片还要考虑良率成本软件",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 394,
      "start": 926.333,
      "end": 930.066,
      "text": "兼容性维护客户迁移和长期供应",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 395,
      "start": 930.266,
      "end": 932.166,
      "text": "north pole最重要的任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 396,
      "start": 932.933,
      "end": 935.666,
      "text": "也许不是成为下一张热门加速卡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 397,
      "start": 935.766,
      "end": 937.333,
      "text": "而是证明一条路线",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 398,
      "start": 937.333,
      "end": 939.1,
      "text": "值得产业继续投入",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 399,
      "start": 939.166,
      "end": 940.733,
      "text": "这条路线已经集中了",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 400,
      "start": 940.733,
      "end": 942.4,
      "text": "英伟达最敏感的问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 401,
      "start": 942.766,
      "end": 945.466,
      "text": "英伟达现在的优势建立在GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 402,
      "start": 945.533,
      "end": 948.933,
      "text": "CUDA网路HBM和整机系统上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 403,
      "start": 949.266,
      "end": 950.866,
      "text": "他卖的不是一颗芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 404,
      "start": 950.866,
      "end": 952.733,
      "text": "而是一整座AI工厂",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 405,
      "start": 952.8,
      "end": 954.533,
      "text": "但随着推理占总算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 406,
      "start": 954.533,
      "end": 956.166,
      "text": "力需求的比例上升",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 407,
      "start": 956.4,
      "end": 958.4,
      "text": "客户会越来越认真地计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 408,
      "start": 958.4,
      "end": 961.5,
      "text": "每个TOKEN的电力延迟和成本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 409,
      "start": 962.066,
      "end": 964.066,
      "text": "只要工作负载足够稳定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 410,
      "start": 964.2,
      "end": 966.566,
      "text": "专用芯片就有机会从GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 411,
      "start": 966.566,
      "end": 969.533,
      "text": "手里切走一部分任务Google TPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 412,
      "start": 969.933,
      "end": 971.933,
      "text": "Amazon Trion和Inferentia",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 413,
      "start": 972.166,
      "end": 975.166,
      "text": "微软Maya以及各种边缘NPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 414,
      "start": 975.2,
      "end": 977.2,
      "text": "本质上都在做类似的选择",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 415,
      "start": 977.3,
      "end": 979.366,
      "text": "不是所有AI计算都必须",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 416,
      "start": 979.366,
      "end": 981.933,
      "text": "交给最通用最昂贵的GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 417,
      "start": 982.2,
      "end": 984.466,
      "text": "Northpointe把这个选择推得更远",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 418,
      "start": 984.733,
      "end": 987.333,
      "text": "它不是简单增加矩阵乘法单元",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 419,
      "start": 987.6,
      "end": 990.4,
      "text": "而是重新安排内存与计算的关系",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 420,
      "start": 990.5,
      "end": 993.366,
      "text": "把少搬数据放在多做计算之前",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 421,
      "start": 993.666,
      "end": 996.733,
      "text": "AI行业过去几年的扩张方式很直接",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 422,
      "start": 996.966,
      "end": 999.4,
      "text": "模型变大就买更多GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 423,
      "start": 999.733,
      "end": 1002.5,
      "text": "上下文变长就上更多HBM",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 424,
      "start": 1002.7,
      "end": 1005.866,
      "text": "机架功率增加就修更大的数据中心",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 425,
      "start": 1006.1,
      "end": 1007.333,
      "text": "这个办法有效",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 426,
      "start": 1007.333,
      "end": 1008.666,
      "text": "却越来越昂贵",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 427,
      "start": 1008.766,
      "end": 1012.2,
      "text": "电网接入时间液冷系统HBM供应",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 428,
      "start": 1012.2,
      "end": 1014.666,
      "text": "和先进封装都开始成为限制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 429,
      "start": 1014.866,
      "end": 1016.466,
      "text": "north pole提醒行业",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 430,
      "start": 1016.666,
      "end": 1019.3,
      "text": "有些增长不必来自更大的蛮力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 431,
      "start": 1019.533,
      "end": 1021.366,
      "text": "也可以来自不做无用功",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 432,
      "start": 1021.5,
      "end": 1023.133,
      "text": "这里还藏着一个经常",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 433,
      "start": 1023.133,
      "end": 1024.8,
      "text": "被忽略的经济问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 434,
      "start": 1025.4,
      "end": 1027.133,
      "text": "数据中心买GPU时",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 435,
      "start": 1027.133,
      "end": 1029.166,
      "text": "采购价格只是一部分成本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 436,
      "start": 1029.766,
      "end": 1031.466,
      "text": "芯片几年生命周期里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 437,
      "start": 1031.466,
      "end": 1033.766,
      "text": "电力冷却机房空间和",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 438,
      "start": 1033.766,
      "end": 1035.8,
      "text": "利用率会持续产生支出",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 439,
      "start": 1035.9,
      "end": 1037.766,
      "text": "一块便宜但经常等待",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 440,
      "start": 1037.766,
      "end": 1040.5,
      "text": "内存的芯片未必真的便宜",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 441,
      "start": 1040.766,
      "end": 1042.733,
      "text": "一套前期适配复杂",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 442,
      "start": 1042.966,
      "end": 1046.0,
      "text": "长期每次推理都更省电的专用系统",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 443,
      "start": 1046.1,
      "end": 1049.0,
      "text": "在高调用量场景里可能更有优势",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 444,
      "start": 1049.6,
      "end": 1051.6,
      "text": "north pole的商业价值要",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 445,
      "start": 1051.6,
      "end": 1053.5,
      "text": "用总拥有成本计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 446,
      "start": 1053.533,
      "end": 1055.333,
      "text": "而不是只看晶圆制程",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 447,
      "start": 1055.5,
      "end": 1058.066,
      "text": "可总拥有成本不能只算电",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 448,
      "start": 1058.2,
      "end": 1061.333,
      "text": "专用芯片的开发工具模型转换",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 449,
      "start": 1061.466,
      "end": 1064.933,
      "text": "工程师培训备件和供应规模都要",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 450,
      "start": 1064.933,
      "end": 1065.733,
      "text": "花钱",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 451,
      "start": 1065.866,
      "end": 1069.466,
      "text": "英伟达GPU贵却很容易找到开发者",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 452,
      "start": 1069.733,
      "end": 1073.166,
      "text": "框架和现成方案企业遇到问题也能",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 453,
      "start": 1073.166,
      "end": 1076.366,
      "text": "从庞大社区和合作伙伴中找到答案",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 454,
      "start": 1076.7,
      "end": 1079.5,
      "text": "north pole若想把能效优势变成订单",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 455,
      "start": 1079.7,
      "end": 1081.366,
      "text": "节省的电费必须大到",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 456,
      "start": 1081.366,
      "end": 1083.1,
      "text": "足以覆盖迁移成本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 457,
      "start": 1083.166,
      "end": 1085.5,
      "text": "还要保证芯片能持续供货",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 458,
      "start": 1085.566,
      "end": 1087.3,
      "text": "这也是先进制程比较",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 459,
      "start": 1087.3,
      "end": 1088.8,
      "text": "最容易误导的地方",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 460,
      "start": 1089.366,
      "end": 1092.866,
      "text": "4纳米GPU承担的是更广泛任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 461,
      "start": 1093.066,
      "end": 1097.566,
      "text": "包含缓存调度图形和通用计算能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 462,
      "start": 1097.9,
      "end": 1100.2,
      "text": "12纳米north pole删除了",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 463,
      "start": 1100.2,
      "end": 1101.7,
      "text": "大量不需要的能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 464,
      "start": 1101.766,
      "end": 1104.666,
      "text": "两者像一辆能在各种道路行驶的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 465,
      "start": 1104.666,
      "end": 1108.1,
      "text": "卡车和一列只跑固定路线的电车",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 466,
      "start": 1108.366,
      "end": 1111.166,
      "text": "电车在自己的轨道上效率极高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 467,
      "start": 1111.266,
      "end": 1114.066,
      "text": "不能据此断言卡车技术落后",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 468,
      "start": 1114.166,
      "end": 1115.4,
      "text": "正确结论是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 469,
      "start": 1115.4,
      "end": 1117.1,
      "text": "当路线足够固定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 470,
      "start": 1117.1,
      "end": 1119.133,
      "text": "专用设计可以用成熟制程",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 471,
      "start": 1119.133,
      "end": 1121.2,
      "text": "赢过先进制程的通用芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 472,
      "start": 1121.533,
      "end": 1123.7,
      "text": "成熟制程还有现实优势",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 473,
      "start": 1124.366,
      "end": 1125.866,
      "text": "晶圆成本更低",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 474,
      "start": 1126.0,
      "end": 1127.3,
      "text": "产能更充足",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 475,
      "start": 1127.466,
      "end": 1129.133,
      "text": "制造经验更稳定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 476,
      "start": 1129.6,
      "end": 1131.7,
      "text": "AI行业把注意力集中在",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 477,
      "start": 1131.7,
      "end": 1133.666,
      "text": "台积电最先进节点",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 478,
      "start": 1133.733,
      "end": 1135.933,
      "text": "但并非所有推理芯片都",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 479,
      "start": 1135.933,
      "end": 1138.0,
      "text": "需要增强最稀缺产能",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 480,
      "start": 1138.066,
      "end": 1140.6,
      "text": "如果架构能在12纳米实现目标",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 481,
      "start": 1140.766,
      "end": 1142.966,
      "text": "同样预算可以购买更多芯片",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 482,
      "start": 1143.266,
      "end": 1144.733,
      "text": "也能降低供应链对",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 483,
      "start": 1144.733,
      "end": 1146.466,
      "text": "单一先进节点的依赖",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 484,
      "start": 1147.3,
      "end": 1149.6,
      "text": "north pole给出的不只是技术路线",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 485,
      "start": 1149.7,
      "end": 1151.533,
      "text": "也是一种供应链选择",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 486,
      "start": 1151.9,
      "end": 1152.8,
      "text": "但反过来",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 487,
      "start": 1152.9,
      "end": 1154.333,
      "text": "专用系统也有被",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 488,
      "start": 1154.333,
      "end": 1156.1,
      "text": "模型变化淘汰的风险",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 489,
      "start": 1156.466,
      "end": 1159.4,
      "text": "如果模型架构每半年就大改一次",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 490,
      "start": 1159.566,
      "end": 1162.166,
      "text": "今天为某种数据流设计的硬件",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 491,
      "start": 1162.2,
      "end": 1163.933,
      "text": "明天可能难以适配",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 492,
      "start": 1164.466,
      "end": 1167.266,
      "text": "GPU虽然没有在每项任务上做到极致",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 493,
      "start": 1167.4,
      "end": 1169.666,
      "text": "却能承受行业快速变化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 494,
      "start": 1169.733,
      "end": 1171.966,
      "text": "客户为通用性支付的溢价",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 495,
      "start": 1172.1,
      "end": 1173.366,
      "text": "实际上是在购买",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 496,
      "start": 1173.366,
      "end": 1175.1,
      "text": "未来的不确定性保险",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 497,
      "start": 1175.333,
      "end": 1176.7,
      "text": "Northpo与英伟达",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 498,
      "start": 1176.933,
      "end": 1179.6,
      "text": "因此不是简单的谁把谁打败",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 499,
      "start": 1179.8,
      "end": 1181.766,
      "text": "它们代表两种不同答案",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 500,
      "start": 1182.266,
      "end": 1185.733,
      "text": "GPU的答案是用强大的通用并行计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 501,
      "start": 1185.8,
      "end": 1189.133,
      "text": "外部高带宽内存和成熟软件生态",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 502,
      "start": 1189.133,
      "end": 1190.766,
      "text": "覆盖尽可能多的任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 503,
      "start": 1191.0,
      "end": 1192.8,
      "text": "north pole的答案是主动",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 504,
      "start": 1192.8,
      "end": 1195.0,
      "text": "放弃训练和部分灵活性",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 505,
      "start": 1195.366,
      "end": 1197.966,
      "text": "把固定推理路径压缩到片上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 506,
      "start": 1197.966,
      "end": 1199.566,
      "text": "用结构换效率",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 507,
      "start": 1199.733,
      "end": 1201.666,
      "text": "未来最可能出现的不是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 508,
      "start": 1201.666,
      "end": 1204.266,
      "text": "专用芯片一夜取代GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 509,
      "start": 1204.366,
      "end": 1207.0,
      "text": "而是AI基础设施开始分层",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 510,
      "start": 1207.366,
      "end": 1210.5,
      "text": "训练和快速实验继续留在GPU上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 511,
      "start": 1210.966,
      "end": 1212.8,
      "text": "稳定的大规模推理转向",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 512,
      "start": 1212.8,
      "end": 1216.1,
      "text": "TPU ASIC或north pole式架构",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 513,
      "start": 1216.266,
      "end": 1218.466,
      "text": "手机汽车和工业设备",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 514,
      "start": 1218.466,
      "end": 1220.333,
      "text": "使用更小的边缘NPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 515,
      "start": 1220.5,
      "end": 1222.266,
      "text": "工作负载被拆开以后",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 516,
      "start": 1222.533,
      "end": 1225.066,
      "text": "英伟达仍然可能是最大赢家",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 517,
      "start": 1225.133,
      "end": 1227.5,
      "text": "却不再天然拥有每一个环节",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 518,
      "start": 1227.766,
      "end": 1230.133,
      "text": "north pole最值得记住的数字",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 519,
      "start": 1230.3,
      "end": 1236.166,
      "text": "也许不是22倍26倍46.9倍或者72.7倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 520,
      "start": 1236.3,
      "end": 1238.1,
      "text": "而是192兆",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 521,
      "start": 1238.2,
      "end": 1240.733,
      "text": "就是这块看起来不大的片上内存",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 522,
      "start": 1241.2,
      "end": 1242.766,
      "text": "让它绕过数据搬运",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 523,
      "start": 1243.2,
      "end": 1245.266,
      "text": "赢得惊人的速度和能效",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 524,
      "start": 1245.466,
      "end": 1247.4,
      "text": "也正是这192招",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 525,
      "start": 1247.7,
      "end": 1250.733,
      "text": "限制了模型大小上下文和灵活性",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 526,
      "start": 1250.9,
      "end": 1252.466,
      "text": "它的优势和代价",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 527,
      "start": 1252.466,
      "end": 1254.133,
      "text": "来自同一个设计决定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 528,
      "start": 1254.266,
      "end": 1256.1,
      "text": "这也回答了开头的问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 529,
      "start": 1256.766,
      "end": 1258.533,
      "text": "Northpole没有成为商品",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 530,
      "start": 1258.566,
      "end": 1260.266,
      "text": "不是因为实验失败",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 531,
      "start": 1260.333,
      "end": 1262.5,
      "text": "而是因为它证明的太专一",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 532,
      "start": 1262.6,
      "end": 1265.533,
      "text": "它证明了12纳米可以在特定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 533,
      "start": 1265.533,
      "end": 1268.2,
      "text": "推理任务上击败4纳米GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 534,
      "start": 1268.533,
      "end": 1271.466,
      "text": "证明了内存墙并非不可改变",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 535,
      "start": 1271.766,
      "end": 1274.1,
      "text": "也证明了AI硬件还有一条",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 536,
      "start": 1274.1,
      "end": 1276.4,
      "text": "不靠无限增加功耗的路",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 537,
      "start": 1276.466,
      "end": 1279.733,
      "text": "可要把这条路变成普遍适用的产品",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 538,
      "start": 1280.266,
      "end": 1284.2,
      "text": "IBM还得解决容量软件和模型适配",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 539,
      "start": 1284.366,
      "end": 1285.6,
      "text": "对英伟达而言",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 540,
      "start": 1285.6,
      "end": 1287.8,
      "text": "真正的威胁不是IBM明天",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 541,
      "start": 1287.8,
      "end": 1289.8,
      "text": "推出一张north pole显卡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 542,
      "start": 1290.133,
      "end": 1291.366,
      "text": "而是越来越多客户",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 543,
      "start": 1291.366,
      "end": 1292.766,
      "text": "开始问同一个问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 544,
      "start": 1292.966,
      "end": 1295.533,
      "text": "既然大部分电力花在搬数据上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 545,
      "start": 1295.533,
      "end": 1297.1,
      "text": "为什么还要为每次推理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 546,
      "start": 1297.1,
      "end": 1299.466,
      "text": "都使用最昂贵的通用GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 547,
      "start": 1299.533,
      "end": 1301.466,
      "text": "只要这个问题被持续追问",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 548,
      "start": 1301.666,
      "end": 1305.2,
      "text": "AI芯片战争就会从制程和算力转向",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 549,
      "start": 1305.2,
      "end": 1308.5,
      "text": "内存位置数据流和每个TOKEN的成本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 550,
      "start": 1308.5,
      "end": 1309.3,
      "text": "所以",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 551,
      "start": 1309.8,
      "end": 1311.766,
      "text": "12纳米并没有推翻4纳米",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 552,
      "start": 1312.4,
      "end": 1313.866,
      "text": "north pole推翻的是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 553,
      "start": 1313.866,
      "end": 1315.8,
      "text": "另一种更顽固的迷信",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 554,
      "start": 1316.266,
      "end": 1317.733,
      "text": "只要制程更新",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 555,
      "start": 1317.8,
      "end": 1319.0,
      "text": "GPU更大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 556,
      "start": 1319.1,
      "end": 1320.133,
      "text": "功耗更高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 557,
      "start": 1320.2,
      "end": 1322.133,
      "text": "AI就一定会更有效率",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 558,
      "start": 1322.333,
      "end": 1325.366,
      "text": "IBM用一颗研究芯片说明",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 559,
      "start": 1325.866,
      "end": 1327.733,
      "text": "架构仍然可以改变规则",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 560,
      "start": 1327.933,
      "end": 1330.466,
      "text": "英伟达今天不会因此失去王座",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 561,
      "start": 1330.6,
      "end": 1333.566,
      "text": "但它必须面对一个更难防守的战场",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 562,
      "start": 1334.0,
      "end": 1335.3,
      "text": "在这个战场上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 563,
      "start": 1335.366,
      "end": 1338.2,
      "text": "竞争对手不需要做出更大的GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 564,
      "start": 1338.366,
      "end": 1341.133,
      "text": "只需要让大量任务不再需要GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 565,
      "start": 1341.366,
      "end": 1342.2,
      "text": "试试核验",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 566,
      "start": 1342.6,
      "end": 1345.333,
      "text": "IBM Northpole采用12纳米工艺",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 567,
      "start": 1345.6,
      "end": 1349.966,
      "text": "拥有220亿个晶体管与256个核心属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 568,
      "start": 1350.366,
      "end": 1352.966,
      "text": "分布式DRAM容量为192兆",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 569,
      "start": 1353.133,
      "end": 1355.733,
      "text": "片上带宽约13太字节每秒",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 570,
      "start": 1356.333,
      "end": 1358.733,
      "text": "属实在限定视觉推理任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 571,
      "start": 1358.733,
      "end": 1360.9,
      "text": "中超过一款4纳米GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 572,
      "start": 1361.2,
      "end": 1364.6,
      "text": "属实相对可比12纳米GPU约25倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 573,
      "start": 1364.6,
      "end": 1367.6,
      "text": "能效和约22倍低延迟属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 574,
      "start": 1368.066,
      "end": 1369.4,
      "text": "每焦耳工作量提高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 575,
      "start": 1369.4,
      "end": 1371.2,
      "text": "22倍的说法不准确",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 576,
      "start": 1371.466,
      "end": 1376.366,
      "text": "16卡系统达到每秒28,356TOKEN属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 577,
      "start": 1376.7,
      "end": 1379.266,
      "text": "288卡研究系统与30千瓦",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 578,
      "start": 1379.266,
      "end": 1381.166,
      "text": "功耗数据有论文支持",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 579,
      "start": 1381.566,
      "end": 1384.333,
      "text": "Northpole全面击败现代英伟达GPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 580,
      "start": 1384.333,
      "end": 1387.533,
      "text": "不属实Northpole已经商业化不属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 581,
      "start": 1387.8,
      "end": 1389.466,
      "text": "IBM已经商业化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 582,
      "start": 1389.6,
      "end": 1392.0,
      "text": "另一款Spire加速器属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    }
  ],
  "transcript_selection": {
    "schema": "videonote.transcript_selection.v1",
    "status": "SELECTED",
    "requested_language": "zh",
    "duration": 1392.0,
    "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260922_091411__12奈米竟虐殺4奈米_IBM這次真的把英偉達逼瘋了_老黃看完直接氣哭/reference_subtitle.vtt",
    "reference_kind": "manual",
    "reference_language": "zh",
    "selected_source": "video_source_subtitle",
    "source_caption_kind": "manual",
    "source_caption_processing": "verbatim_direct",
    "youtube_overlap_clamps": [],
    "general_high_confidence_repairs": [],
    "decision_reason": "complete_source_subtitle_fast_path_no_whisper",
    "whisper_rejected": false,
    "source_profile": {
      "segment_count": 583,
      "duration_coverage": 0.926203
    }
  }
}