Huawei AI CloudMatrix 384是一款针对英国伟达GB200 NVL72的应用程序

Apri1,2025 HuaweiACoudMatrix8ChinaʼsAnswertoNvidiaGB200NVL72//ChinaAbundanceofPower,100%Optics,0%Copper,Powernefficiency,2.xowerFLOPperWatt,1TransceiversperChip,LinearPuggabeOptics华为A云矩阵8中国对英伟达GB200NVL72的回应中国电⼒资源丰富，100%光互联，零铜缆，能效较低，每⽡特FLOP性能低2.倍，每芯⽚配备1个光模块，采⽤线性可插拔光学技术 By,,,,,,,and作者：DyanPate、DanieNishba、MyronXie、WegaChu、PatrickZhou、vanChiam、AJKourabi、ChristopherSeife和DougO'LaughinHuaweiismakingwaveswithitsnewAacceeratorandrackscaearchitecture.MeetChinaʼsnewestandmostpowerfuChinesedomesticsoution,theCoudMatrix8buitusingtheAscend10C.ThissoutioncompetesdirectywiththeGB200NVL72,andinsomemetricsismoreadvancedthanNvidiaʼsrackscaesoution.Theengineeringadvantageisatthesystemevenotjustatthechipeve,withinnovationattheacceerator,networking,optics,andsoftwareayers.DyanPateDanieNishbaMyronXieWegaChu PatrickZhouvanChiam AJKourabi ChristopherSeifeDougO'Laughin 华为正凭借其全新A加速器及机柜级架构掀起波澜。这款名为CoudMatrix8的中国本⼟最强解决⽅案，基于昇腾10C打造，直接对标GB200NVL72，部分指标甚⾄超越英伟达的机柜级⽅案。其⼯程优势不仅体现在芯⽚层⾯，更贯穿加速器、⽹络、光模块及软件层的系统级创新。 Source:Huawei来源：华为 TheHuaweiAscendchipisnotnewtoSemiAnaysis,butina,HuaweiispushingtheimitsofAsystemperformance.Therearetrade-offs,butgivenexportcontrosandackusterdomesticyieds,itʼscearthattherearefurtheroophoesintheChineseexportcontros.wordwheresystemsmattermorethanmicroarchitecture 华为昇腾芯⽚对SemiAnaysis⽽⾔并不陌⽣，但在⼀个系统性能⽐微架构更重要的世界⾥，华为正在突破A系统性能的极限。虽然存在权衡取舍，但考虑到出⼝管制和国内芯⽚良率不尽如⼈意，显然中国出⼝管制还存在更多漏洞。 GoogleAIInfrastructureSupremacy:SystemsMatterMoreThanMicroarchitecture ⾕歌AI基础设施霸权：系统⽐微架构更重要 FromDLRMtoLLM,internalworkloadswin,buthowdoesGooglefareinexternalworkloads?ThedawnoftheAIeraishere,anditiscrucialtounderstandthatthecoststructureofAI-drivensoftwaredeviatesconsiderablyfrom traditionalsoftware.Chipmicroarchitectureandsystemarchitectureplayavitalroleinthedevelopmentandscalability…Continuereading 从DLRM到LLM内部⼯作负载占据优势但Google在外 WhietheAscendchipcanbefabricatedatSMC,wenotethatthisisagobachipthathas,,andisfabricatedby.WedoadeepdiveintowhatispossibefordomesticChineseproductionwhatisanaggressiveskirtingoftheexportcontros,andwhytheUSgovernmentneedstofocusontheseHBMfromKorea primarywaferproductionfromTSMC10sofbiionsofwaferfabricationequipmentfromtheUS,Netherands,andJapan keynewareastoimitChinaʼsAcapabiities. 尽管昇腾芯⽚可以在中芯国际⽣产，但我们注意到这是⼀款全球化芯⽚——其HBM来⾃韩国，主要晶圆⽣产依赖台积电，并由价值数百亿美元的美国、荷兰和⽇本晶圆制造设备完成加⼯。我们将深⼊探讨中国本⼟⽣产的可能性边界、这种激进规避出⼝管制的⾏为本质，以及美国政府为何需要聚焦这些关键新领域以限制中国A能⼒发展。 FabWhack-AMole:ChineseCompaniesareEvadingU.S.Sanctions 芯⽚制造猫⿏游戏：中国企业如何规避美国制裁 HuaweiFabNetwork,WFEVendorsCryWolf,FrameworkforFutureControlsAIcompetitivenessisakeynationalsecurityconcern.When“expert-levelscienceandengineeringˮorevenAGIarepossibleoutcomes,itiscrucialthattheU.S.maintainsorextendsitsleadintheseareas.Considertheopposite:whatifChinaweretoachieve Huaweiisagenerationbehindinchips,butitsscae-upsoutionisarguabyagenerationaheadofNvidiaandAMDʼscurrentproductsonthemarket.SowhatwoudbethespecificationsforHuaweiʼsCoudMatrix8CM8? 华为在芯⽚上落后⼀代，但其扩展解决⽅案可以说⽐英伟达和AMD当前市⾯上的产品领先⼀代。那么华为的CoudMatrix8（CM8）规格会是怎样的呢？ TheCoudMatrix8consistsof8Ascend10Cchipsconnectedthroughana-to-atopoogy.Thetradeoffissimpe:havingfivetimesasmanyAscendsmorethanoffsetseachGPUbeingonyone-thirdtheperformanceofanNvidiaBackwe. CoudMatrix8由8颗Ascend10C芯⽚通过全互联拓扑连接⽽成。其取舍很简单：Ascend芯⽚数量是NvidiaBackwe的五倍之多，⾜以弥补单个GPU性能仅为后者三分之⼀的事实。 来源：SemiAnaysis、Nvidia、华为 AfuCoudMatrixsystemcannowdeiver00PFLOPsofdenseBF1compute,amostdoubethatoftheGB200NVL72.Withmorethan.xaggregatememorycapacityand2.1xmorememorybandwidth,HuaweiandChinanowhaveAsystemcapabiitiesthatcanbeatNvidiaʼs. 如今，⼀套完整的CoudMatrix系统可提供00PFLOPs的密集BF1算⼒，⼏乎是GB200NVL72的两倍。凭借超过.倍的总内存容量和2.1倍的内存带宽，华为与中国现已拥有能超越英伟达的A系统能⼒。 Whatʼsmore,istheCM8isuniqueysuitedtoChinaʼsstrengths,whichisdomesticnetworkingproduction,infrastructuresoftwaretopreventnetworkfaiures,andwithfurtheryiedimprovements,anabiitytoscaeuptoevenargerdomains. 更重要的是，CM8完美契合中国的优势领域，包括本⼟⽹络⽣产能⼒、预防⽹络故障的基础设施软件，以及随着良率进⼀步提升，还能扩展⾄更⼤型的领域。 Thedrawbackhereisthatittakes.xthepowerofaGB200NVL72,with2.xworsepowerperFLOP,1.8xworsepowerperTB/smemorybandwidth,and1.1xworsepowerperTBHBMmemorycapacity. 此⽅案的缺点在于其功耗是GB200NVL72的.倍，每FLOP能效⽐低2.倍，每TB/s内存带宽能效⽐低1.8倍，每TBHBM内存容量能效⽐低1.1倍。 ThedeficienciesinpowerarereevantbutnotaimitingfactorinChina. 能效⽅⾯的不⾜确实存在，但在中国并⾮限制性因素。 ChinahasNoPowerConstraints,justSiiconConstraints 中国没有电⼒限制，只有硅限制 ThecommonrefrainintheWestisthat,butinChina,thisistheopposite.TheWesthasspenttheastdecadeshiftingaprimariycoa-basedpowerinfrastructuretogreenernaturagasandrenewabepowergenerationpairedwithmoreefficientenergyusageonapercapitabasis.ThisistheoppositeinChina,whererisingifestyesandcontinuedheavyinvestmentmeanmassivepowergenerationdemand.Aispower-imited 西⽅普遍认为⼈⼯智能受限于电⼒供应，但在中国情况恰恰相反。过去⼗年间，西⽅致⼒于将主要依赖煤炭的电⼒基础设施转向更环保的天然⽓和可再⽣能源发电，并提⾼⼈均能源使⽤效率。⽽中国则因⽣活⽔平提升和持续⼤规模投资，电⼒需求呈现爆发式增⻓。 来源：SemiAnaysis数据中⼼模型 Mostofthishasbeenpoweredbycoa,butChinaasohastheargestinsta