微軟-汽車生成式AI云端應用- Automotive GenAI Applications on the Cloud 2024

AutomotiveGenAI

ApplicationsontheCloud

October9th,2024

EstiMorYosef

SeniorProgramManager-AzureCore

GenAIinfluenceonindustries

MANUFACTURINGHEALTHCAREAND

PHARMACEUTICALS

RETAILMEDIAAND

ENTERTAINMENT

FINANCIAL

SERVICESSECTOR

Digitalassistant

WhatisGenerative

AI?

Atypeofartificial

intelligencethatcancreate

originalcontentinresponsetoauser’sprompt.

AD/ADASdata

challenges

Datacollection:

EdgecasesDiversity

LogisticallycomplicatedTimeconsuming

Expensive

Datasearch:

Findingspecificscenariosand

events

Whatarethe

solutions?

?Datageneration

?Dataaugmentation

?Datadescriptionandsearch

DataGeneration

IntroducingGAIA-1:ACutting-EdgeGenerativeAIModelforAutonomy-Wayve

DataAugmentation

DataDescriptionandSearch

HowisMicrosoft

enablingusers,customersandpartners

Microsoftasaplatform

步cognata/nsysCSPACESYNOPSYS"SIEMENS

partner

solution

Ecosystem

CI/CD

Toolchain

Github

ML

Toolchain

AzureInfrastructure

AML

AzureAIstudio

ModelasaService

AzureNVV710v5

EmpoweringReal-TimeAI/MLInferencingandAdvanced

VisualizationintheCloud

IntroducingAzureNVV710v5:Real-TimeAI/MLInferencing&Visualization|MicrosoftAzure

Thankyou

EstiMorYosef

esthermo@

[Public]

AMDCollaborationwithMicrosoft&Cognata

ROCmEnablementonAzureNVv5-V710

DockerImageDevelopment

√DevelopedOpenGLDockerImageforCognataSimCloud&SimRendering

√OptimizedtheAMDOpenGL-PdriverforCognataSimCloud

√DevelopedPyTorchDockerImageforCognataSimCloudengine

√PackagedlatestAMDdrivers&tools

ModelOptimization

√HIPifiedtheBEVFusionmodelforexecutiononAMDV710

√Verifiedoptimalinference

performancewithBEVFusion

√Developedthefirst-ever

ROCmdriverforRadeon

√EnabledROCmstackforcloudSR-IOVvirtualization

√VerifiedPyTorch&MIGraphXworkflows

√OptimizedforCognatamodelsandapplications

1|

[Public]

AzureNVadsV710v5–AcceleratedVDI&

MLInferenceintheCloud

VMName

NV4ads_V710_v5

NV8ads_V710_v5

NV12ads_V710_v5

NV24ads_V710_v5

NV28adms_V710_v5

GPU

1/6V710PCIe

1/3V710PCIe

1/2V710PCIe

1xV710PCIe

1xV710PCIe

GPUMemoryperVM

4GB

8GB

12GB

24GB

24GB

vCPU

(Threads)

4

8

12

24

28

RAM

16GB

32GB

64GB

128GB

160GB

TempDisk

180GB

360GB

720GB

1024GB

1024GB

NVV710

Comingsoon…

?PublicPreviewinQ32024

?GAinQ42024

?SuccessortoAzureNVv4

Designedformixedworkloads

?GPU-acceleratedRemoteDesktop

?Multi-sessionVDI

?MachineLearningInference

?VideoTranscoding

?ProfessionalRendering

2|AMDINTERNALUSEONLY|2024

[Public]

AMDRadeonPROV710GPU

AMDRDNATM3Architecture

Virtualization

SR-IOVGPU

28GBGDDR6

Units

memory

54Compute

DataPoisoning,

processors

FBRemapping

3456Stream

SupportforAV1

Encoding

448GB/s

MemoryBandwidth

Single-slot@158W

RobustAMDROCmSupportforAI/ML

3|AMDINTERNALUSEONLY|2024

[Public]

AMDCloudGPUsforAI/HPC

TwoGPUArchitecturesfortwodifferentmarkets

AMDRadeonPROV710

AMDInstinct

Hardwarevirtualizationformulti-userDesktops,Gaming,MachineLearning

GPUCapability

OptimizedforAI/ComputeatScale(nographicssupport)

GPUArchitecture

AMDCDNA(Compute

DNA)

AMD

RDNA(Radeon

DNA)

VirtualDesktopInfrastructureCloudGaming

Personal/BatchedInference

ExampleUse-Cases

LargeModelTrainingSupercomputing

Real-TimeInferenceasaService

SoftwarePlatformAMDROCmTMAMDROCm

4|AMDINTERNALUSEONLY|2024

[Public]

LpenivoicwarerlavormrorIrulomnpuie

unockeaHUHowerio

Acceleraieoinpuiauoniallasks

Lpuinizeuioimr?aiuLeepLealnligvvoik.oausalbcale

?penivouiceciaoluniyiiniovauoi,uiierenuauon,anuLolaooiauon1

5|

bcicililairv?hppuuppuil

upclauilyuyblcinbuuppuii

iusiciLepIuyiiic1il

rlaliICWUIhuupvuil

Liorarles

rIuyiaunnuilyivuucls

LcvclupiiIciit1uulciiaul

LepluyliIcit1UUIS

vunpiciprofilerTracerhipiiyGPUFort

[Public]

AMDHIP:ASTRONGCUDAALTERNATIVE

HIPisaC++runtimeAPIandkernellanguagethat

allowsdeveloperstocreateportableapplicationsthatcanrunonbothAMDandCUDAdevices.

KeyBenefitsofHIP:

?HIPcompiler(hipcc)canbuildmostCUDAapplicationswithminimaltonoportingeffort

?HIPissyntacticallyverysimilartoCUDA

?Open-source

?RunsonbothAMDand