英伟达2026 GTC大会 黄仁勋演讲课件_第1页
英伟达2026 GTC大会 黄仁勋演讲课件_第2页
英伟达2026 GTC大会 黄仁勋演讲课件_第3页
英伟达2026 GTC大会 黄仁勋演讲课件_第4页
英伟达2026 GTC大会 黄仁勋演讲课件_第5页
已阅读5页,还剩74页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

2026AnnouncingDLSS

53D-GuidedNeuralRenderingStructuredDataistheGroundTruthofAI$120BStructuredDataEcosystem–TheGroundTruthofEnterprisesSource:Gartner

DBMS,2025Unstructured

Data

is

the

Context

of

AI100’sofZettabytesPerYearofUnstructuredData–GrowingExponentiallySource:IDC

Global

DataSphere,

2025“Working

with

IBM

and

NVIDIA,early

results

show

we

can

refresh

global

operationsdata

in

minutes

atreducedcost—makingacapabilitythatcanbeturned

intotangiblebusinessimpactinareassuchasmanufacturingorwarehousing.

”―ChrisWright,ChiefInformationandDigitalOfficerof

NestléIBM

watsonx.dataData

Engines:NVIDIAcuDFOpen

Formats:DataGovernanceandAccess

Control

ThroughoutNVIDIA

A100IBM

Storage

Scale83%CostSavingsCPU

GPU1,0007505002500Order-to-CashDataMartUpdates3min

vs

15min“Dell

AI

Data

Platform

with

NVIDIA

helps

us

build

scalableandrepeatabledatapipelines

that

drive

automation—processing

massive

data

volumes

inminutesinsteadofhoursanddeliveringtransformative

valueforourclients.”__AbhijitDubey,CEOof

NTT

DATADellAI

Data

PlatformDataOrchestration

EngineData

EnginesNVIDIAcuVS

NVIDIAcuDFHigh

PerformanceScale-OutStorageNVIDIACMXData

ProtectionandCyber

ResilienceNVIDIARTX

PRO6000

BlackwellDellAI

Factorywith

NVIDIAEnterpriseDataProcessing

FromHoursto

Minutes10TB

Dataset100806040200CPU

GPU3XFaster“Our

collaboration

with

NVIDIA

and

Google

Cloud

helps

us

innovate

faster

for

more

than

abillion

Snapchatters

worldwide.By

lowering

costs

and

scaling

experiments

across

petabytesof

data,we’re

delivering

AI-powered

experiences

more

quickly

andefficiently.”―SaralJain,CIO

ofSnapGoogleCloudAI

HypercomputerFlexibleConsumptionOpenSoftwareApacheSpark

EngineApacheSparkJobsNVIDIAcuDFComputeG2VM

(NVIDIA

L4GPUs)

G4VM

(NVIDIA

RTX

PRO

6000)GoogleKubernetes

EngineControl

plane,APIs,Scheduling

and

monitoringPerformance-optimized

HardwareDataprocManagedApacheSpark

and

HadoopNetworkingJupiter

NetworkStorageGCS

RapidGCS

Lustre50,00040,00030,00020,0000A/B

Experimentation

Platform

45kCPUsto

1k

GPUs76%CostSavingsCPU

GPU10,000AIOptimizedInfrastructureOpenSoftware,

FrameworksLiquid

CoolingStorageNetworkingGoogleDistributedCloudVertexAIGoogleKubernetesEngineCloud

RunGoogleComputeEngineBigQueryDataprocDataflowBigtableJAX

XLA

vLLMSystemBuildersandOEMsDataScience

andAnalyticsAIPlatformsNVIDIAAIPlatformAmazonSageMakerAISystemBuildersandOEMsAWSParallelClusterAWSNitroEnclavesAmazonBedrockAmazonEC2AmazonEKSAmazonEMRAmazonEFAAmazonBraketNVIDIAAIPlatformAWSBatchAzureKubernetesServiceSystemBuildersandOEMsAzureContainer

AppsMicrosoft365

CopilotMicrosoftFoundryMicrosoftBingWindows

365MicrosoftFabricFoundry

LocalNVIDIAAIPlatformOracleAutonomousAILakehouseSystemBuildersandOEMsOCIAIAcceleratorPacksOCISuperclustersOracle

AIDataPlatformOCIGenerativeAIOCIDataScienceOracleAIDatabaseNVIDIAAIPlatformOCIOKECoreWeaveKubernetesKubernetesSystemBuildersandOEMsCoreWeaveMissionControlCoreWeaveStorageCoreWeaveComputeCoreWeaveInferenceW&BInferenceW&BTrainingNVIDIAAIPlatformW&BModelsW&BWeaveCoreWeaveSlurm

onServiceRubix

Ontology

Pipeline

Builder

Fusion

Agent

Studio

Model

Studio

AIP

Logic

Model

Catalog

Workshop

AI

FDEDellNVIDIAAIPlatformInference

Frameworks

Agent

Frameworks/Protocols

Frontier

Model

Builders

Model

to

ProductionDL

FrameworksAI

for

Auto

AI

for

Customer

Support

AI

for

Engineering

AI

for

Healthcare

AI

for

Robotics

AI

for

Search

AI

for

Software

Development

AI

NativesNVIDIAAIPlatformInferenceInflectionArrives10,000X

ChatGPT

ComputeInferenceInflectionDrivesStrongGrowthNVIDIAFull-StackExpandingAItoAllRegionsand

IndustriesTokensperWattDrivesFactory

RevenuePerformanceDrivesTokenCostInferenceXbyGB300NVL7250XHigherPerf/

WattH200NVL8Competition50100

150

200Interactivity(TPS/User)9,000,0008,000,0007,000,0006,000,0005,000,0004,000,0003,000,0002,000,0001,000,00000H200NVL8Competition35XLowerCostGB300NVL720

50

100

150

200NVIDIAExtremeCo-DesignRevolutionizedTokenCost“GB

NVL72Inference

King”$4.50$4.00$3.50$3.00$2.50$2.00$1.50$1.00$0.50$0.00DeepSeekR10528·FP4·1K/1K·

Source:

SemiAnalysis

InferenceXInteractivity

(TPS/User)NVIDIAistheGlobalStandardforAIInferenceatScaleAllLeadingInference

End-PointsRun

on

NVIDIAKimi

K2.5ReasoningAIFactoriesaretheIndustrialInfrastructureof

theAI

EraInferenceistheWorkloadTokensaretheNewCommodityComputeis

RevenueInferencePerformanceandEfficiencyDriveCompanyResults50

1002004006008001,000+InteractivityTPS/UserQwen

3235Billion32KKimi

K2.51Trillion128KGPT

MoE2Trillion128KGPT

MoE2Trillion400KGPT

MoE2Trillion400K$0$3$6$45$150InferencePerformanceandEfficiency

DriveCompanyResults1.8

M1.6

M1.4

M1.2

M1.0M0.8

M0.6

M0.4

M0.2

MFreeMediumHighPremium

UltraThroughput

TPS/MWInferencePerformanceandEfficiency

DriveCompanyResults1.8

M1.6

M1.4

M1.2

M1.0M0.8

M0.6

M0.4

M0.2

M100InteractivityTPS/UserMediumKimi

K2.51Trillion128K$3FreeQwen

3235Billion32K$0Throughput

TPS/MWHopper20050Blackwell

NVL72HopperInteractivityTPS/UserMediumKimi

K2.51Trillion128K50FreeQwen

3235Billion32K0.8

M0.6

M0.4

M0.2

M1.8

M1.6

M1.4

M1.2

M200HighGPT

MoE2Trillion128K35XInferencePerformanceandEfficiency

DriveCompanyResults1.0M$0

$3

$6Throughput

TPS/MWInferencePerformanceandEfficiency

DriveCompanyResultsBlackwellNVL721.8

M1.6

M1.4

M1.2

M1.0M0.8

M0.6

M0.4

M0.2

MPremiumGPT

MoE2Trillion

400K$45200InteractivityTPS/UserMediumKimi

K2.51Trillion128K$3FreeQwen

3235Billion32K$0HighGPT

MoE2Trillion128K$6Throughput

TPS/MWHopper35X40010050RubinNVL722X2XBlackwellNVL723XHopper10X50100200

400InteractivityTPS/User1.6

M1.4

M1.2

M0.8

M0.6

M0.4

M0.2

MPremiumGPT

MoE2Trillion

400K$45MediumKimi

K2.51Trillion128K$3FreeQwen

3235Billion32K$0HighGPT

MoE2Trillion128K$61.8

MInferencePerformanceandEfficiency

DriveCompanyResults1.0MThroughput

TPS/MWFreeMediumHighPremiumTotalAnnualQwen

3Kimi

K2.5GPT

MoEGPT

MoERevenue235

Billion32K1Trillion128K2Trillion128K2Trillion400K$0$3$6$45$150B$100BAnnual

RevenuePerGigawatt$50B$0BNVIDIAVeraRubin

Unlocks$150BRevenueOpportunity

5X

$30B

Blackwell

Rubin$150BRubinNVL722X2XBlackwellNVL723XHopper10X50

100

200

400InteractivityTPS/User1.6

M1.4

M1.2

M0.8

M0.6

M0.4

M0.2

MPremiumGPT

MoE2Trillion

400K$45MediumKimi

K2.51Trillion128K$3FreeQwen

3235Billion32K$0HighGPT

MoE2Trillion128K$61.8

MInferencePerformanceandEfficiency

DriveCompanyResults1.0MThroughput

TPS/MWRubinNVL72

2X

2X

BlackwellNVL72

3X

35XHopper10XInferencePerformanceandEfficiency

DriveCompanyResults1.8

M1.6

M1.4

M1.2

M1.0M0.8

M0.6

M0.4

M0.2

M50

100200

400PremiumGPT

MoE2Trillion

400K$45MediumKimi

K2.51Trillion128K$3FreeQwen

3235Billion32K$0HighGPT

MoE2Trillion128K$6InteractivityTPS/UserThroughput

TPS/MW0.0M3XHopper

35X50

100

20040035XPerformanceEfficiencyDrivesCompanyResults1.8

M1.6

M1.4

M1.2

M1.0M0.8

M0.6

M0.4

M0.2

M$6

$45

$1502XBlackwell

NVL72FreeMediumQwen

3

Kimi

K2.5235

Billion

1

Trillion32K

128K$0

$3PremiumGPT

MoE2Trillion

400K35X600InteractivityTPS/UserUltraGPT

MoE2Trillion400KHighGPT

MoE2Trillion128KRubin

+

LPXRubin

NVL72Throughput

TPS/MW1,000+8002XFreeQwen

3235

Billion32K$0MediumKimi

K2.51Trillion128K$3HighGPT

MoE2Trillion128K$6PremiumGPT

MoE2Trillion400K$45UltraGPT

MoE2Trillion400K$150NVIDIAVeraRubin

+

LPXUnlocks$300B

RevenueOpportunity$300B$150B$30B$300B$250B$200B$150B$100B$50B$0B Blackwell

Rubin

VR+LPX10XAnnual

RevenuePerGigawattTotalAnnualRevenueGroq

3

LPU500MB

SRAM150

TB/s

SRAM

Bandwidth1.2

PFLOPs

(FP8)98B

TransistorsUnitingProcessorsofExtremePerformances288GB

HBM422TB/s50PFLOPs

(NVFP4)336B

Transistors

+

2.5T

(HBM4)RubinGPUGroq

3

LPU4GB

SRAM1,200TB/s

SRAM

Bandwidth

55X

9.6

PFLOPs

(FP8)784B

TransistorsUnitingProcessorsofExtremePerformances288GB

HBM422TB/s50PFLOPs

(NVFP4)336B

Transistors

+

2.5T

(HBM4)RubinGPUKV

CacheDecodeRemaining

Tokens1stTokenPrefillUnitingProcessorsofExtremeFLOPSand

BandwidthNVIDIADynamoVeraRubin

NVL72Groq3

LPXInputFPGALPU

C2C

Spine

Connectors8

LPUsBF4AIInferenceComputeSRAM

CapacityMemoryBandwidthScale-UpDensityScale-UpBandwidth315

PFLOPS128

GB40PB/s256

Chips640

TB/sNVIDIAGroq

3

LPXHostCPULPUC2C

LinksAvailable2H26AI

LabsOEMsandODMsAnnouncingNVIDIAVeraRubinNVL72LaunchPartners10X

Perf/W|

3.6

EF

NVFP4

|1.6

PB/s

HBM4

|

260

TB/s

NVLink6CloudOEMsandODMsCloudAnnouncingNVIDIAVeraCPULaunchPartners256Vera

CPUs|300

TB/s

LPDDR5X

|

ETL

Spine

|

6.5X

ThroughputAnnouncingNVIDIABlueField-4STXLaunchPartnersCloud5x

Tokens/sec|50

Tb/s

Networking

BW

|

16TB

Shared

Context/

GPUOEMsRubinGPUVeraCPUCX9BF4NVLinkSwitchSpectrumCPOGroq

3

LPUVera

Rubin

Compute

Tray

NVLink

Switch

Tray

Vera

Compute

Tray

BF4STX

Server

Spectrum

Switch

Groq3Compute

Tray1GW

AI

FactoryX86

+

HopperVera

Rubin#of

GPUs600K300KAI

FLOPS1.2ZFLOPS16ZFLOPSAll-to-AllScale-up7.2TB/s260

TB/sMemoryBW-per-Domain

(GROQSRAM)2

EB/s100EB/sTokens

per

Second2M700MNVIDIAVeraRubin7Chips–5

Rack

SystemsAIFactoryfortheAgenticAIFrontierNVIDIAExtremeCo-DesignDeliveringX-FactorsEveryYearFromChipstoRacks

toAI

FactoriesFeynmanRubinCX91600G2026Spectrum7204T,CPOLP30

LP35NVFP4GraceCPUVeraCPU2024OberonNVL72OberonNVL72NVL576BlueField-3BlueField-4Spectrum6102T,

CPOKyberNVL144LP40NVLinkRosaCPUFeynmanDieStacking

CustomHBMNVLink6

Switch

3600GB/sNVLink

7Switch

3600GB/sKyberNVL144NVL1152NVLink8

CPO2028OberonETL256OberonETL256OberonNVL72Rubin

UltraHBM4eBlueField-5BlackwellUltraHBM3eBlackwellNVLink

5

1800GB/sSpectrum5

51TRubin

HBM4Blackwell

HBM3eCX10CX8

800GChips,Systems,

FacilitiesReference

Designs,Best

Practices,

MethodologiesNVIDIADSXAIFactory

PlatformExtremeCo-DesignatInfrastructureScalePower

LiquidCoolingDSXMax-Q

DSXFlexDSX

ExchangeDSXSimDSXLibraries,APIs,SoftwareRubin

DSXReferenceDesignsOmniverseDSX

BlueprintDSXEcosystemNVIDIADSXAIFactory

PlatformAcceleratesScalable,Energy-EfficientAIFactoryDeployment

DSX

Max-Q

DSX

Exchange

DSX

Flex

DSXSim

AnnouncingNVIDIASpace-1VeraRubin

ModuleInferenceInflectionArrivesAnnouncingNVIDIANemoClawforOpenClawAgents–

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论