1.MeasurementsconductedbyAMDPerformanceLabsasofJune2,2017ontheRadeonInstinct™MI25“Vega”architecturebasedaccelerator.

Resultsareestimatesonlyandmayvary.Performancemayvarybasedonuseoflatestdrivers.PC/systemmanufacturersmayvaryconfigurations

yieldingdifferentresults.TheresultscalculatedforRadeonInstinctMI25resultedin24.6TFLOPSpeakhalfprecision(FP16)and12.3TFLOPSpeak

singleprecision(FP32)floating-pointperformance.

AMDTFLOPScalculationsconductedwiththefollowingequation:FLOPScalculationsareperformedbytakingtheengineclockfromthehighest

DPMstateandmultiplyingitbyxxCUsperGPU.Then,multiplyingthatnumberbyxxstreamprocessors,whichexistineachCU.Then,thatnumber

ismultipliedby2FLOPSperclockforFP32.TocalculateTFLOPSforFP16,4FLOPSperclockwereused. TheFP64TFLOPSrateiscalculatedusing

1/16thrate.

ExternalresultsontheNVidiaTeslaP100-16(16GBcard)GPUAcceleratorresultedin18.7TFLOPSpeakhalfprecision(FP16)and9.3TFLOPSpeak

singleprecision(FP32)floating-pointperformance.

Resultsfoundat:https://images.nvidia.com/content/tesla/pdf/nvidia-tesla-p100-PCIe-datasheet.pdf.

ExternalresultsontheNVidiaTeslaP100-SXM2GPUAcceleratorresultedin21.2TFLOPSpeakhalfprecision(FP16)and10.6TFLOPSpeaksingle

precision(FP32)floating-pointperformance.

Resultsfoundat:http://www.nvidia.com/object/tesla-p100.html

AMDhasnotindependentlytestedorverifiedexternal/thirdpartyresults/dataandbearsnoresponsibilityforanyerrorsoromissionstherein.RIV-1

2.MeasurementsconductedbyAMDPerformanceLabsasofJune2,2017ontheRadeonInstinct™MI25“Vega”architecturebasedaccelerator.

Resultsareestimatesonlyandmayvary.Performancemayvarybasedonuseoflatestdrivers.PC/systemmanufacturersmayvaryconfigurations

yieldingdifferentresults.

TheresultscalculatedforRadeonInstinctMI25resultedin82GFLOPS/wattpeakhalfprecision(FP16)or41GFLOPS/wattpeaksingleprecision

(FP32)floating-pointperformance.

AMDGFLOPSperwattcalculationsconductedwiththefollowingequation:FLOPScalculationsareperformedbytakingtheengineclockfrom

thehighestDPMstateandmultiplyingitbyxxCUsperGPU.Then,multiplyingthatnumberbyxxstreamprocessors,whichexistineachCU.Then,

thatnumberismultipliedby2FLOPSperclockforFP32.TocalculateTFLOPSforFP16,4FLOPSperclockwereused. TheFP64TFLOPSrate

iscalculatedusing1/16thrate.OncetheTFLOPsarecalculated,thenumberisdividedbythexxxwattsTDPpowerandmultipliedby1,000to

determinetheGFLOPSperwatt.

CalculationsconductedbyAMDPerformanceLabsasofJune2,2017ontheNVidiaTeslaP100-16(16GBcard)GPUAcceleratortodetermine

GFLOPS/wattbydividingTFLOPSresultsby250wattsTDPresultedin75GFLOPSperwattpeakhalfprecision(FP16)and37GFLOPSperwattpeak

singleprecision(FP32)floating-pointperformance.

Sources:https://images.nvidia.com/content/tesla/pdf/nvidia-tesla-p100-PCIe-datasheet.pdf

CalculationsconductedbyAMDPerformanceLabsasofJune2,2017ontheNVidiaTeslaP100-SXM2GPUAcceleratortodetermineGFLOPS/

wattbydividingTFLOPSresultsby300wattsTDPresultedin71GFLOPSperwattpeakhalfprecision(FP16)and35GFLOPSperwattpeaksingle

precision(FP32)floating-pointperformance.

Sources:

http://www.nvidia.com/object/tesla-p100.html

AMDhasnotindependentlytestedorverifiedexternal/thirdpartyresults/dataandbearsnoresponsibilityforanyerrorsoromissionstherein.RIV-4

3.ECCsupportislimitedtotheHBM2memoryandECCprotectionisnotprovidedforinternalGPUstructures.

4.TheRadeonInstinctGPUacceleratorproductscomewithathreeyearlimitedwarranty.Pleasevisitwww.AMD.com/warrantyfordetailsonthe

specificgraphicsproductspurchased.Toll-freephoneserviceavailableintheU.S.andCanadaonly,emailaccessisglobal.

5.SupportforPythonisplanned,butstillunderdevelopment.

6.DatabasedonAMDEngineeringdesignofVega. RadeonR9FuryXhas4geometryenginesandapeakof4polygonsperclock.

Vegaisdesignedtohandleupto11polygonsperclockwith4geometryengines. Thisrepresentsanincreaseof2.6x. VG-3

FOOTNOTES