8.AM仇径-面向神经网络的RISC-V架构矩阵扩展创新(2).pdf

《8.AM仇径-面向神经网络的RISC-V架构矩阵扩展创新(2).pdf》由会员分享，可在线阅读，更多相关《8.AM仇径-面向神经网络的RISC-V架构矩阵扩展创新(2).pdf（22页珍藏版）》请在三个皮匠报告上搜索。

1、面向神经网络的RISC-V架构矩阵扩展创新仇径01020304GeneralIntroDetailTalksMatrixProposalFutureRoadmapGeneral Introduction矩阵扩展The Architecture of Alexnet*Imagenet classification with deep convolutional neural networksJ.The Architecture of Swim Transformer*Swin transformer:Hierarchical vision transformer using shifted wi

2、ndowsVectorN elements-N operationsMatrixN2elements-N3operationsPerformance improvement2x 8x performance boostReduced memory bandwidth requirementless data to generate more operations(data memory bandwidth)one instruction contains N3operations(inst memory bandwidth)Inherently more power efficiency wi

3、th less data movementMatrix is All You NeedRISC-V 矩阵扩展CustomizeOpen and Free ArchitectureLightweight and EfficientSoftware CompatibilityModularityGoals for RISC-V Matrix ExtensionEfficiencyImprove AI performance by accelerating matrix operationsVersatility&Scalability From low-powered edge devices t

4、o high-performance data centerFlexibilityMultiple data types and matrix shapesPortabilityBinary code portable across register size and hardware implementationFuture-ProofEasily extensible for future AI Matrix+ExtensionDetail Talks处理器中的矩阵扩展Fully integrated facilityReuse vector registers for source op

5、erandsReuse vector registers for accumulatorsHybrid facilityReuse vector registers for source operandsIndependent accumulator registers Attached facilityAdditional matrix registersWhat others doReuse vector registers for source operands and accumulators SiFive Intelligence ExtensionReuse vector regi

6、sters for source operandsindependent accumulator registerPower MMAArm SMEindependent source operands and accumulators Intel AMXApple AMXStream Computing Matrix Extension独立的矩阵扩展Diversified market needsParallel processingImplementation FlexibilityAgile development(both hardware and software)Power Opti