11.打造使用RISC-V的可扩展AIML软件平台：从PyTorch到SiFive XM平台的完整部署流程.pdf

《11.打造使用RISC-V的可扩展AIML软件平台：从PyTorch到SiFive XM平台的完整部署流程.pdf》由会员分享，可在线阅读，更多相关《11.打造使用RISC-V的可扩展AIML软件平台：从PyTorch到SiFive XM平台的完整部署流程.pdf（25页珍藏版）》请在三个皮匠报告上搜索。

1、2025 RISC-V China SummitPhoebe ChenAIML Senior Software EngineerBuilding a Scalable AI/ML Software Stack for RISC-VFrom PyTorch to Deployment on SiFive Intelligence XM PlatformsShowcase of AchievementsIntroduce the SiFive AI/ML Software StackSiFive Intelligence XM PlatformsIntegrate matmul accelerat

2、or into IREEEnd-to-end deploymentPytorch BEV perception models to RISC-VAGENDASummary&Future works1234522025 SiFive,Inc.All rights reserved.SiFive Confidential Subject to NDAShowcase of Achievements3SiFive Confidential Subject to NDABEV Former Perception models on RISC-VHW Platform:SiFive XM Emulato

3、r GPU PredictionSiFive XM PredictionSiFive Confidential Subject to NDAIntroduce the SiFive AI/ML Software Stack5SiFive Confidential Subject to NDASiFive AI/ML Software StackVCIX MLIR DialectSiFive Kernel Library(SKL)Open Source LibrariesSiFiveSystem SoftwareIPC LibraryML Interpreter(TFLite,XNNPACK,O

4、NNXRuntime,Llama.cpp)SiFive Intelligence core seriesCustomer AcceleratorsSiFive collaboration with othersSiFive owned and maintained componentComponent provided by others SiFiveLLVM CompilerAll critical components enabled and ready to run high profile e2e models on SiFive platformsIREE-Open-sourced

5、MLIR compiler and runtime with SiFive uarch specific optimizationUtilizing highly tuned SiFive LLVM Compilers and SKL LibrariesFront end supports popular AI/ML models like Pytorch LLMTechnology foundation to be extended easily by customers as needed with IREE source code providedAI Models(Llama,BERT

6、,CNN based models,BEVDepth,BEVFormer,etc.)AI/ML engineeringDevelopment SoftwareDevelopment SoftwareHardwareML CompilerML RuntimeIREEIREEMLIRAI Frameworks6SiFive Confidential Subject to NDAIREE(Intermediate Representation Execution Environment)MLIR e2e based compiler&runtime7 Support various front-en