ROCmOpeneCosystem

OptimizedDeepLearning

Operations

The ROCm Open eCosystem, is an open-source HPC/Hyperscale-class platform for GPU computing that’s also

programming-languageindependent.ROCmbringschoice,minimalismandmodularsoftwaredevelopmenttoGPUcomputing.

Optimized Deep Learning Operations with �exible

mixed-precision FP16, FP32, INT8 & INT4

capabilitiesbringscustomerssuperchargedcompute

performance to meet today’s demanding system

requirements of handling large data e�ciently for

trainingcomplexneuralnetworksandrunninginference

againstthosenetworksusedindeeplearning.

World’sFastestDoublePrecision

PCIe®Accelerator¹

Ultra-fast double precision performance with up to 7.4

TFLOPS FP64 performance on the Radeon Instinct™ MI60

Compute GPU, allows scientists and researches across the

globe to more e�ciently process HPC parallel codes across

several industries including life sciences, energy, �nance,

automotive and aerospace, academics, government, defense

andmore.

AMDIn�nityFabric™Link

Two In�nity Fabric™ Links per GPU for high speed

Direct-ConnectGPUhivesdeliveringupto200GB/s

GPU peer-to-peer bandwidth – 6x faster than using

PCIe3.0alone.⁷

Ultra-FastHBM2Memory

With 32 GB HBM2 memory utilizing a four-stack memory

con�guration, the new AMD Radeon Instinct MI60 delivers

ultra-high memory bandwidth with up to 1TB/s memory

bandwidth.

AMDRADEONINSTINCT™MI60

1.Calculated on Oct 22, 2018, the Radeon Instinct MI60 GPU resulted in 7.4 TFLOPS peak theoretical double precision �oating-point (FP64) performance. AMD TFLOPS calculations conducted with the

followingequation:FLOPScalculationsareperformedbytakingtheengineclockfromthehighestDPMstateandmultiplyingitbyxxCUsperGPU.Then,multiplyingthatnumberbyxxstreamprocessors,which

exist in each CU. Then, that number is multiplied by 1/2 FLOPS per clock for FP64. TFLOP calculations for MI60 can be found at https://www.amd.com/en/products/professional-graphics/instinct-mi60

External results on the NVidia Tesla V100 (16GB card) GPU accelerator resulted in 7 TFLOPS peak double precision (FP64) �oating-point performance. Results found at:

https://images.nvidia.com/content/technologies/volta/pdf/437317-Volta-V100-DS-NV-US-WEB.pdf AMD has not independently tested or veri�ed external/third party results/data and bears no

responsibilityforanyerrorsoromissionstherein.

2.Works with PCIe® Gen 4.0 and Gen 3.0 compliant motherboards. Performance may vary from motherboard to motherboard. Refer to system or motherboard provider for individual product

performanceandfeatures.

3.ECCsupporton2ndGenRadeonInstinct™GPUcards,basedonthe“Vega7nm”technologyhasbeenextendedtofull-chipECCincludingHBM2memoryandinternalGPUstructures.

4.ExpandedRAS(Reliability,availabilityandserviceability)attributeshavebeenaddedtoAMD’s2ndGenRadeonInstinct™Vega7nmtechnologybasedGPUcardsandtheirsupportingecosystemincluding

software,�rmwareandsystemlevelfeatures.AMD’sremotemanageabilitycapabilitiesusingadvancedout-of-bandcircuitryallowforeasierGPUmonitoringviaI2C,regardlessoftheGPUstate.Forfullsystem

RAScapabilities,refertothesystemmanufacturer’sguidelinesforspeci�csystemmodels

5.MxGPUSR-IOVtechnologyforhardware-basedvirtualizedcomputedeploymentsonlyavailableonspeci�edRadeonInstinct™serverGPUcards(MI50andMI60)whenusingKVMasahypervisor.MxGPU

SR-IOVvirtualizationusagebeyondthisrequiresfurtherdevelopmentwithAMDfordeployment.Currentplansfor“Vega7nm”basedproductsupportwillinclude1VMperGPUupto8GPUsinaVM.

6.TheRadeonInstinctGPUacceleratorproductscomewithathree-yearlimitedwarranty.Pleasevisitwww.AMD.com/warrantyforwarrantydetailsonthespeci�cgraphicsproductspurchased.Toll-freephone

serviceavailableintheU.S.andCanadaonly,emailaccessisglobal.

7.AsofOct22,2018.RadeonInstinct™MI50andMI60“Vega7nm”technology-basedacceleratorsarePCIe®Gen4.0*capableprovidingupto64GB/speaktheoreticaltransportdatabandwidthfromCPUto

GPU per card with PCIe Gen 4.0 x16 certi�ed servers. Previous Gen Radeon Instinct compute GPU cards are based on PCIe Gen 3.0 providing up to 32 GB/s peak theoretical transport rate bandwidth

performance.PeaktheoreticaltransportrateperformanceiscalculatedbyBaudRate*widthinbytes*#directions=GB/spercard.PCIeGen3:8*2*2=32GB/s.PCIeGen4:16*2*2=64GB/s.Radeon

Instinct™MI50andMI60“Vega7nm”technology-basedacceleratorsincludedualIn�nityFabric™Linksprovidingupto200GB/speaktheoreticalGPUtoGPUorPeer-to-Peer(P2P)transportratebandwidth

performanceperGPUcard.CombinedwithPCIeGen4compatibilityprovidinganaggregateGPUcardI/Opeakbandwidthofupto264GB/s.Performanceguidelinesareestimatedonlyandmayvary.Previous

GenRadeonInstinctcomputeGPUcardsprovideupto32GB/speakPCIeGen3.0bandwidthperformance.In�nityFabric™LinktechnologypeaktheoreticaltransportrateperformanceiscalculatedbyBaud

Rate*widthinbytes*#directions*#links=GB/spercard.In�nityFabricLink:25*2*2=100GB/s.MI50|MI60eachhavetwolinks:100GB/s*2linksperGPU=200GB/s.Refertoservermanufacture

PCIe Gen 4.0 compatibility and performance guidelines for potential peak performance of the speci�ed server model numbers. Server manufacturers may vary con�guration o�erings

yielding di�erent results. https://pcisig.com/ , https://www.chipestimate.com/PCI-Express-Gen-4-a-Big-Pipe-for-Big-Data/Cadence/Technical-Article/2014/04/15 ,

https://www.tomshardware.com/news/pcie-4.0-power-speed-express,32525.html AMD has not independently tested or veri�ed external/third party results/data and bears no

responsibility for any errors or omissions therein. *Pending

©2018AdvancedMicroDevices,Inc. Allrightsreserved.AMD,theAMDArrowlogo,Radeon,andcombinationsthereofaretrademarksofAdvancedMicroDevices,Inc. OpenCLisatrademarkofAppleInc.

usedbypermissionbyKhronos.Otherproductnamesusedinthispublicationareforidenti�cationpurposesonlyandmaybetrademarksoftheirrespectivecompanies.

ROCmOpeneCosystem

DEEPLEARNINGANDHPCAPPLICATIONS

HARDWAREABSTRACTION

OPTIMIZEDECOSYSTEM

DeepLearningandHPCLibraries

OPENFRAMEWORKS

Energy

Life Sciences

Automotive

Financial Services

HPC

Cloud / Hyperscale